OpenCloudOS-Kernel/mm/slub.c

5857 lines
144 KiB
C
Raw Normal View History

License cleanup: add SPDX GPL-2.0 license identifier to files with no license Many source files in the tree are missing licensing information, which makes it harder for compliance tools to determine the correct license. By default all files without license information are under the default license of the kernel, which is GPL version 2. Update the files which contain no license information with the 'GPL-2.0' SPDX license identifier. The SPDX identifier is a legally binding shorthand, which can be used instead of the full boiler plate text. This patch is based on work done by Thomas Gleixner and Kate Stewart and Philippe Ombredanne. How this work was done: Patches were generated and checked against linux-4.14-rc6 for a subset of the use cases: - file had no licensing information it it. - file was a */uapi/* one with no licensing information in it, - file was a */uapi/* one with existing licensing information, Further patches will be generated in subsequent months to fix up cases where non-standard license headers were used, and references to license had to be inferred by heuristics based on keywords. The analysis to determine which SPDX License Identifier to be applied to a file was done in a spreadsheet of side by side results from of the output of two independent scanners (ScanCode & Windriver) producing SPDX tag:value files created by Philippe Ombredanne. Philippe prepared the base worksheet, and did an initial spot review of a few 1000 files. The 4.13 kernel was the starting point of the analysis with 60,537 files assessed. Kate Stewart did a file by file comparison of the scanner results in the spreadsheet to determine which SPDX license identifier(s) to be applied to the file. She confirmed any determination that was not immediately clear with lawyers working with the Linux Foundation. Criteria used to select files for SPDX license identifier tagging was: - Files considered eligible had to be source code files. - Make and config files were included as candidates if they contained >5 lines of source - File already had some variant of a license header in it (even if <5 lines). All documentation files were explicitly excluded. The following heuristics were used to determine which SPDX license identifiers to apply. - when both scanners couldn't find any license traces, file was considered to have no license information in it, and the top level COPYING file license applied. For non */uapi/* files that summary was: SPDX license identifier # files ---------------------------------------------------|------- GPL-2.0 11139 and resulted in the first patch in this series. If that file was a */uapi/* path one, it was "GPL-2.0 WITH Linux-syscall-note" otherwise it was "GPL-2.0". Results of that was: SPDX license identifier # files ---------------------------------------------------|------- GPL-2.0 WITH Linux-syscall-note 930 and resulted in the second patch in this series. - if a file had some form of licensing information in it, and was one of the */uapi/* ones, it was denoted with the Linux-syscall-note if any GPL family license was found in the file or had no licensing in it (per prior point). Results summary: SPDX license identifier # files ---------------------------------------------------|------ GPL-2.0 WITH Linux-syscall-note 270 GPL-2.0+ WITH Linux-syscall-note 169 ((GPL-2.0 WITH Linux-syscall-note) OR BSD-2-Clause) 21 ((GPL-2.0 WITH Linux-syscall-note) OR BSD-3-Clause) 17 LGPL-2.1+ WITH Linux-syscall-note 15 GPL-1.0+ WITH Linux-syscall-note 14 ((GPL-2.0+ WITH Linux-syscall-note) OR BSD-3-Clause) 5 LGPL-2.0+ WITH Linux-syscall-note 4 LGPL-2.1 WITH Linux-syscall-note 3 ((GPL-2.0 WITH Linux-syscall-note) OR MIT) 3 ((GPL-2.0 WITH Linux-syscall-note) AND MIT) 1 and that resulted in the third patch in this series. - when the two scanners agreed on the detected license(s), that became the concluded license(s). - when there was disagreement between the two scanners (one detected a license but the other didn't, or they both detected different licenses) a manual inspection of the file occurred. - In most cases a manual inspection of the information in the file resulted in a clear resolution of the license that should apply (and which scanner probably needed to revisit its heuristics). - When it was not immediately clear, the license identifier was confirmed with lawyers working with the Linux Foundation. - If there was any question as to the appropriate license identifier, the file was flagged for further research and to be revisited later in time. In total, over 70 hours of logged manual review was done on the spreadsheet to determine the SPDX license identifiers to apply to the source files by Kate, Philippe, Thomas and, in some cases, confirmation by lawyers working with the Linux Foundation. Kate also obtained a third independent scan of the 4.13 code base from FOSSology, and compared selected files where the other two scanners disagreed against that SPDX file, to see if there was new insights. The Windriver scanner is based on an older version of FOSSology in part, so they are related. Thomas did random spot checks in about 500 files from the spreadsheets for the uapi headers and agreed with SPDX license identifier in the files he inspected. For the non-uapi files Thomas did random spot checks in about 15000 files. In initial set of patches against 4.14-rc6, 3 files were found to have copy/paste license identifier errors, and have been fixed to reflect the correct identifier. Additionally Philippe spent 10 hours this week doing a detailed manual inspection and review of the 12,461 patched files from the initial patch version early this week with: - a full scancode scan run, collecting the matched texts, detected license ids and scores - reviewing anything where there was a license detected (about 500+ files) to ensure that the applied SPDX license was correct - reviewing anything where there was no detection but the patch license was not GPL-2.0 WITH Linux-syscall-note to ensure that the applied SPDX license was correct This produced a worksheet with 20 files needing minor correction. This worksheet was then exported into 3 different .csv files for the different types of files to be modified. These .csv files were then reviewed by Greg. Thomas wrote a script to parse the csv files and add the proper SPDX tag to the file, in the format that the file expected. This script was further refined by Greg based on the output to detect more types of files automatically and to distinguish between header and source .c files (which need different comment types.) Finally Greg ran the script using the .csv files to generate the patches. Reviewed-by: Kate Stewart <kstewart@linuxfoundation.org> Reviewed-by: Philippe Ombredanne <pombredanne@nexb.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2017-11-01 22:07:57 +08:00
// SPDX-License-Identifier: GPL-2.0
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/*
* SLUB: A slab allocator that limits cache line use instead of queuing
* objects in per cpu and per node lists.
*
* The allocator synchronizes using per slab locks or atomic operations
* and only uses a centralized lock to manage a pool of partial slabs.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*
* (C) 2007 SGI, Christoph Lameter
* (C) 2011 Linux Foundation, Christoph Lameter
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*/
#include <linux/mm.h>
#include <linux/swap.h> /* struct reclaim_state */
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
#include <linux/module.h>
#include <linux/bit_spinlock.h>
#include <linux/interrupt.h>
#include <linux/bitops.h>
#include <linux/slab.h>
#include "slab.h"
#include <linux/proc_fs.h>
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
#include <linux/seq_file.h>
mm: slub: introduce metadata_access_enable()/metadata_access_disable() It's ok for slub to access memory that marked by kasan as inaccessible (object's metadata). Kasan shouldn't print report in that case because these accesses are valid. Disabling instrumentation of slub.c code is not enough to achieve this because slub passes pointer to object's metadata into external functions like memchr_inv(). We don't want to disable instrumentation for memchr_inv() because this is quite generic function, and we don't want to miss bugs. metadata_access_enable/metadata_access_disable used to tell KASan where accesses to metadata starts/end, so we could temporarily disable KASan reports. Signed-off-by: Andrey Ryabinin <a.ryabinin@samsung.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Konstantin Serebryany <kcc@google.com> Cc: Dmitry Chernenkov <dmitryc@google.com> Signed-off-by: Andrey Konovalov <adech.fo@gmail.com> Cc: Yuri Gribov <tetra2005@gmail.com> Cc: Konstantin Khlebnikov <koct9i@gmail.com> Cc: Sasha Levin <sasha.levin@oracle.com> Cc: Christoph Lameter <cl@linux.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Ingo Molnar <mingo@elte.hu> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-14 06:39:38 +08:00
#include <linux/kasan.h>
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
#include <linux/cpu.h>
#include <linux/cpuset.h>
#include <linux/mempolicy.h>
#include <linux/ctype.h>
infrastructure to debug (dynamic) objects We can see an ever repeating problem pattern with objects of any kind in the kernel: 1) freeing of active objects 2) reinitialization of active objects Both problems can be hard to debug because the crash happens at a point where we have no chance to decode the root cause anymore. One problem spot are kernel timers, where the detection of the problem often happens in interrupt context and usually causes the machine to panic. While working on a timer related bug report I had to hack specialized code into the timer subsystem to get a reasonable hint for the root cause. This debug hack was fine for temporary use, but far from a mergeable solution due to the intrusiveness into the timer code. The code further lacked the ability to detect and report the root cause instantly and keep the system operational. Keeping the system operational is important to get hold of the debug information without special debugging aids like serial consoles and special knowledge of the bug reporter. The problems described above are not restricted to timers, but timers tend to expose it usually in a full system crash. Other objects are less explosive, but the symptoms caused by such mistakes can be even harder to debug. Instead of creating specialized debugging code for the timer subsystem a generic infrastructure is created which allows developers to verify their code and provides an easy to enable debug facility for users in case of trouble. The debugobjects core code keeps track of operations on static and dynamic objects by inserting them into a hashed list and sanity checking them on object operations and provides additional checks whenever kernel memory is freed. The tracked object operations are: - initializing an object - adding an object to a subsystem list - deleting an object from a subsystem list Each operation is sanity checked before the operation is executed and the subsystem specific code can provide a fixup function which allows to prevent the damage of the operation. When the sanity check triggers a warning message and a stack trace is printed. The list of operations can be extended if the need arises. For now it's limited to the requirements of the first user (timers). The core code enqueues the objects into hash buckets. The hash index is generated from the address of the object to simplify the lookup for the check on kfree/vfree. Each bucket has it's own spinlock to avoid contention on a global lock. The debug code can be compiled in without being active. The runtime overhead is minimal and could be optimized by asm alternatives. A kernel command line option enables the debugging code. Thanks to Ingo Molnar for review, suggestions and cleanup patches. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@elte.hu> Cc: Greg KH <greg@kroah.com> Cc: Randy Dunlap <randy.dunlap@oracle.com> Cc: Kay Sievers <kay.sievers@vrfy.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-30 15:55:01 +08:00
#include <linux/debugobjects.h>
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
#include <linux/kallsyms.h>
mm, kfence: insert KFENCE hooks for SLUB Inserts KFENCE hooks into the SLUB allocator. To pass the originally requested size to KFENCE, add an argument 'orig_size' to slab_alloc*(). The additional argument is required to preserve the requested original size for kmalloc() allocations, which uses size classes (e.g. an allocation of 272 bytes will return an object of size 512). Therefore, kmem_cache::size does not represent the kmalloc-caller's requested size, and we must introduce the argument 'orig_size' to propagate the originally requested size to KFENCE. Without the originally requested size, we would not be able to detect out-of-bounds accesses for objects placed at the end of a KFENCE object page if that object is not equal to the kmalloc-size class it was bucketed into. When KFENCE is disabled, there is no additional overhead, since slab_alloc*() functions are __always_inline. Link: https://lkml.kernel.org/r/20201103175841.3495947-6-elver@google.com Signed-off-by: Marco Elver <elver@google.com> Signed-off-by: Alexander Potapenko <glider@google.com> Reviewed-by: Dmitry Vyukov <dvyukov@google.com> Reviewed-by: Jann Horn <jannh@google.com> Co-developed-by: Marco Elver <elver@google.com> Cc: Andrey Konovalov <andreyknvl@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christopher Lameter <cl@linux.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: David Rientjes <rientjes@google.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Hillf Danton <hdanton@sina.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Joern Engel <joern@purestorage.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kees Cook <keescook@chromium.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Paul E. McKenney <paulmck@kernel.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: SeongJae Park <sjpark@amazon.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-26 09:19:16 +08:00
#include <linux/kfence.h>
#include <linux/memory.h>
#include <linux/math64.h>
#include <linux/fault-inject.h>
#include <linux/stacktrace.h>
#include <linux/prefetch.h>
#include <linux/memcontrol.h>
mm: add SLUB free list pointer obfuscation This SLUB free list pointer obfuscation code is modified from Brad Spengler/PaX Team's code in the last public patch of grsecurity/PaX based on my understanding of the code. Changes or omissions from the original code are mine and don't reflect the original grsecurity/PaX code. This adds a per-cache random value to SLUB caches that is XORed with their freelist pointer address and value. This adds nearly zero overhead and frustrates the very common heap overflow exploitation method of overwriting freelist pointers. A recent example of the attack is written up here: http://cyseclabs.com/blog/cve-2016-6187-heap-off-by-one-exploit and there is a section dedicated to the technique the book "A Guide to Kernel Exploitation: Attacking the Core". This is based on patches by Daniel Micay, and refactored to minimize the use of #ifdef. With 200-count cycles of "hackbench -g 20 -l 1000" I saw the following run times: before: mean 10.11882499999999999995 variance .03320378329145728642 stdev .18221905304181911048 after: mean 10.12654000000000000014 variance .04700556623115577889 stdev .21680767106160192064 The difference gets lost in the noise, but if the above is to be taken literally, using CONFIG_FREELIST_HARDENED is 0.07% slower. Link: http://lkml.kernel.org/r/20170802180609.GA66807@beast Signed-off-by: Kees Cook <keescook@chromium.org> Suggested-by: Daniel Micay <danielmicay@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Tycho Andersen <tycho@docker.com> Cc: Alexander Popov <alex.popov@linux.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:19:18 +08:00
#include <linux/random.h>
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
slub tracing: move trace calls out of always inlined functions to reduce kernel code size Having the trace calls defined in the always inlined kmalloc functions in include/linux/slub_def.h causes a lot of code duplication as the trace functions get instantiated for each kamalloc call site. This can simply be removed by pushing the trace calls down into the functions in slub.c. On my x86_64 built this patch shrinks the code size of the kernel by approx 36K and also shrinks the code size of many modules -- too many to list here ;) size vmlinux (2.6.36) reports text data bss dec hex filename 5410611 743172 828928 6982711 6a8c37 vmlinux 5373738 744244 828928 6946910 6a005e vmlinux + patch The resulting kernel has had some testing & kmalloc trace still seems to work. This patch - moves trace_kmalloc out of the inlined kmalloc() and pushes it down into kmem_cache_alloc_trace() so this it only get instantiated once. - rename kmem_cache_alloc_notrace() to kmem_cache_alloc_trace() to indicate that now is does have tracing. (maybe this would better being called something like kmalloc_kmem_cache ?) - adds a new function kmalloc_order() to handle allocation and tracing of large allocations of page order. - removes tracing from the inlined kmalloc_large() replacing them with a call to kmalloc_order(); - move tracing out of inlined kmalloc_node() and pushing it down into kmem_cache_alloc_node_trace - rename kmem_cache_alloc_node_notrace() to kmem_cache_alloc_node_trace() - removes the include of trace/events/kmem.h from slub_def.h. v2 - keep kmalloc_order_trace inline when !CONFIG_TRACE Signed-off-by: Richard Kennedy <richard@rsk.demon.co.uk> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2010-10-21 17:29:19 +08:00
#include <trace/events/kmem.h>
mm: sl[au]b: add knowledge of PFMEMALLOC reserve pages When a user or administrator requires swap for their application, they create a swap partition and file, format it with mkswap and activate it with swapon. Swap over the network is considered as an option in diskless systems. The two likely scenarios are when blade servers are used as part of a cluster where the form factor or maintenance costs do not allow the use of disks and thin clients. The Linux Terminal Server Project recommends the use of the Network Block Device (NBD) for swap according to the manual at https://sourceforge.net/projects/ltsp/files/Docs-Admin-Guide/LTSPManual.pdf/download There is also documentation and tutorials on how to setup swap over NBD at places like https://help.ubuntu.com/community/UbuntuLTSP/EnableNBDSWAP The nbd-client also documents the use of NBD as swap. Despite this, the fact is that a machine using NBD for swap can deadlock within minutes if swap is used intensively. This patch series addresses the problem. The core issue is that network block devices do not use mempools like normal block devices do. As the host cannot control where they receive packets from, they cannot reliably work out in advance how much memory they might need. Some years ago, Peter Zijlstra developed a series of patches that supported swap over an NFS that at least one distribution is carrying within their kernels. This patch series borrows very heavily from Peter's work to support swapping over NBD as a pre-requisite to supporting swap-over-NFS. The bulk of the complexity is concerned with preserving memory that is allocated from the PFMEMALLOC reserves for use by the network layer which is needed for both NBD and NFS. Patch 1 adds knowledge of the PFMEMALLOC reserves to SLAB and SLUB to preserve access to pages allocated under low memory situations to callers that are freeing memory. Patch 2 optimises the SLUB fast path to avoid pfmemalloc checks Patch 3 introduces __GFP_MEMALLOC to allow access to the PFMEMALLOC reserves without setting PFMEMALLOC. Patch 4 opens the possibility for softirqs to use PFMEMALLOC reserves for later use by network packet processing. Patch 5 only sets page->pfmemalloc when ALLOC_NO_WATERMARKS was required Patch 6 ignores memory policies when ALLOC_NO_WATERMARKS is set. Patches 7-12 allows network processing to use PFMEMALLOC reserves when the socket has been marked as being used by the VM to clean pages. If packets are received and stored in pages that were allocated under low-memory situations and are unrelated to the VM, the packets are dropped. Patch 11 reintroduces __skb_alloc_page which the networking folk may object to but is needed in some cases to propogate pfmemalloc from a newly allocated page to an skb. If there is a strong objection, this patch can be dropped with the impact being that swap-over-network will be slower in some cases but it should not fail. Patch 13 is a micro-optimisation to avoid a function call in the common case. Patch 14 tags NBD sockets as being SOCK_MEMALLOC so they can use PFMEMALLOC if necessary. Patch 15 notes that it is still possible for the PFMEMALLOC reserve to be depleted. To prevent this, direct reclaimers get throttled on a waitqueue if 50% of the PFMEMALLOC reserves are depleted. It is expected that kswapd and the direct reclaimers already running will clean enough pages for the low watermark to be reached and the throttled processes are woken up. Patch 16 adds a statistic to track how often processes get throttled Some basic performance testing was run using kernel builds, netperf on loopback for UDP and TCP, hackbench (pipes and sockets), iozone and sysbench. Each of them were expected to use the sl*b allocators reasonably heavily but there did not appear to be significant performance variances. For testing swap-over-NBD, a machine was booted with 2G of RAM with a swapfile backed by NBD. 8*NUM_CPU processes were started that create anonymous memory mappings and read them linearly in a loop. The total size of the mappings were 4*PHYSICAL_MEMORY to use swap heavily under memory pressure. Without the patches and using SLUB, the machine locks up within minutes and runs to completion with them applied. With SLAB, the story is different as an unpatched kernel run to completion. However, the patched kernel completed the test 45% faster. MICRO 3.5.0-rc2 3.5.0-rc2 vanilla swapnbd Unrecognised test vmscan-anon-mmap-write MMTests Statistics: duration Sys Time Running Test (seconds) 197.80 173.07 User+Sys Time Running Test (seconds) 206.96 182.03 Total Elapsed Time (seconds) 3240.70 1762.09 This patch: mm: sl[au]b: add knowledge of PFMEMALLOC reserve pages Allocations of pages below the min watermark run a risk of the machine hanging due to a lack of memory. To prevent this, only callers who have PF_MEMALLOC or TIF_MEMDIE set and are not processing an interrupt are allowed to allocate with ALLOC_NO_WATERMARKS. Once they are allocated to a slab though, nothing prevents other callers consuming free objects within those slabs. This patch limits access to slab pages that were alloced from the PFMEMALLOC reserves. When this patch is applied, pages allocated from below the low watermark are returned with page->pfmemalloc set and it is up to the caller to determine how the page should be protected. SLAB restricts access to any page with page->pfmemalloc set to callers which are known to able to access the PFMEMALLOC reserve. If one is not available, an attempt is made to allocate a new page rather than use a reserve. SLUB is a bit more relaxed in that it only records if the current per-CPU page was allocated from PFMEMALLOC reserve and uses another partial slab if the caller does not have the necessary GFP or process flags. This was found to be sufficient in tests to avoid hangs due to SLUB generally maintaining smaller lists than SLAB. In low-memory conditions it does mean that !PFMEMALLOC allocators can fail a slab allocation even though free objects are available because they are being preserved for callers that are freeing pages. [a.p.zijlstra@chello.nl: Original implementation] [sebastian@breakpoint.cc: Correct order of page flag clearing] Signed-off-by: Mel Gorman <mgorman@suse.de> Cc: David Miller <davem@davemloft.net> Cc: Neil Brown <neilb@suse.de> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Christie <michaelc@cs.wisc.edu> Cc: Eric B Munson <emunson@mgebm.net> Cc: Eric Dumazet <eric.dumazet@gmail.com> Cc: Sebastian Andrzej Siewior <sebastian@breakpoint.cc> Cc: Mel Gorman <mgorman@suse.de> Cc: Christoph Lameter <cl@linux.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-08-01 07:43:58 +08:00
#include "internal.h"
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/*
* Lock order:
* 1. slab_mutex (Global Mutex)
* 2. node->list_lock
* 3. slab_lock(page) (Only on some arches and for debugging)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*
* slab_mutex
*
* The role of the slab_mutex is to protect the list of all the slabs
* and to synchronize major metadata changes to slab cache structures.
*
* The slab_lock is only used for debugging and on arches that do not
* have the ability to do a cmpxchg_double. It only protects:
* A. page->freelist -> List of object free in a page
* B. page->inuse -> Number of objects in use
* C. page->objects -> Number of objects in page
* D. page->frozen -> frozen state
*
* If a slab is frozen then it is exempt from list management. It is not
* on any list except per cpu partial list. The processor that froze the
* slab is the one who can perform list operations on the page. Other
* processors may put objects onto the freelist but the processor that
* froze the slab is the only one that can retrieve the objects from the
* page's freelist.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*
* The list_lock protects the partial and full list on each node and
* the partial slab counter. If taken then no new slabs may be added or
* removed from the lists nor make the number of partial slabs be modified.
* (Note that the total number of slabs is an atomic value that may be
* modified without taking the list lock).
*
* The list_lock is a centralized lock and thus we avoid taking it as
* much as possible. As long as SLUB does not have to handle partial
* slabs, operations can continue without any centralized lock. F.e.
* allocating a long series of objects that fill up slabs does not require
* the list lock.
* Interrupts are disabled during allocation and deallocation in order to
* make the slab allocator safe to use in the context of an irq. In addition
* interrupts are disabled to ensure that the processor does not change
* while handling per_cpu slabs, due to kernel preemption.
*
* SLUB assigns one slab for allocation to each processor.
* Allocations only occur from these slabs called cpu slabs.
*
* Slabs with free elements are kept on a partial list and during regular
* operations no list for full slabs is used. If an object in a full slab is
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
* freed then the slab will show up again on the partial lists.
* We track full slabs for debugging purposes though because otherwise we
* cannot scan all objects.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*
* Slabs are freed when they become empty. Teardown and setup is
* minimal so we rely on the page allocators per cpu caches for
* fast frees and allocs.
*
* page->frozen The slab is frozen and exempt from list processing.
* This means that the slab is dedicated to a purpose
* such as satisfying allocations for a specific
* processor. Objects may be freed in the slab while
* it is frozen but slab_free will then skip the usual
* list operations. It is up to the processor holding
* the slab to integrate the slab into the slab lists
* when the slab is no longer needed.
*
* One use of this flag is to mark slabs that are
* used for allocations. Then such a slab becomes a cpu
* slab. The cpu slab may be equipped with an additional
SLUB: Avoid page struct cacheline bouncing due to remote frees to cpu slab A remote free may access the same page struct that also contains the lockless freelist for the cpu slab. If objects have a short lifetime and are freed by a different processor then remote frees back to the slab from which we are currently allocating are frequent. The cacheline with the page struct needs to be repeately acquired in exclusive mode by both the allocating thread and the freeing thread. If this is frequent enough then performance will suffer because of cacheline bouncing. This patchset puts the lockless_freelist pointer in its own cacheline. In order to make that happen we introduce a per cpu structure called kmem_cache_cpu. Instead of keeping an array of pointers to page structs we now keep an array to a per cpu structure that--among other things--contains the pointer to the lockless freelist. The freeing thread can then keep possession of exclusive access to the page struct cacheline while the allocating thread keeps its exclusive access to the cacheline containing the per cpu structure. This works as long as the allocating cpu is able to service its request from the lockless freelist. If the lockless freelist runs empty then the allocating thread needs to acquire exclusive access to the cacheline with the page struct lock the slab. The allocating thread will then check if new objects were freed to the per cpu slab. If so it will keep the slab as the cpu slab and continue with the recently remote freed objects. So the allocating thread can take a series of just freed remote pages and dish them out again. Ideally allocations could be just recycling objects in the same slab this way which will lead to an ideal allocation / remote free pattern. The number of objects that can be handled in this way is limited by the capacity of one slab. Increasing slab size via slub_min_objects/ slub_max_order may increase the number of objects and therefore performance. If the allocating thread runs out of objects and finds that no objects were put back by the remote processor then it will retrieve a new slab (from the partial lists or from the page allocator) and start with a whole new set of objects while the remote thread may still be freeing objects to the old cpu slab. This may then repeat until the new slab is also exhausted. If remote freeing has freed objects in the earlier slab then that earlier slab will now be on the partial freelist and the allocating thread will pick that slab next for allocation. So the loop is extended. However, both threads need to take the list_lock to make the swizzling via the partial list happen. It is likely that this kind of scheme will keep the objects being passed around to a small set that can be kept in the cpu caches leading to increased performance. More code cleanups become possible: - Instead of passing a cpu we can now pass a kmem_cache_cpu structure around. Allows reducing the number of parameters to various functions. - Can define a new node_match() function for NUMA to encapsulate locality checks. Effect on allocations: Cachelines touched before this patch: Write: page cache struct and first cacheline of object Cachelines touched after this patch: Write: kmem_cache_cpu cacheline and first cacheline of object Read: page cache struct (but see later patch that avoids touching that cacheline) The handling when the lockless alloc list runs empty gets to be a bit more complicated since another cacheline has now to be written to. But that is halfway out of the hot path. Effect on freeing: Cachelines touched before this patch: Write: page_struct and first cacheline of object Cachelines touched after this patch depending on how we free: Write(to cpu_slab): kmem_cache_cpu struct and first cacheline of object Write(to other): page struct and first cacheline of object Read(to cpu_slab): page struct to id slab etc. (but see later patch that avoids touching the page struct on free) Read(to other): cpu local kmem_cache_cpu struct to verify its not the cpu slab. Summary: Pro: - Distinct cachelines so that concurrent remote frees and local allocs on a cpuslab can occur without cacheline bouncing. - Avoids potential bouncing cachelines because of neighboring per cpu pointer updates in kmem_cache's cpu_slab structure since it now grows to a cacheline (Therefore remove the comment that talks about that concern). Cons: - Freeing objects now requires the reading of one additional cacheline. That can be mitigated for some cases by the following patches but its not possible to completely eliminate these references. - Memory usage grows slightly. The size of each per cpu object is blown up from one word (pointing to the page_struct) to one cacheline with various data. So this is NR_CPUS*NR_SLABS*L1_BYTES more memory use. Lets say NR_SLABS is 100 and a cache line size of 128 then we have just increased SLAB metadata requirements by 12.8k per cpu. (Another later patch reduces these requirements) Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 16:26:05 +08:00
* freelist that allows lockless access to
slub: support concurrent local and remote frees and allocs on a slab Avoid atomic overhead in slab_alloc and slab_free SLUB needs to use the slab_lock for the per cpu slabs to synchronize with potential kfree operations. This patch avoids that need by moving all free objects onto a lockless_freelist. The regular freelist continues to exist and will be used to free objects. So while we consume the lockless_freelist the regular freelist may build up objects. If we are out of objects on the lockless_freelist then we may check the regular freelist. If it has objects then we move those over to the lockless_freelist and do this again. There is a significant savings in terms of atomic operations that have to be performed. We can even free directly to the lockless_freelist if we know that we are running on the same processor. So this speeds up short lived objects. They may be allocated and freed without taking the slab_lock. This is particular good for netperf. In order to maximize the effect of the new faster hotpath we extract the hottest performance pieces into inlined functions. These are then inlined into kmem_cache_alloc and kmem_cache_free. So hotpath allocation and freeing no longer requires a subroutine call within SLUB. [I am not sure that it is worth doing this because it changes the easy to read structure of slub just to reduce atomic ops. However, there is someone out there with a benchmark on 4 way and 8 way processor systems that seems to show a 5% regression vs. Slab. Seems that the regression is due to increased atomic operations use vs. SLAB in SLUB). I wonder if this is applicable or discernable at all in a real workload?] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-10 18:15:16 +08:00
* free objects in addition to the regular freelist
* that requires the slab lock.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*
* SLAB_DEBUG_FLAGS Slab requires special handling due to debug
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
* options set. This moves slab handling out of
slub: support concurrent local and remote frees and allocs on a slab Avoid atomic overhead in slab_alloc and slab_free SLUB needs to use the slab_lock for the per cpu slabs to synchronize with potential kfree operations. This patch avoids that need by moving all free objects onto a lockless_freelist. The regular freelist continues to exist and will be used to free objects. So while we consume the lockless_freelist the regular freelist may build up objects. If we are out of objects on the lockless_freelist then we may check the regular freelist. If it has objects then we move those over to the lockless_freelist and do this again. There is a significant savings in terms of atomic operations that have to be performed. We can even free directly to the lockless_freelist if we know that we are running on the same processor. So this speeds up short lived objects. They may be allocated and freed without taking the slab_lock. This is particular good for netperf. In order to maximize the effect of the new faster hotpath we extract the hottest performance pieces into inlined functions. These are then inlined into kmem_cache_alloc and kmem_cache_free. So hotpath allocation and freeing no longer requires a subroutine call within SLUB. [I am not sure that it is worth doing this because it changes the easy to read structure of slub just to reduce atomic ops. However, there is someone out there with a benchmark on 4 way and 8 way processor systems that seems to show a 5% regression vs. Slab. Seems that the regression is due to increased atomic operations use vs. SLAB in SLUB). I wonder if this is applicable or discernable at all in a real workload?] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-10 18:15:16 +08:00
* the fast path and disables lockless freelists.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*/
#ifdef CONFIG_SLUB_DEBUG
#ifdef CONFIG_SLUB_DEBUG_ON
DEFINE_STATIC_KEY_TRUE(slub_debug_enabled);
#else
DEFINE_STATIC_KEY_FALSE(slub_debug_enabled);
#endif
#endif
static inline bool kmem_cache_debug(struct kmem_cache *s)
{
return kmem_cache_debug_flags(s, SLAB_DEBUG_FLAGS);
}
void *fixup_red_left(struct kmem_cache *s, void *p)
{
if (kmem_cache_debug_flags(s, SLAB_RED_ZONE))
p += s->red_left_pad;
return p;
}
static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
{
#ifdef CONFIG_SLUB_CPU_PARTIAL
return !kmem_cache_debug(s);
#else
return false;
#endif
}
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/*
* Issues still to be resolved:
*
* - Support PAGE_ALLOC_DEBUG. Should be easy to do.
*
* - Variable sizing of the per node arrays
*/
/* Enable to test recovery from slab corruption on boot */
#undef SLUB_RESILIENCY_TEST
/* Enable to log cmpxchg failures */
#undef SLUB_DEBUG_CMPXCHG
/*
* Minimum number of partial slabs. These will be left on the partial
* lists even if they are empty. kmem_cache_shrink may reclaim them.
*/
#define MIN_PARTIAL 5
/*
* Maximum number of desirable partial slabs.
* The existence of more partial slabs makes kmem_cache_shrink
* sort the partial list by the number of objects in use.
*/
#define MAX_PARTIAL 10
#define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
SLAB_POISON | SLAB_STORE_USER)
/*
* These debug flags cannot use CMPXCHG because there might be consistency
* issues when checking or reading debug information
*/
#define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
SLAB_TRACE)
/*
* Debugging flags that require metadata to be stored in the slab. These get
* disabled when slub_debug=O is used and a cache's min order increases with
* metadata.
*/
#define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
#define OO_SHIFT 16
#define OO_MASK ((1 << OO_SHIFT) - 1)
#define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/* Internal SLUB flags */
/* Poison object */
#define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
/* Use cmpxchg_double */
#define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/*
* Tracking user of a slab.
*/
slub: Enable backtrace for create/delete points This patch attempts to grab a backtrace for the creation and deletion points of the slub object. When a fault is detected, we can then get a better idea of where the item was deleted. Example output from debugging some funky nfs/rpc behaviour: ============================================================================= BUG kmalloc-64: Object is on free-list ----------------------------------------------------------------------------- INFO: Allocated in rpcb_getport_async+0x39c/0x5a5 [sunrpc] age=381 cpu=3 pid=3750 __slab_alloc+0x348/0x3ba kmem_cache_alloc_trace+0x67/0xe7 rpcb_getport_async+0x39c/0x5a5 [sunrpc] call_bind+0x70/0x75 [sunrpc] __rpc_execute+0x78/0x24b [sunrpc] rpc_execute+0x3d/0x42 [sunrpc] rpc_run_task+0x79/0x81 [sunrpc] rpc_call_sync+0x3f/0x60 [sunrpc] rpc_ping+0x42/0x58 [sunrpc] rpc_create+0x4aa/0x527 [sunrpc] nfs_create_rpc_client+0xb1/0xf6 [nfs] nfs_init_client+0x3b/0x7d [nfs] nfs_get_client+0x453/0x5ab [nfs] nfs_create_server+0x10b/0x437 [nfs] nfs_fs_mount+0x4ca/0x708 [nfs] mount_fs+0x6b/0x152 INFO: Freed in rpcb_map_release+0x3f/0x44 [sunrpc] age=30 cpu=2 pid=29049 __slab_free+0x57/0x150 kfree+0x107/0x13a rpcb_map_release+0x3f/0x44 [sunrpc] rpc_release_calldata+0x12/0x14 [sunrpc] rpc_free_task+0x59/0x61 [sunrpc] rpc_final_put_task+0x82/0x8a [sunrpc] __rpc_execute+0x23c/0x24b [sunrpc] rpc_async_schedule+0x10/0x12 [sunrpc] process_one_work+0x230/0x41d worker_thread+0x133/0x217 kthread+0x7d/0x85 kernel_thread_helper+0x4/0x10 INFO: Slab 0xffffea00029aa470 objects=20 used=9 fp=0xffff8800be7830d8 flags=0x20000000004081 INFO: Object 0xffff8800be7830d8 @offset=4312 fp=0xffff8800be7827a8 Bytes b4 0xffff8800be7830c8: 87 a8 96 00 01 00 00 00 5a 5a 5a 5a 5a 5a 5a 5a .�......ZZZZZZZZ Object 0xffff8800be7830d8: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk Object 0xffff8800be7830e8: 6b 6b 6b 6b 01 08 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkk..kkkkkkkkkk Object 0xffff8800be7830f8: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk Object 0xffff8800be783108: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b a5 kkkkkkkkkkkkkkk� Redzone 0xffff8800be783118: bb bb bb bb bb bb bb bb ������������� Padding 0xffff8800be783258: 5a 5a 5a 5a 5a 5a 5a 5a ZZZZZZZZ Pid: 29049, comm: kworker/2:2 Not tainted 3.0.0-rc4+ #8 Call Trace: [<ffffffff811055c3>] print_trailer+0x131/0x13a [<ffffffff81105601>] object_err+0x35/0x3e [<ffffffff8110746f>] verify_mem_not_deleted+0x7a/0xb7 [<ffffffffa02851b5>] rpcb_getport_done+0x23/0x126 [sunrpc] [<ffffffffa027d0ba>] rpc_exit_task+0x3f/0x6d [sunrpc] [<ffffffffa027d4ab>] __rpc_execute+0x78/0x24b [sunrpc] [<ffffffffa027d6c0>] ? rpc_execute+0x42/0x42 [sunrpc] [<ffffffffa027d6d0>] rpc_async_schedule+0x10/0x12 [sunrpc] [<ffffffff810611b7>] process_one_work+0x230/0x41d [<ffffffff81061102>] ? process_one_work+0x17b/0x41d [<ffffffff81063613>] worker_thread+0x133/0x217 [<ffffffff810634e0>] ? manage_workers+0x191/0x191 [<ffffffff81066e10>] kthread+0x7d/0x85 [<ffffffff81485924>] kernel_thread_helper+0x4/0x10 [<ffffffff8147eb18>] ? retint_restore_args+0x13/0x13 [<ffffffff81066d93>] ? __init_kthread_worker+0x56/0x56 [<ffffffff81485920>] ? gs_change+0x13/0x13 Acked-by: Christoph Lameter <cl@linux.com> Signed-off-by: Ben Greear <greearb@candelatech.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-07-08 02:36:36 +08:00
#define TRACK_ADDRS_COUNT 16
struct track {
unsigned long addr; /* Called from address */
slub: Enable backtrace for create/delete points This patch attempts to grab a backtrace for the creation and deletion points of the slub object. When a fault is detected, we can then get a better idea of where the item was deleted. Example output from debugging some funky nfs/rpc behaviour: ============================================================================= BUG kmalloc-64: Object is on free-list ----------------------------------------------------------------------------- INFO: Allocated in rpcb_getport_async+0x39c/0x5a5 [sunrpc] age=381 cpu=3 pid=3750 __slab_alloc+0x348/0x3ba kmem_cache_alloc_trace+0x67/0xe7 rpcb_getport_async+0x39c/0x5a5 [sunrpc] call_bind+0x70/0x75 [sunrpc] __rpc_execute+0x78/0x24b [sunrpc] rpc_execute+0x3d/0x42 [sunrpc] rpc_run_task+0x79/0x81 [sunrpc] rpc_call_sync+0x3f/0x60 [sunrpc] rpc_ping+0x42/0x58 [sunrpc] rpc_create+0x4aa/0x527 [sunrpc] nfs_create_rpc_client+0xb1/0xf6 [nfs] nfs_init_client+0x3b/0x7d [nfs] nfs_get_client+0x453/0x5ab [nfs] nfs_create_server+0x10b/0x437 [nfs] nfs_fs_mount+0x4ca/0x708 [nfs] mount_fs+0x6b/0x152 INFO: Freed in rpcb_map_release+0x3f/0x44 [sunrpc] age=30 cpu=2 pid=29049 __slab_free+0x57/0x150 kfree+0x107/0x13a rpcb_map_release+0x3f/0x44 [sunrpc] rpc_release_calldata+0x12/0x14 [sunrpc] rpc_free_task+0x59/0x61 [sunrpc] rpc_final_put_task+0x82/0x8a [sunrpc] __rpc_execute+0x23c/0x24b [sunrpc] rpc_async_schedule+0x10/0x12 [sunrpc] process_one_work+0x230/0x41d worker_thread+0x133/0x217 kthread+0x7d/0x85 kernel_thread_helper+0x4/0x10 INFO: Slab 0xffffea00029aa470 objects=20 used=9 fp=0xffff8800be7830d8 flags=0x20000000004081 INFO: Object 0xffff8800be7830d8 @offset=4312 fp=0xffff8800be7827a8 Bytes b4 0xffff8800be7830c8: 87 a8 96 00 01 00 00 00 5a 5a 5a 5a 5a 5a 5a 5a .�......ZZZZZZZZ Object 0xffff8800be7830d8: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk Object 0xffff8800be7830e8: 6b 6b 6b 6b 01 08 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkk..kkkkkkkkkk Object 0xffff8800be7830f8: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk Object 0xffff8800be783108: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b a5 kkkkkkkkkkkkkkk� Redzone 0xffff8800be783118: bb bb bb bb bb bb bb bb ������������� Padding 0xffff8800be783258: 5a 5a 5a 5a 5a 5a 5a 5a ZZZZZZZZ Pid: 29049, comm: kworker/2:2 Not tainted 3.0.0-rc4+ #8 Call Trace: [<ffffffff811055c3>] print_trailer+0x131/0x13a [<ffffffff81105601>] object_err+0x35/0x3e [<ffffffff8110746f>] verify_mem_not_deleted+0x7a/0xb7 [<ffffffffa02851b5>] rpcb_getport_done+0x23/0x126 [sunrpc] [<ffffffffa027d0ba>] rpc_exit_task+0x3f/0x6d [sunrpc] [<ffffffffa027d4ab>] __rpc_execute+0x78/0x24b [sunrpc] [<ffffffffa027d6c0>] ? rpc_execute+0x42/0x42 [sunrpc] [<ffffffffa027d6d0>] rpc_async_schedule+0x10/0x12 [sunrpc] [<ffffffff810611b7>] process_one_work+0x230/0x41d [<ffffffff81061102>] ? process_one_work+0x17b/0x41d [<ffffffff81063613>] worker_thread+0x133/0x217 [<ffffffff810634e0>] ? manage_workers+0x191/0x191 [<ffffffff81066e10>] kthread+0x7d/0x85 [<ffffffff81485924>] kernel_thread_helper+0x4/0x10 [<ffffffff8147eb18>] ? retint_restore_args+0x13/0x13 [<ffffffff81066d93>] ? __init_kthread_worker+0x56/0x56 [<ffffffff81485920>] ? gs_change+0x13/0x13 Acked-by: Christoph Lameter <cl@linux.com> Signed-off-by: Ben Greear <greearb@candelatech.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-07-08 02:36:36 +08:00
#ifdef CONFIG_STACKTRACE
unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
#endif
int cpu; /* Was running on cpu */
int pid; /* Pid context */
unsigned long when; /* When did the operation occur */
};
enum track_item { TRACK_ALLOC, TRACK_FREE };
#ifdef CONFIG_SYSFS
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
static int sysfs_slab_add(struct kmem_cache *);
static int sysfs_slab_alias(struct kmem_cache *, const char *);
#else
static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
{ return 0; }
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
#endif
static inline void stat(const struct kmem_cache *s, enum stat_item si)
SLUB: Support for performance statistics The statistics provided here allow the monitoring of allocator behavior but at the cost of some (minimal) loss of performance. Counters are placed in SLUB's per cpu data structure. The per cpu structure may be extended by the statistics to grow larger than one cacheline which will increase the cache footprint of SLUB. There is a compile option to enable/disable the inclusion of the runtime statistics and its off by default. The slabinfo tool is enhanced to support these statistics via two options: -D Switches the line of information displayed for a slab from size mode to activity mode. -A Sorts the slabs displayed by activity. This allows the display of the slabs most important to the performance of a certain load. -r Report option will report detailed statistics on Example (tbench load): slabinfo -AD ->Shows the most active slabs Name Objects Alloc Free %Fast skbuff_fclone_cache 33 111953835 111953835 99 99 :0000192 2666 5283688 5281047 99 99 :0001024 849 5247230 5246389 83 83 vm_area_struct 1349 119642 118355 91 22 :0004096 15 66753 66751 98 98 :0000064 2067 25297 23383 98 78 dentry 10259 28635 18464 91 45 :0000080 11004 18950 8089 98 98 :0000096 1703 12358 10784 99 98 :0000128 762 10582 9875 94 18 :0000512 184 9807 9647 95 81 :0002048 479 9669 9195 83 65 anon_vma 777 9461 9002 99 71 kmalloc-8 6492 9981 5624 99 97 :0000768 258 7174 6931 58 15 So the skbuff_fclone_cache is of highest importance for the tbench load. Pretty high load on the 192 sized slab. Look for the aliases slabinfo -a | grep 000192 :0000192 <- xfs_btree_cur filp kmalloc-192 uid_cache tw_sock_TCP request_sock_TCPv6 tw_sock_TCPv6 skbuff_head_cache xfs_ili Likely skbuff_head_cache. Looking into the statistics of the skbuff_fclone_cache is possible through slabinfo skbuff_fclone_cache ->-r option implied if cache name is mentioned .... Usual output ... Slab Perf Counter Alloc Free %Al %Fr -------------------------------------------------- Fastpath 111953360 111946981 99 99 Slowpath 1044 7423 0 0 Page Alloc 272 264 0 0 Add partial 25 325 0 0 Remove partial 86 264 0 0 RemoteObj/SlabFrozen 350 4832 0 0 Total 111954404 111954404 Flushes 49 Refill 0 Deactivate Full=325(92%) Empty=0(0%) ToHead=24(6%) ToTail=1(0%) Looks good because the fastpath is overwhelmingly taken. skbuff_head_cache: Slab Perf Counter Alloc Free %Al %Fr -------------------------------------------------- Fastpath 5297262 5259882 99 99 Slowpath 4477 39586 0 0 Page Alloc 937 824 0 0 Add partial 0 2515 0 0 Remove partial 1691 824 0 0 RemoteObj/SlabFrozen 2621 9684 0 0 Total 5301739 5299468 Deactivate Full=2620(100%) Empty=0(0%) ToHead=0(0%) ToTail=0(0%) Descriptions of the output: Total: The total number of allocation and frees that occurred for a slab Fastpath: The number of allocations/frees that used the fastpath. Slowpath: Other allocations Page Alloc: Number of calls to the page allocator as a result of slowpath processing Add Partial: Number of slabs added to the partial list through free or alloc (occurs during cpuslab flushes) Remove Partial: Number of slabs removed from the partial list as a result of allocations retrieving a partial slab or by a free freeing the last object of a slab. RemoteObj/Froz: How many times were remotely freed object encountered when a slab was about to be deactivated. Frozen: How many times was free able to skip list processing because the slab was in use as the cpuslab of another processor. Flushes: Number of times the cpuslab was flushed on request (kmem_cache_shrink, may result from races in __slab_alloc) Refill: Number of times we were able to refill the cpuslab from remotely freed objects for the same slab. Deactivate: Statistics how slabs were deactivated. Shows how they were put onto the partial list. In general fastpath is very good. Slowpath without partial list processing is also desirable. Any touching of partial list uses node specific locks which may potentially cause list lock contention. Signed-off-by: Christoph Lameter <clameter@sgi.com>
2008-02-08 09:47:41 +08:00
{
#ifdef CONFIG_SLUB_STATS
/*
* The rmw is racy on a preemptible kernel but this is acceptable, so
* avoid this_cpu_add()'s irq-disable overhead.
*/
raw_cpu_inc(s->cpu_slab->stat[si]);
SLUB: Support for performance statistics The statistics provided here allow the monitoring of allocator behavior but at the cost of some (minimal) loss of performance. Counters are placed in SLUB's per cpu data structure. The per cpu structure may be extended by the statistics to grow larger than one cacheline which will increase the cache footprint of SLUB. There is a compile option to enable/disable the inclusion of the runtime statistics and its off by default. The slabinfo tool is enhanced to support these statistics via two options: -D Switches the line of information displayed for a slab from size mode to activity mode. -A Sorts the slabs displayed by activity. This allows the display of the slabs most important to the performance of a certain load. -r Report option will report detailed statistics on Example (tbench load): slabinfo -AD ->Shows the most active slabs Name Objects Alloc Free %Fast skbuff_fclone_cache 33 111953835 111953835 99 99 :0000192 2666 5283688 5281047 99 99 :0001024 849 5247230 5246389 83 83 vm_area_struct 1349 119642 118355 91 22 :0004096 15 66753 66751 98 98 :0000064 2067 25297 23383 98 78 dentry 10259 28635 18464 91 45 :0000080 11004 18950 8089 98 98 :0000096 1703 12358 10784 99 98 :0000128 762 10582 9875 94 18 :0000512 184 9807 9647 95 81 :0002048 479 9669 9195 83 65 anon_vma 777 9461 9002 99 71 kmalloc-8 6492 9981 5624 99 97 :0000768 258 7174 6931 58 15 So the skbuff_fclone_cache is of highest importance for the tbench load. Pretty high load on the 192 sized slab. Look for the aliases slabinfo -a | grep 000192 :0000192 <- xfs_btree_cur filp kmalloc-192 uid_cache tw_sock_TCP request_sock_TCPv6 tw_sock_TCPv6 skbuff_head_cache xfs_ili Likely skbuff_head_cache. Looking into the statistics of the skbuff_fclone_cache is possible through slabinfo skbuff_fclone_cache ->-r option implied if cache name is mentioned .... Usual output ... Slab Perf Counter Alloc Free %Al %Fr -------------------------------------------------- Fastpath 111953360 111946981 99 99 Slowpath 1044 7423 0 0 Page Alloc 272 264 0 0 Add partial 25 325 0 0 Remove partial 86 264 0 0 RemoteObj/SlabFrozen 350 4832 0 0 Total 111954404 111954404 Flushes 49 Refill 0 Deactivate Full=325(92%) Empty=0(0%) ToHead=24(6%) ToTail=1(0%) Looks good because the fastpath is overwhelmingly taken. skbuff_head_cache: Slab Perf Counter Alloc Free %Al %Fr -------------------------------------------------- Fastpath 5297262 5259882 99 99 Slowpath 4477 39586 0 0 Page Alloc 937 824 0 0 Add partial 0 2515 0 0 Remove partial 1691 824 0 0 RemoteObj/SlabFrozen 2621 9684 0 0 Total 5301739 5299468 Deactivate Full=2620(100%) Empty=0(0%) ToHead=0(0%) ToTail=0(0%) Descriptions of the output: Total: The total number of allocation and frees that occurred for a slab Fastpath: The number of allocations/frees that used the fastpath. Slowpath: Other allocations Page Alloc: Number of calls to the page allocator as a result of slowpath processing Add Partial: Number of slabs added to the partial list through free or alloc (occurs during cpuslab flushes) Remove Partial: Number of slabs removed from the partial list as a result of allocations retrieving a partial slab or by a free freeing the last object of a slab. RemoteObj/Froz: How many times were remotely freed object encountered when a slab was about to be deactivated. Frozen: How many times was free able to skip list processing because the slab was in use as the cpuslab of another processor. Flushes: Number of times the cpuslab was flushed on request (kmem_cache_shrink, may result from races in __slab_alloc) Refill: Number of times we were able to refill the cpuslab from remotely freed objects for the same slab. Deactivate: Statistics how slabs were deactivated. Shows how they were put onto the partial list. In general fastpath is very good. Slowpath without partial list processing is also desirable. Any touching of partial list uses node specific locks which may potentially cause list lock contention. Signed-off-by: Christoph Lameter <clameter@sgi.com>
2008-02-08 09:47:41 +08:00
#endif
}
mm, slab, slub: stop taking memory hotplug lock Since commit 03afc0e25f7f ("slab: get_online_mems for kmem_cache_{create,destroy,shrink}") we are taking memory hotplug lock for SLAB and SLUB when creating, destroying or shrinking a cache. It is quite a heavy lock and it's best to avoid it if possible, as we had several issues with lockdep complaining about ordering in the past, see e.g. e4f8e513c3d3 ("mm/slub: fix a deadlock in show_slab_objects()"). The problem scenario in 03afc0e25f7f (solved by the memory hotplug lock) can be summarized as follows: while there's slab_mutex synchronizing new kmem cache creation and SLUB's MEM_GOING_ONLINE callback slab_mem_going_online_callback(), we may miss creation of kmem_cache_node for the hotplugged node in the new kmem cache, because the hotplug callback doesn't yet see the new cache, and cache creation in init_kmem_cache_nodes() only inits kmem_cache_node for nodes in the N_NORMAL_MEMORY nodemask, which however may not yet include the new node, as that happens only later after the MEM_GOING_ONLINE callback. Instead of using get/put_online_mems(), the problem can be solved by SLUB maintaining its own nodemask of nodes for which it has allocated the per-node kmem_cache_node structures. This nodemask would generally mirror the N_NORMAL_MEMORY nodemask, but would be updated only in under SLUB's control in its memory hotplug callbacks under the slab_mutex. This patch adds such nodemask and its handling. Commit 03afc0e25f7f mentiones "issues like [the one above]", but there don't appear to be further issues. All the paths (shared for SLAB and SLUB) taking the memory hotplug locks are also taking the slab_mutex, except kmem_cache_shrink() where 03afc0e25f7f replaced slab_mutex with get/put_online_mems(). We however cannot simply restore slab_mutex in kmem_cache_shrink(), as SLUB can enters the function from a write to sysfs 'shrink' file, thus holding kernfs lock, and in kmem_cache_create() the kernfs lock is nested within slab_mutex. But on closer inspection we don't actually need to protect kmem_cache_shrink() from hotplug callbacks: While SLUB's __kmem_cache_shrink() does for_each_kmem_cache_node(), missing a new node added in parallel hotplug is not fatal, and parallel hotremove does not free kmem_cache_node's anymore after the previous patch, so use-after free cannot happen. The per-node shrinking itself is protected by n->list_lock. Same is true for SLAB, and SLOB is no-op. SLAB also doesn't need the memory hotplug locking, which it only gained by 03afc0e25f7f through the shared paths in slab_common.c. Its memory hotplug callbacks are also protected by slab_mutex against races with these paths. The problem of SLUB relying on N_NORMAL_MEMORY doesn't apply to SLAB, as its setup_kmem_cache_nodes relies on N_ONLINE, and the new node is already set there during the MEM_GOING_ONLINE callback, so no special care is needed for SLAB. As such, this patch removes all get/put_online_mems() usage by the slab subsystem. Link: https://lkml.kernel.org/r/20210113131634.3671-3-vbabka@suse.cz Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Cc: Christoph Lameter <cl@linux.com> Cc: David Hildenbrand <david@redhat.com> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: Qian Cai <cai@redhat.com> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-25 04:01:12 +08:00
/*
* Tracks for which NUMA nodes we have kmem_cache_nodes allocated.
* Corresponds to node_state[N_NORMAL_MEMORY], but can temporarily
* differ during memory hotplug/hotremove operations.
* Protected by slab_mutex.
*/
static nodemask_t slab_nodes;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/********************************************************************
* Core slab cache functions
*******************************************************************/
mm: add SLUB free list pointer obfuscation This SLUB free list pointer obfuscation code is modified from Brad Spengler/PaX Team's code in the last public patch of grsecurity/PaX based on my understanding of the code. Changes or omissions from the original code are mine and don't reflect the original grsecurity/PaX code. This adds a per-cache random value to SLUB caches that is XORed with their freelist pointer address and value. This adds nearly zero overhead and frustrates the very common heap overflow exploitation method of overwriting freelist pointers. A recent example of the attack is written up here: http://cyseclabs.com/blog/cve-2016-6187-heap-off-by-one-exploit and there is a section dedicated to the technique the book "A Guide to Kernel Exploitation: Attacking the Core". This is based on patches by Daniel Micay, and refactored to minimize the use of #ifdef. With 200-count cycles of "hackbench -g 20 -l 1000" I saw the following run times: before: mean 10.11882499999999999995 variance .03320378329145728642 stdev .18221905304181911048 after: mean 10.12654000000000000014 variance .04700556623115577889 stdev .21680767106160192064 The difference gets lost in the noise, but if the above is to be taken literally, using CONFIG_FREELIST_HARDENED is 0.07% slower. Link: http://lkml.kernel.org/r/20170802180609.GA66807@beast Signed-off-by: Kees Cook <keescook@chromium.org> Suggested-by: Daniel Micay <danielmicay@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Tycho Andersen <tycho@docker.com> Cc: Alexander Popov <alex.popov@linux.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:19:18 +08:00
/*
* Returns freelist pointer (ptr). With hardening, this is obfuscated
* with an XOR of the address where the pointer is held and a per-cache
* random number.
*/
static inline void *freelist_ptr(const struct kmem_cache *s, void *ptr,
unsigned long ptr_addr)
{
#ifdef CONFIG_SLAB_FREELIST_HARDENED
/*
kasan, mm: reset tags when accessing metadata Kernel allocator code accesses metadata for slab objects, that may lie out-of-bounds of the object itself, or be accessed when an object is freed. Such accesses trigger tag faults and lead to false-positive reports with hardware tag-based KASAN. Software KASAN modes disable instrumentation for allocator code via KASAN_SANITIZE Makefile macro, and rely on kasan_enable/disable_current() annotations which are used to ignore KASAN reports. With hardware tag-based KASAN neither of those options are available, as it doesn't use compiler instrumetation, no tag faults are ignored, and MTE is disabled after the first one. Instead, reset tags when accessing metadata (currently only for SLUB). Link: https://lkml.kernel.org/r/a0f3cefbc49f34c843b664110842de4db28179d0.1606161801.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Signed-off-by: Vincenzo Frascino <vincenzo.frascino@arm.com> Acked-by: Marco Elver <elver@google.com> Reviewed-by: Alexander Potapenko <glider@google.com> Tested-by: Vincenzo Frascino <vincenzo.frascino@arm.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Branislav Rankov <Branislav.Rankov@arm.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Evgenii Stepanov <eugenis@google.com> Cc: Kevin Brodsky <kevin.brodsky@arm.com> Cc: Vasily Gorbik <gor@linux.ibm.com> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-23 04:02:17 +08:00
* When CONFIG_KASAN_SW/HW_TAGS is enabled, ptr_addr might be tagged.
* Normally, this doesn't cause any issues, as both set_freepointer()
* and get_freepointer() are called with a pointer with the same tag.
* However, there are some issues with CONFIG_SLUB_DEBUG code. For
* example, when __free_slub() iterates over objects in a cache, it
* passes untagged pointers to check_object(). check_object() in turns
* calls get_freepointer() with an untagged pointer, which causes the
* freepointer to be restored incorrectly.
*/
return (void *)((unsigned long)ptr ^ s->random ^
slub: improve bit diffusion for freelist ptr obfuscation Under CONFIG_SLAB_FREELIST_HARDENED=y, the obfuscation was relatively weak in that the ptr and ptr address were usually so close that the first XOR would result in an almost entirely 0-byte value[1], leaving most of the "secret" number ultimately being stored after the third XOR. A single blind memory content exposure of the freelist was generally sufficient to learn the secret. Add a swab() call to mix bits a little more. This is a cheap way (1 cycle) to make attacks need more than a single exposure to learn the secret (or to know _where_ the exposure is in memory). kmalloc-32 freelist walk, before: ptr ptr_addr stored value secret ffff90c22e019020@ffff90c22e019000 is 86528eb656b3b5bd (86528eb656b3b59d) ffff90c22e019040@ffff90c22e019020 is 86528eb656b3b5fd (86528eb656b3b59d) ffff90c22e019060@ffff90c22e019040 is 86528eb656b3b5bd (86528eb656b3b59d) ffff90c22e019080@ffff90c22e019060 is 86528eb656b3b57d (86528eb656b3b59d) ffff90c22e0190a0@ffff90c22e019080 is 86528eb656b3b5bd (86528eb656b3b59d) ... after: ptr ptr_addr stored value secret ffff9eed6e019020@ffff9eed6e019000 is 793d1135d52cda42 (86528eb656b3b59d) ffff9eed6e019040@ffff9eed6e019020 is 593d1135d52cda22 (86528eb656b3b59d) ffff9eed6e019060@ffff9eed6e019040 is 393d1135d52cda02 (86528eb656b3b59d) ffff9eed6e019080@ffff9eed6e019060 is 193d1135d52cdae2 (86528eb656b3b59d) ffff9eed6e0190a0@ffff9eed6e019080 is f93d1135d52cdac2 (86528eb656b3b59d) [1] https://blog.infosectcbr.com.au/2020/03/weaknesses-in-linux-kernel-heap.html Fixes: 2482ddec670f ("mm: add SLUB free list pointer obfuscation") Reported-by: Silvio Cesare <silvio.cesare@gmail.com> Signed-off-by: Kees Cook <keescook@chromium.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: <stable@vger.kernel.org> Link: http://lkml.kernel.org/r/202003051623.AF4F8CB@keescook Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 12:04:23 +08:00
swab((unsigned long)kasan_reset_tag((void *)ptr_addr)));
mm: add SLUB free list pointer obfuscation This SLUB free list pointer obfuscation code is modified from Brad Spengler/PaX Team's code in the last public patch of grsecurity/PaX based on my understanding of the code. Changes or omissions from the original code are mine and don't reflect the original grsecurity/PaX code. This adds a per-cache random value to SLUB caches that is XORed with their freelist pointer address and value. This adds nearly zero overhead and frustrates the very common heap overflow exploitation method of overwriting freelist pointers. A recent example of the attack is written up here: http://cyseclabs.com/blog/cve-2016-6187-heap-off-by-one-exploit and there is a section dedicated to the technique the book "A Guide to Kernel Exploitation: Attacking the Core". This is based on patches by Daniel Micay, and refactored to minimize the use of #ifdef. With 200-count cycles of "hackbench -g 20 -l 1000" I saw the following run times: before: mean 10.11882499999999999995 variance .03320378329145728642 stdev .18221905304181911048 after: mean 10.12654000000000000014 variance .04700556623115577889 stdev .21680767106160192064 The difference gets lost in the noise, but if the above is to be taken literally, using CONFIG_FREELIST_HARDENED is 0.07% slower. Link: http://lkml.kernel.org/r/20170802180609.GA66807@beast Signed-off-by: Kees Cook <keescook@chromium.org> Suggested-by: Daniel Micay <danielmicay@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Tycho Andersen <tycho@docker.com> Cc: Alexander Popov <alex.popov@linux.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:19:18 +08:00
#else
return ptr;
#endif
}
/* Returns the freelist pointer recorded at location ptr_addr. */
static inline void *freelist_dereference(const struct kmem_cache *s,
void *ptr_addr)
{
return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr),
(unsigned long)ptr_addr);
}
static inline void *get_freepointer(struct kmem_cache *s, void *object)
{
kasan, mm: reset tags when accessing metadata Kernel allocator code accesses metadata for slab objects, that may lie out-of-bounds of the object itself, or be accessed when an object is freed. Such accesses trigger tag faults and lead to false-positive reports with hardware tag-based KASAN. Software KASAN modes disable instrumentation for allocator code via KASAN_SANITIZE Makefile macro, and rely on kasan_enable/disable_current() annotations which are used to ignore KASAN reports. With hardware tag-based KASAN neither of those options are available, as it doesn't use compiler instrumetation, no tag faults are ignored, and MTE is disabled after the first one. Instead, reset tags when accessing metadata (currently only for SLUB). Link: https://lkml.kernel.org/r/a0f3cefbc49f34c843b664110842de4db28179d0.1606161801.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Signed-off-by: Vincenzo Frascino <vincenzo.frascino@arm.com> Acked-by: Marco Elver <elver@google.com> Reviewed-by: Alexander Potapenko <glider@google.com> Tested-by: Vincenzo Frascino <vincenzo.frascino@arm.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Branislav Rankov <Branislav.Rankov@arm.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Evgenii Stepanov <eugenis@google.com> Cc: Kevin Brodsky <kevin.brodsky@arm.com> Cc: Vasily Gorbik <gor@linux.ibm.com> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-23 04:02:17 +08:00
object = kasan_reset_tag(object);
mm: add SLUB free list pointer obfuscation This SLUB free list pointer obfuscation code is modified from Brad Spengler/PaX Team's code in the last public patch of grsecurity/PaX based on my understanding of the code. Changes or omissions from the original code are mine and don't reflect the original grsecurity/PaX code. This adds a per-cache random value to SLUB caches that is XORed with their freelist pointer address and value. This adds nearly zero overhead and frustrates the very common heap overflow exploitation method of overwriting freelist pointers. A recent example of the attack is written up here: http://cyseclabs.com/blog/cve-2016-6187-heap-off-by-one-exploit and there is a section dedicated to the technique the book "A Guide to Kernel Exploitation: Attacking the Core". This is based on patches by Daniel Micay, and refactored to minimize the use of #ifdef. With 200-count cycles of "hackbench -g 20 -l 1000" I saw the following run times: before: mean 10.11882499999999999995 variance .03320378329145728642 stdev .18221905304181911048 after: mean 10.12654000000000000014 variance .04700556623115577889 stdev .21680767106160192064 The difference gets lost in the noise, but if the above is to be taken literally, using CONFIG_FREELIST_HARDENED is 0.07% slower. Link: http://lkml.kernel.org/r/20170802180609.GA66807@beast Signed-off-by: Kees Cook <keescook@chromium.org> Suggested-by: Daniel Micay <danielmicay@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Tycho Andersen <tycho@docker.com> Cc: Alexander Popov <alex.popov@linux.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:19:18 +08:00
return freelist_dereference(s, object + s->offset);
}
slub: prefetch next freelist pointer in slab_alloc() Recycling a page is a problem, since freelist link chain is hot on cpu(s) which freed objects, and possibly very cold on cpu currently owning slab. Adding a prefetch of cache line containing the pointer to next object in slab_alloc() helps a lot in many workloads, in particular on assymetric ones (allocations done on one cpu, frees on another cpus). Added cost is three machine instructions only. Examples on my dual socket quad core ht machine (Intel CPU E5540 @2.53GHz) (16 logical cpus, 2 memory nodes), 64bit kernel. Before patch : # perf stat -r 32 hackbench 50 process 4000 >/dev/null Performance counter stats for 'hackbench 50 process 4000' (32 runs): 327577,471718 task-clock # 15,821 CPUs utilized ( +- 0,64% ) 28 866 491 context-switches # 0,088 M/sec ( +- 1,80% ) 1 506 929 CPU-migrations # 0,005 M/sec ( +- 3,24% ) 127 151 page-faults # 0,000 M/sec ( +- 0,16% ) 829 399 813 448 cycles # 2,532 GHz ( +- 0,64% ) 580 664 691 740 stalled-cycles-frontend # 70,01% frontend cycles idle ( +- 0,71% ) 197 431 700 448 stalled-cycles-backend # 23,80% backend cycles idle ( +- 1,03% ) 503 548 648 975 instructions # 0,61 insns per cycle # 1,15 stalled cycles per insn ( +- 0,46% ) 95 780 068 471 branches # 292,389 M/sec ( +- 0,48% ) 1 426 407 916 branch-misses # 1,49% of all branches ( +- 1,35% ) 20,705679994 seconds time elapsed ( +- 0,64% ) After patch : # perf stat -r 32 hackbench 50 process 4000 >/dev/null Performance counter stats for 'hackbench 50 process 4000' (32 runs): 286236,542804 task-clock # 15,786 CPUs utilized ( +- 1,32% ) 19 703 372 context-switches # 0,069 M/sec ( +- 4,99% ) 1 658 249 CPU-migrations # 0,006 M/sec ( +- 6,62% ) 126 776 page-faults # 0,000 M/sec ( +- 0,12% ) 724 636 593 213 cycles # 2,532 GHz ( +- 1,32% ) 499 320 714 837 stalled-cycles-frontend # 68,91% frontend cycles idle ( +- 1,47% ) 156 555 126 809 stalled-cycles-backend # 21,60% backend cycles idle ( +- 2,22% ) 463 897 792 661 instructions # 0,64 insns per cycle # 1,08 stalled cycles per insn ( +- 0,94% ) 87 717 352 563 branches # 306,451 M/sec ( +- 0,99% ) 941 738 280 branch-misses # 1,07% of all branches ( +- 3,35% ) 18,132070670 seconds time elapsed ( +- 1,30% ) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Acked-by: Christoph Lameter <cl@linux.com> CC: Matt Mackall <mpm@selenic.com> CC: David Rientjes <rientjes@google.com> CC: "Alex,Shi" <alex.shi@intel.com> CC: Shaohua Li <shaohua.li@intel.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-12-16 23:25:34 +08:00
static void prefetch_freepointer(const struct kmem_cache *s, void *object)
{
prefetch(object + s->offset);
slub: prefetch next freelist pointer in slab_alloc() Recycling a page is a problem, since freelist link chain is hot on cpu(s) which freed objects, and possibly very cold on cpu currently owning slab. Adding a prefetch of cache line containing the pointer to next object in slab_alloc() helps a lot in many workloads, in particular on assymetric ones (allocations done on one cpu, frees on another cpus). Added cost is three machine instructions only. Examples on my dual socket quad core ht machine (Intel CPU E5540 @2.53GHz) (16 logical cpus, 2 memory nodes), 64bit kernel. Before patch : # perf stat -r 32 hackbench 50 process 4000 >/dev/null Performance counter stats for 'hackbench 50 process 4000' (32 runs): 327577,471718 task-clock # 15,821 CPUs utilized ( +- 0,64% ) 28 866 491 context-switches # 0,088 M/sec ( +- 1,80% ) 1 506 929 CPU-migrations # 0,005 M/sec ( +- 3,24% ) 127 151 page-faults # 0,000 M/sec ( +- 0,16% ) 829 399 813 448 cycles # 2,532 GHz ( +- 0,64% ) 580 664 691 740 stalled-cycles-frontend # 70,01% frontend cycles idle ( +- 0,71% ) 197 431 700 448 stalled-cycles-backend # 23,80% backend cycles idle ( +- 1,03% ) 503 548 648 975 instructions # 0,61 insns per cycle # 1,15 stalled cycles per insn ( +- 0,46% ) 95 780 068 471 branches # 292,389 M/sec ( +- 0,48% ) 1 426 407 916 branch-misses # 1,49% of all branches ( +- 1,35% ) 20,705679994 seconds time elapsed ( +- 0,64% ) After patch : # perf stat -r 32 hackbench 50 process 4000 >/dev/null Performance counter stats for 'hackbench 50 process 4000' (32 runs): 286236,542804 task-clock # 15,786 CPUs utilized ( +- 1,32% ) 19 703 372 context-switches # 0,069 M/sec ( +- 4,99% ) 1 658 249 CPU-migrations # 0,006 M/sec ( +- 6,62% ) 126 776 page-faults # 0,000 M/sec ( +- 0,12% ) 724 636 593 213 cycles # 2,532 GHz ( +- 1,32% ) 499 320 714 837 stalled-cycles-frontend # 68,91% frontend cycles idle ( +- 1,47% ) 156 555 126 809 stalled-cycles-backend # 21,60% backend cycles idle ( +- 2,22% ) 463 897 792 661 instructions # 0,64 insns per cycle # 1,08 stalled cycles per insn ( +- 0,94% ) 87 717 352 563 branches # 306,451 M/sec ( +- 0,99% ) 941 738 280 branch-misses # 1,07% of all branches ( +- 3,35% ) 18,132070670 seconds time elapsed ( +- 1,30% ) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Acked-by: Christoph Lameter <cl@linux.com> CC: Matt Mackall <mpm@selenic.com> CC: David Rientjes <rientjes@google.com> CC: "Alex,Shi" <alex.shi@intel.com> CC: Shaohua Li <shaohua.li@intel.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-12-16 23:25:34 +08:00
}
static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
{
mm: add SLUB free list pointer obfuscation This SLUB free list pointer obfuscation code is modified from Brad Spengler/PaX Team's code in the last public patch of grsecurity/PaX based on my understanding of the code. Changes or omissions from the original code are mine and don't reflect the original grsecurity/PaX code. This adds a per-cache random value to SLUB caches that is XORed with their freelist pointer address and value. This adds nearly zero overhead and frustrates the very common heap overflow exploitation method of overwriting freelist pointers. A recent example of the attack is written up here: http://cyseclabs.com/blog/cve-2016-6187-heap-off-by-one-exploit and there is a section dedicated to the technique the book "A Guide to Kernel Exploitation: Attacking the Core". This is based on patches by Daniel Micay, and refactored to minimize the use of #ifdef. With 200-count cycles of "hackbench -g 20 -l 1000" I saw the following run times: before: mean 10.11882499999999999995 variance .03320378329145728642 stdev .18221905304181911048 after: mean 10.12654000000000000014 variance .04700556623115577889 stdev .21680767106160192064 The difference gets lost in the noise, but if the above is to be taken literally, using CONFIG_FREELIST_HARDENED is 0.07% slower. Link: http://lkml.kernel.org/r/20170802180609.GA66807@beast Signed-off-by: Kees Cook <keescook@chromium.org> Suggested-by: Daniel Micay <danielmicay@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Tycho Andersen <tycho@docker.com> Cc: Alexander Popov <alex.popov@linux.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:19:18 +08:00
unsigned long freepointer_addr;
void *p;
mm, debug_pagealloc: don't rely on static keys too early Commit 96a2b03f281d ("mm, debug_pagelloc: use static keys to enable debugging") has introduced a static key to reduce overhead when debug_pagealloc is compiled in but not enabled. It relied on the assumption that jump_label_init() is called before parse_early_param() as in start_kernel(), so when the "debug_pagealloc=on" option is parsed, it is safe to enable the static key. However, it turns out multiple architectures call parse_early_param() earlier from their setup_arch(). x86 also calls jump_label_init() even earlier, so no issue was found while testing the commit, but same is not true for e.g. ppc64 and s390 where the kernel would not boot with debug_pagealloc=on as found by our QA. To fix this without tricky changes to init code of multiple architectures, this patch partially reverts the static key conversion from 96a2b03f281d. Init-time and non-fastpath calls (such as in arch code) of debug_pagealloc_enabled() will again test a simple bool variable. Fastpath mm code is converted to a new debug_pagealloc_enabled_static() variant that relies on the static key, which is enabled in a well-defined point in mm_init() where it's guaranteed that jump_label_init() has been called, regardless of architecture. [sfr@canb.auug.org.au: export _debug_pagealloc_enabled_early] Link: http://lkml.kernel.org/r/20200106164944.063ac07b@canb.auug.org.au Link: http://lkml.kernel.org/r/20191219130612.23171-1-vbabka@suse.cz Fixes: 96a2b03f281d ("mm, debug_pagelloc: use static keys to enable debugging") Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.au> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Matthew Wilcox <willy@infradead.org> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Qian Cai <cai@lca.pw> Cc: <stable@vger.kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-01-14 08:29:20 +08:00
if (!debug_pagealloc_enabled_static())
return get_freepointer(s, object);
object = kasan_reset_tag(object);
mm: add SLUB free list pointer obfuscation This SLUB free list pointer obfuscation code is modified from Brad Spengler/PaX Team's code in the last public patch of grsecurity/PaX based on my understanding of the code. Changes or omissions from the original code are mine and don't reflect the original grsecurity/PaX code. This adds a per-cache random value to SLUB caches that is XORed with their freelist pointer address and value. This adds nearly zero overhead and frustrates the very common heap overflow exploitation method of overwriting freelist pointers. A recent example of the attack is written up here: http://cyseclabs.com/blog/cve-2016-6187-heap-off-by-one-exploit and there is a section dedicated to the technique the book "A Guide to Kernel Exploitation: Attacking the Core". This is based on patches by Daniel Micay, and refactored to minimize the use of #ifdef. With 200-count cycles of "hackbench -g 20 -l 1000" I saw the following run times: before: mean 10.11882499999999999995 variance .03320378329145728642 stdev .18221905304181911048 after: mean 10.12654000000000000014 variance .04700556623115577889 stdev .21680767106160192064 The difference gets lost in the noise, but if the above is to be taken literally, using CONFIG_FREELIST_HARDENED is 0.07% slower. Link: http://lkml.kernel.org/r/20170802180609.GA66807@beast Signed-off-by: Kees Cook <keescook@chromium.org> Suggested-by: Daniel Micay <danielmicay@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Tycho Andersen <tycho@docker.com> Cc: Alexander Popov <alex.popov@linux.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:19:18 +08:00
freepointer_addr = (unsigned long)object + s->offset;
copy_from_kernel_nofault(&p, (void **)freepointer_addr, sizeof(p));
mm: add SLUB free list pointer obfuscation This SLUB free list pointer obfuscation code is modified from Brad Spengler/PaX Team's code in the last public patch of grsecurity/PaX based on my understanding of the code. Changes or omissions from the original code are mine and don't reflect the original grsecurity/PaX code. This adds a per-cache random value to SLUB caches that is XORed with their freelist pointer address and value. This adds nearly zero overhead and frustrates the very common heap overflow exploitation method of overwriting freelist pointers. A recent example of the attack is written up here: http://cyseclabs.com/blog/cve-2016-6187-heap-off-by-one-exploit and there is a section dedicated to the technique the book "A Guide to Kernel Exploitation: Attacking the Core". This is based on patches by Daniel Micay, and refactored to minimize the use of #ifdef. With 200-count cycles of "hackbench -g 20 -l 1000" I saw the following run times: before: mean 10.11882499999999999995 variance .03320378329145728642 stdev .18221905304181911048 after: mean 10.12654000000000000014 variance .04700556623115577889 stdev .21680767106160192064 The difference gets lost in the noise, but if the above is to be taken literally, using CONFIG_FREELIST_HARDENED is 0.07% slower. Link: http://lkml.kernel.org/r/20170802180609.GA66807@beast Signed-off-by: Kees Cook <keescook@chromium.org> Suggested-by: Daniel Micay <danielmicay@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Tycho Andersen <tycho@docker.com> Cc: Alexander Popov <alex.popov@linux.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:19:18 +08:00
return freelist_ptr(s, p, freepointer_addr);
}
static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
{
mm: add SLUB free list pointer obfuscation This SLUB free list pointer obfuscation code is modified from Brad Spengler/PaX Team's code in the last public patch of grsecurity/PaX based on my understanding of the code. Changes or omissions from the original code are mine and don't reflect the original grsecurity/PaX code. This adds a per-cache random value to SLUB caches that is XORed with their freelist pointer address and value. This adds nearly zero overhead and frustrates the very common heap overflow exploitation method of overwriting freelist pointers. A recent example of the attack is written up here: http://cyseclabs.com/blog/cve-2016-6187-heap-off-by-one-exploit and there is a section dedicated to the technique the book "A Guide to Kernel Exploitation: Attacking the Core". This is based on patches by Daniel Micay, and refactored to minimize the use of #ifdef. With 200-count cycles of "hackbench -g 20 -l 1000" I saw the following run times: before: mean 10.11882499999999999995 variance .03320378329145728642 stdev .18221905304181911048 after: mean 10.12654000000000000014 variance .04700556623115577889 stdev .21680767106160192064 The difference gets lost in the noise, but if the above is to be taken literally, using CONFIG_FREELIST_HARDENED is 0.07% slower. Link: http://lkml.kernel.org/r/20170802180609.GA66807@beast Signed-off-by: Kees Cook <keescook@chromium.org> Suggested-by: Daniel Micay <danielmicay@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Tycho Andersen <tycho@docker.com> Cc: Alexander Popov <alex.popov@linux.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:19:18 +08:00
unsigned long freeptr_addr = (unsigned long)object + s->offset;
#ifdef CONFIG_SLAB_FREELIST_HARDENED
BUG_ON(object == fp); /* naive detection of double free or corruption */
#endif
kasan, mm: reset tags when accessing metadata Kernel allocator code accesses metadata for slab objects, that may lie out-of-bounds of the object itself, or be accessed when an object is freed. Such accesses trigger tag faults and lead to false-positive reports with hardware tag-based KASAN. Software KASAN modes disable instrumentation for allocator code via KASAN_SANITIZE Makefile macro, and rely on kasan_enable/disable_current() annotations which are used to ignore KASAN reports. With hardware tag-based KASAN neither of those options are available, as it doesn't use compiler instrumetation, no tag faults are ignored, and MTE is disabled after the first one. Instead, reset tags when accessing metadata (currently only for SLUB). Link: https://lkml.kernel.org/r/a0f3cefbc49f34c843b664110842de4db28179d0.1606161801.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Signed-off-by: Vincenzo Frascino <vincenzo.frascino@arm.com> Acked-by: Marco Elver <elver@google.com> Reviewed-by: Alexander Potapenko <glider@google.com> Tested-by: Vincenzo Frascino <vincenzo.frascino@arm.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Branislav Rankov <Branislav.Rankov@arm.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Evgenii Stepanov <eugenis@google.com> Cc: Kevin Brodsky <kevin.brodsky@arm.com> Cc: Vasily Gorbik <gor@linux.ibm.com> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-23 04:02:17 +08:00
freeptr_addr = (unsigned long)kasan_reset_tag((void *)freeptr_addr);
mm: add SLUB free list pointer obfuscation This SLUB free list pointer obfuscation code is modified from Brad Spengler/PaX Team's code in the last public patch of grsecurity/PaX based on my understanding of the code. Changes or omissions from the original code are mine and don't reflect the original grsecurity/PaX code. This adds a per-cache random value to SLUB caches that is XORed with their freelist pointer address and value. This adds nearly zero overhead and frustrates the very common heap overflow exploitation method of overwriting freelist pointers. A recent example of the attack is written up here: http://cyseclabs.com/blog/cve-2016-6187-heap-off-by-one-exploit and there is a section dedicated to the technique the book "A Guide to Kernel Exploitation: Attacking the Core". This is based on patches by Daniel Micay, and refactored to minimize the use of #ifdef. With 200-count cycles of "hackbench -g 20 -l 1000" I saw the following run times: before: mean 10.11882499999999999995 variance .03320378329145728642 stdev .18221905304181911048 after: mean 10.12654000000000000014 variance .04700556623115577889 stdev .21680767106160192064 The difference gets lost in the noise, but if the above is to be taken literally, using CONFIG_FREELIST_HARDENED is 0.07% slower. Link: http://lkml.kernel.org/r/20170802180609.GA66807@beast Signed-off-by: Kees Cook <keescook@chromium.org> Suggested-by: Daniel Micay <danielmicay@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Tycho Andersen <tycho@docker.com> Cc: Alexander Popov <alex.popov@linux.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:19:18 +08:00
*(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr);
}
/* Loop over all objects in a slab */
#define for_each_object(__p, __s, __addr, __objects) \
for (__p = fixup_red_left(__s, __addr); \
__p < (__addr) + (__objects) * (__s)->size; \
__p += (__s)->size)
static inline unsigned int order_objects(unsigned int order, unsigned int size)
{
return ((unsigned int)PAGE_SIZE << order) / size;
}
static inline struct kmem_cache_order_objects oo_make(unsigned int order,
unsigned int size)
{
struct kmem_cache_order_objects x = {
(order << OO_SHIFT) + order_objects(order, size)
};
return x;
}
static inline unsigned int oo_order(struct kmem_cache_order_objects x)
{
return x.x >> OO_SHIFT;
}
static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
{
return x.x & OO_MASK;
}
/*
* Per slab locking using the pagelock
*/
static __always_inline void slab_lock(struct page *page)
{
VM_BUG_ON_PAGE(PageTail(page), page);
bit_spin_lock(PG_locked, &page->flags);
}
static __always_inline void slab_unlock(struct page *page)
{
VM_BUG_ON_PAGE(PageTail(page), page);
__bit_spin_unlock(PG_locked, &page->flags);
}
/* Interrupts must be disabled (for the fallback code to work right) */
static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
void *freelist_old, unsigned long counters_old,
void *freelist_new, unsigned long counters_new,
const char *n)
{
VM_BUG_ON(!irqs_disabled());
#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
if (s->flags & __CMPXCHG_DOUBLE) {
if (cmpxchg_double(&page->freelist, &page->counters,
freelist_old, counters_old,
freelist_new, counters_new))
return true;
} else
#endif
{
slab_lock(page);
if (page->freelist == freelist_old &&
page->counters == counters_old) {
page->freelist = freelist_new;
page->counters = counters_new;
slab_unlock(page);
return true;
}
slab_unlock(page);
}
cpu_relax();
stat(s, CMPXCHG_DOUBLE_FAIL);
#ifdef SLUB_DEBUG_CMPXCHG
pr_info("%s %s: cmpxchg double redo ", n, s->name);
#endif
return false;
}
static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
void *freelist_old, unsigned long counters_old,
void *freelist_new, unsigned long counters_new,
const char *n)
{
#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
if (s->flags & __CMPXCHG_DOUBLE) {
if (cmpxchg_double(&page->freelist, &page->counters,
freelist_old, counters_old,
freelist_new, counters_new))
return true;
} else
#endif
{
unsigned long flags;
local_irq_save(flags);
slab_lock(page);
if (page->freelist == freelist_old &&
page->counters == counters_old) {
page->freelist = freelist_new;
page->counters = counters_new;
slab_unlock(page);
local_irq_restore(flags);
return true;
}
slab_unlock(page);
local_irq_restore(flags);
}
cpu_relax();
stat(s, CMPXCHG_DOUBLE_FAIL);
#ifdef SLUB_DEBUG_CMPXCHG
pr_info("%s %s: cmpxchg double redo ", n, s->name);
#endif
return false;
}
#ifdef CONFIG_SLUB_DEBUG
static unsigned long object_map[BITS_TO_LONGS(MAX_OBJS_PER_PAGE)];
static DEFINE_SPINLOCK(object_map_lock);
/*
* Determine a map of object in use on a page.
*
* Node listlock must be held to guarantee that the page does
* not vanish from under us.
*/
static unsigned long *get_map(struct kmem_cache *s, struct page *page)
__acquires(&object_map_lock)
{
void *p;
void *addr = page_address(page);
VM_BUG_ON(!irqs_disabled());
spin_lock(&object_map_lock);
bitmap_zero(object_map, page->objects);
for (p = page->freelist; p; p = get_freepointer(s, p))
set_bit(__obj_to_index(s, addr, p), object_map);
return object_map;
}
static void put_map(unsigned long *map) __releases(&object_map_lock)
{
VM_BUG_ON(map != object_map);
spin_unlock(&object_map_lock);
}
static inline unsigned int size_from_object(struct kmem_cache *s)
{
if (s->flags & SLAB_RED_ZONE)
return s->size - s->red_left_pad;
return s->size;
}
static inline void *restore_red_left(struct kmem_cache *s, void *p)
{
if (s->flags & SLAB_RED_ZONE)
p -= s->red_left_pad;
return p;
}
/*
* Debug settings:
*/
#if defined(CONFIG_SLUB_DEBUG_ON)
static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
#else
static slab_flags_t slub_debug;
#endif
2020-08-07 14:18:35 +08:00
static char *slub_debug_string;
static int disable_higher_order_debug;
mm: slub: introduce metadata_access_enable()/metadata_access_disable() It's ok for slub to access memory that marked by kasan as inaccessible (object's metadata). Kasan shouldn't print report in that case because these accesses are valid. Disabling instrumentation of slub.c code is not enough to achieve this because slub passes pointer to object's metadata into external functions like memchr_inv(). We don't want to disable instrumentation for memchr_inv() because this is quite generic function, and we don't want to miss bugs. metadata_access_enable/metadata_access_disable used to tell KASan where accesses to metadata starts/end, so we could temporarily disable KASan reports. Signed-off-by: Andrey Ryabinin <a.ryabinin@samsung.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Konstantin Serebryany <kcc@google.com> Cc: Dmitry Chernenkov <dmitryc@google.com> Signed-off-by: Andrey Konovalov <adech.fo@gmail.com> Cc: Yuri Gribov <tetra2005@gmail.com> Cc: Konstantin Khlebnikov <koct9i@gmail.com> Cc: Sasha Levin <sasha.levin@oracle.com> Cc: Christoph Lameter <cl@linux.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Ingo Molnar <mingo@elte.hu> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-14 06:39:38 +08:00
/*
* slub is about to manipulate internal object metadata. This memory lies
* outside the range of the allocated object, so accessing it would normally
* be reported by kasan as a bounds error. metadata_access_enable() is used
* to tell kasan that these accesses are OK.
*/
static inline void metadata_access_enable(void)
{
kasan_disable_current();
}
static inline void metadata_access_disable(void)
{
kasan_enable_current();
}
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/*
* Object debugging
*/
/* Verify that a pointer has an address that is valid within a slab page */
static inline int check_valid_pointer(struct kmem_cache *s,
struct page *page, void *object)
{
void *base;
if (!object)
return 1;
base = page_address(page);
slub: fix SLAB_CONSISTENCY_CHECKS + KASAN_SW_TAGS Enabling SLUB_DEBUG's SLAB_CONSISTENCY_CHECKS with KASAN_SW_TAGS triggers endless false positives during boot below due to check_valid_pointer() checks tagged pointers which have no addresses that is valid within slab pages: BUG radix_tree_node (Tainted: G B ): Freelist Pointer check fails ----------------------------------------------------------------------------- INFO: Slab objects=69 used=69 fp=0x (null) flags=0x7ffffffc000200 INFO: Object @offset=15060037153926966016 fp=0x Redzone: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb ................ Object : 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ Object : 00 00 00 00 00 00 00 00 18 6b 06 00 08 80 ff d0 .........k...... Object : 18 6b 06 00 08 80 ff d0 00 00 00 00 00 00 00 00 .k.............. Object : 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ Object : 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ Object : 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ Object : 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ Object : 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ Object : 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ Object : 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ Object : 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ Object : 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ Object : 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ Object : 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ Object : 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ Object : 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ Object : 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ Object : 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ Object : 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ Object : 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ Object : 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ Object : 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ Object : 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ Object : 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ Object : 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ Object : 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ Object : 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ Object : 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ Object : 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ Object : 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ Object : 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ Object : 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ Object : 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ Object : 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ Object : 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ Object : 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ Redzone: bb bb bb bb bb bb bb bb ........ Padding: 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a ZZZZZZZZZZZZZZZZ CPU: 0 PID: 0 Comm: swapper/0 Tainted: G B 5.0.0-rc5+ #18 Call trace: dump_backtrace+0x0/0x450 show_stack+0x20/0x2c __dump_stack+0x20/0x28 dump_stack+0xa0/0xfc print_trailer+0x1bc/0x1d0 object_err+0x40/0x50 alloc_debug_processing+0xf0/0x19c ___slab_alloc+0x554/0x704 kmem_cache_alloc+0x2f8/0x440 radix_tree_node_alloc+0x90/0x2fc idr_get_free+0x1e8/0x6d0 idr_alloc_u32+0x11c/0x2a4 idr_alloc+0x74/0xe0 worker_pool_assign_id+0x5c/0xbc workqueue_init_early+0x49c/0xd50 start_kernel+0x52c/0xac4 FIX radix_tree_node: Marking all objects used Link: http://lkml.kernel.org/r/20190209044128.3290-1-cai@lca.pw Signed-off-by: Qian Cai <cai@lca.pw> Reviewed-by: Andrey Konovalov <andreyknvl@google.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-02-21 14:19:36 +08:00
object = kasan_reset_tag(object);
object = restore_red_left(s, object);
if (object < base || object >= base + page->objects * s->size ||
(object - base) % s->size) {
return 0;
}
return 1;
}
static void print_section(char *level, char *text, u8 *addr,
unsigned int length)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
{
mm: slub: introduce metadata_access_enable()/metadata_access_disable() It's ok for slub to access memory that marked by kasan as inaccessible (object's metadata). Kasan shouldn't print report in that case because these accesses are valid. Disabling instrumentation of slub.c code is not enough to achieve this because slub passes pointer to object's metadata into external functions like memchr_inv(). We don't want to disable instrumentation for memchr_inv() because this is quite generic function, and we don't want to miss bugs. metadata_access_enable/metadata_access_disable used to tell KASan where accesses to metadata starts/end, so we could temporarily disable KASan reports. Signed-off-by: Andrey Ryabinin <a.ryabinin@samsung.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Konstantin Serebryany <kcc@google.com> Cc: Dmitry Chernenkov <dmitryc@google.com> Signed-off-by: Andrey Konovalov <adech.fo@gmail.com> Cc: Yuri Gribov <tetra2005@gmail.com> Cc: Konstantin Khlebnikov <koct9i@gmail.com> Cc: Sasha Levin <sasha.levin@oracle.com> Cc: Christoph Lameter <cl@linux.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Ingo Molnar <mingo@elte.hu> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-14 06:39:38 +08:00
metadata_access_enable();
kasan, mm: reset tags when accessing metadata Kernel allocator code accesses metadata for slab objects, that may lie out-of-bounds of the object itself, or be accessed when an object is freed. Such accesses trigger tag faults and lead to false-positive reports with hardware tag-based KASAN. Software KASAN modes disable instrumentation for allocator code via KASAN_SANITIZE Makefile macro, and rely on kasan_enable/disable_current() annotations which are used to ignore KASAN reports. With hardware tag-based KASAN neither of those options are available, as it doesn't use compiler instrumetation, no tag faults are ignored, and MTE is disabled after the first one. Instead, reset tags when accessing metadata (currently only for SLUB). Link: https://lkml.kernel.org/r/a0f3cefbc49f34c843b664110842de4db28179d0.1606161801.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Signed-off-by: Vincenzo Frascino <vincenzo.frascino@arm.com> Acked-by: Marco Elver <elver@google.com> Reviewed-by: Alexander Potapenko <glider@google.com> Tested-by: Vincenzo Frascino <vincenzo.frascino@arm.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Branislav Rankov <Branislav.Rankov@arm.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Evgenii Stepanov <eugenis@google.com> Cc: Kevin Brodsky <kevin.brodsky@arm.com> Cc: Vasily Gorbik <gor@linux.ibm.com> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-23 04:02:17 +08:00
print_hex_dump(level, kasan_reset_tag(text), DUMP_PREFIX_ADDRESS,
16, 1, addr, length, 1);
mm: slub: introduce metadata_access_enable()/metadata_access_disable() It's ok for slub to access memory that marked by kasan as inaccessible (object's metadata). Kasan shouldn't print report in that case because these accesses are valid. Disabling instrumentation of slub.c code is not enough to achieve this because slub passes pointer to object's metadata into external functions like memchr_inv(). We don't want to disable instrumentation for memchr_inv() because this is quite generic function, and we don't want to miss bugs. metadata_access_enable/metadata_access_disable used to tell KASan where accesses to metadata starts/end, so we could temporarily disable KASan reports. Signed-off-by: Andrey Ryabinin <a.ryabinin@samsung.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Konstantin Serebryany <kcc@google.com> Cc: Dmitry Chernenkov <dmitryc@google.com> Signed-off-by: Andrey Konovalov <adech.fo@gmail.com> Cc: Yuri Gribov <tetra2005@gmail.com> Cc: Konstantin Khlebnikov <koct9i@gmail.com> Cc: Sasha Levin <sasha.levin@oracle.com> Cc: Christoph Lameter <cl@linux.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Ingo Molnar <mingo@elte.hu> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-14 06:39:38 +08:00
metadata_access_disable();
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
mm/slub: fix incorrect interpretation of s->offset In a couple of places in the slub memory allocator, the code uses "s->offset" as a check to see if the free pointer is put right after the object. That check is no longer true with commit 3202fa62fb43 ("slub: relocate freelist pointer to middle of object"). As a result, echoing "1" into the validate sysfs file, e.g. of dentry, may cause a bunch of "Freepointer corrupt" error reports like the following to appear with the system in panic afterwards. ============================================================================= BUG dentry(666:pmcd.service) (Tainted: G B): Freepointer corrupt ----------------------------------------------------------------------------- To fix it, use the check "s->offset == s->inuse" in the new helper function freeptr_outside_object() instead. Also add another helper function get_info_end() to return the end of info block (inuse + free pointer if not overlapping with object). Fixes: 3202fa62fb43 ("slub: relocate freelist pointer to middle of object") Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Acked-by: Rafael Aquini <aquini@redhat.com> Cc: Christoph Lameter <cl@linux.com> Cc: Vitaly Nikolenko <vnik@duasynt.com> Cc: Silvio Cesare <silvio.cesare@gmail.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Markus Elfring <Markus.Elfring@web.de> Cc: Changbin Du <changbin.du@gmail.com> Link: http://lkml.kernel.org/r/20200429135328.26976-1-longman@redhat.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-05-08 09:36:06 +08:00
/*
* See comment in calculate_sizes().
*/
static inline bool freeptr_outside_object(struct kmem_cache *s)
{
return s->offset >= s->inuse;
}
/*
* Return offset of the end of info block which is inuse + free pointer if
* not overlapping with object.
*/
static inline unsigned int get_info_end(struct kmem_cache *s)
{
if (freeptr_outside_object(s))
return s->inuse + sizeof(void *);
else
return s->inuse;
}
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
static struct track *get_track(struct kmem_cache *s, void *object,
enum track_item alloc)
{
struct track *p;
mm/slub: fix incorrect interpretation of s->offset In a couple of places in the slub memory allocator, the code uses "s->offset" as a check to see if the free pointer is put right after the object. That check is no longer true with commit 3202fa62fb43 ("slub: relocate freelist pointer to middle of object"). As a result, echoing "1" into the validate sysfs file, e.g. of dentry, may cause a bunch of "Freepointer corrupt" error reports like the following to appear with the system in panic afterwards. ============================================================================= BUG dentry(666:pmcd.service) (Tainted: G B): Freepointer corrupt ----------------------------------------------------------------------------- To fix it, use the check "s->offset == s->inuse" in the new helper function freeptr_outside_object() instead. Also add another helper function get_info_end() to return the end of info block (inuse + free pointer if not overlapping with object). Fixes: 3202fa62fb43 ("slub: relocate freelist pointer to middle of object") Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Acked-by: Rafael Aquini <aquini@redhat.com> Cc: Christoph Lameter <cl@linux.com> Cc: Vitaly Nikolenko <vnik@duasynt.com> Cc: Silvio Cesare <silvio.cesare@gmail.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Markus Elfring <Markus.Elfring@web.de> Cc: Changbin Du <changbin.du@gmail.com> Link: http://lkml.kernel.org/r/20200429135328.26976-1-longman@redhat.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-05-08 09:36:06 +08:00
p = object + get_info_end(s);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
kasan, mm: reset tags when accessing metadata Kernel allocator code accesses metadata for slab objects, that may lie out-of-bounds of the object itself, or be accessed when an object is freed. Such accesses trigger tag faults and lead to false-positive reports with hardware tag-based KASAN. Software KASAN modes disable instrumentation for allocator code via KASAN_SANITIZE Makefile macro, and rely on kasan_enable/disable_current() annotations which are used to ignore KASAN reports. With hardware tag-based KASAN neither of those options are available, as it doesn't use compiler instrumetation, no tag faults are ignored, and MTE is disabled after the first one. Instead, reset tags when accessing metadata (currently only for SLUB). Link: https://lkml.kernel.org/r/a0f3cefbc49f34c843b664110842de4db28179d0.1606161801.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Signed-off-by: Vincenzo Frascino <vincenzo.frascino@arm.com> Acked-by: Marco Elver <elver@google.com> Reviewed-by: Alexander Potapenko <glider@google.com> Tested-by: Vincenzo Frascino <vincenzo.frascino@arm.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Branislav Rankov <Branislav.Rankov@arm.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Evgenii Stepanov <eugenis@google.com> Cc: Kevin Brodsky <kevin.brodsky@arm.com> Cc: Vasily Gorbik <gor@linux.ibm.com> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-23 04:02:17 +08:00
return kasan_reset_tag(p + alloc);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
static void set_track(struct kmem_cache *s, void *object,
enum track_item alloc, unsigned long addr)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
{
struct track *p = get_track(s, object, alloc);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (addr) {
slub: Enable backtrace for create/delete points This patch attempts to grab a backtrace for the creation and deletion points of the slub object. When a fault is detected, we can then get a better idea of where the item was deleted. Example output from debugging some funky nfs/rpc behaviour: ============================================================================= BUG kmalloc-64: Object is on free-list ----------------------------------------------------------------------------- INFO: Allocated in rpcb_getport_async+0x39c/0x5a5 [sunrpc] age=381 cpu=3 pid=3750 __slab_alloc+0x348/0x3ba kmem_cache_alloc_trace+0x67/0xe7 rpcb_getport_async+0x39c/0x5a5 [sunrpc] call_bind+0x70/0x75 [sunrpc] __rpc_execute+0x78/0x24b [sunrpc] rpc_execute+0x3d/0x42 [sunrpc] rpc_run_task+0x79/0x81 [sunrpc] rpc_call_sync+0x3f/0x60 [sunrpc] rpc_ping+0x42/0x58 [sunrpc] rpc_create+0x4aa/0x527 [sunrpc] nfs_create_rpc_client+0xb1/0xf6 [nfs] nfs_init_client+0x3b/0x7d [nfs] nfs_get_client+0x453/0x5ab [nfs] nfs_create_server+0x10b/0x437 [nfs] nfs_fs_mount+0x4ca/0x708 [nfs] mount_fs+0x6b/0x152 INFO: Freed in rpcb_map_release+0x3f/0x44 [sunrpc] age=30 cpu=2 pid=29049 __slab_free+0x57/0x150 kfree+0x107/0x13a rpcb_map_release+0x3f/0x44 [sunrpc] rpc_release_calldata+0x12/0x14 [sunrpc] rpc_free_task+0x59/0x61 [sunrpc] rpc_final_put_task+0x82/0x8a [sunrpc] __rpc_execute+0x23c/0x24b [sunrpc] rpc_async_schedule+0x10/0x12 [sunrpc] process_one_work+0x230/0x41d worker_thread+0x133/0x217 kthread+0x7d/0x85 kernel_thread_helper+0x4/0x10 INFO: Slab 0xffffea00029aa470 objects=20 used=9 fp=0xffff8800be7830d8 flags=0x20000000004081 INFO: Object 0xffff8800be7830d8 @offset=4312 fp=0xffff8800be7827a8 Bytes b4 0xffff8800be7830c8: 87 a8 96 00 01 00 00 00 5a 5a 5a 5a 5a 5a 5a 5a .�......ZZZZZZZZ Object 0xffff8800be7830d8: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk Object 0xffff8800be7830e8: 6b 6b 6b 6b 01 08 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkk..kkkkkkkkkk Object 0xffff8800be7830f8: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk Object 0xffff8800be783108: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b a5 kkkkkkkkkkkkkkk� Redzone 0xffff8800be783118: bb bb bb bb bb bb bb bb ������������� Padding 0xffff8800be783258: 5a 5a 5a 5a 5a 5a 5a 5a ZZZZZZZZ Pid: 29049, comm: kworker/2:2 Not tainted 3.0.0-rc4+ #8 Call Trace: [<ffffffff811055c3>] print_trailer+0x131/0x13a [<ffffffff81105601>] object_err+0x35/0x3e [<ffffffff8110746f>] verify_mem_not_deleted+0x7a/0xb7 [<ffffffffa02851b5>] rpcb_getport_done+0x23/0x126 [sunrpc] [<ffffffffa027d0ba>] rpc_exit_task+0x3f/0x6d [sunrpc] [<ffffffffa027d4ab>] __rpc_execute+0x78/0x24b [sunrpc] [<ffffffffa027d6c0>] ? rpc_execute+0x42/0x42 [sunrpc] [<ffffffffa027d6d0>] rpc_async_schedule+0x10/0x12 [sunrpc] [<ffffffff810611b7>] process_one_work+0x230/0x41d [<ffffffff81061102>] ? process_one_work+0x17b/0x41d [<ffffffff81063613>] worker_thread+0x133/0x217 [<ffffffff810634e0>] ? manage_workers+0x191/0x191 [<ffffffff81066e10>] kthread+0x7d/0x85 [<ffffffff81485924>] kernel_thread_helper+0x4/0x10 [<ffffffff8147eb18>] ? retint_restore_args+0x13/0x13 [<ffffffff81066d93>] ? __init_kthread_worker+0x56/0x56 [<ffffffff81485920>] ? gs_change+0x13/0x13 Acked-by: Christoph Lameter <cl@linux.com> Signed-off-by: Ben Greear <greearb@candelatech.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-07-08 02:36:36 +08:00
#ifdef CONFIG_STACKTRACE
mm/slub: Simplify stack trace retrieval Replace the indirection through struct stack_trace with an invocation of the storage array based interface. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Reviewed-by: Josh Poimboeuf <jpoimboe@redhat.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Andy Lutomirski <luto@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: linux-mm@kvack.org Cc: David Rientjes <rientjes@google.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Alexander Potapenko <glider@google.com> Cc: Alexey Dobriyan <adobriyan@gmail.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: kasan-dev@googlegroups.com Cc: Mike Rapoport <rppt@linux.vnet.ibm.com> Cc: Akinobu Mita <akinobu.mita@gmail.com> Cc: Christoph Hellwig <hch@lst.de> Cc: iommu@lists.linux-foundation.org Cc: Robin Murphy <robin.murphy@arm.com> Cc: Marek Szyprowski <m.szyprowski@samsung.com> Cc: Johannes Thumshirn <jthumshirn@suse.de> Cc: David Sterba <dsterba@suse.com> Cc: Chris Mason <clm@fb.com> Cc: Josef Bacik <josef@toxicpanda.com> Cc: linux-btrfs@vger.kernel.org Cc: dm-devel@redhat.com Cc: Mike Snitzer <snitzer@redhat.com> Cc: Alasdair Kergon <agk@redhat.com> Cc: Daniel Vetter <daniel@ffwll.ch> Cc: intel-gfx@lists.freedesktop.org Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com> Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> Cc: dri-devel@lists.freedesktop.org Cc: David Airlie <airlied@linux.ie> Cc: Jani Nikula <jani.nikula@linux.intel.com> Cc: Rodrigo Vivi <rodrigo.vivi@intel.com> Cc: Tom Zanussi <tom.zanussi@linux.intel.com> Cc: Miroslav Benes <mbenes@suse.cz> Cc: linux-arch@vger.kernel.org Link: https://lkml.kernel.org/r/20190425094801.771410441@linutronix.de
2019-04-25 17:45:00 +08:00
unsigned int nr_entries;
slub: Enable backtrace for create/delete points This patch attempts to grab a backtrace for the creation and deletion points of the slub object. When a fault is detected, we can then get a better idea of where the item was deleted. Example output from debugging some funky nfs/rpc behaviour: ============================================================================= BUG kmalloc-64: Object is on free-list ----------------------------------------------------------------------------- INFO: Allocated in rpcb_getport_async+0x39c/0x5a5 [sunrpc] age=381 cpu=3 pid=3750 __slab_alloc+0x348/0x3ba kmem_cache_alloc_trace+0x67/0xe7 rpcb_getport_async+0x39c/0x5a5 [sunrpc] call_bind+0x70/0x75 [sunrpc] __rpc_execute+0x78/0x24b [sunrpc] rpc_execute+0x3d/0x42 [sunrpc] rpc_run_task+0x79/0x81 [sunrpc] rpc_call_sync+0x3f/0x60 [sunrpc] rpc_ping+0x42/0x58 [sunrpc] rpc_create+0x4aa/0x527 [sunrpc] nfs_create_rpc_client+0xb1/0xf6 [nfs] nfs_init_client+0x3b/0x7d [nfs] nfs_get_client+0x453/0x5ab [nfs] nfs_create_server+0x10b/0x437 [nfs] nfs_fs_mount+0x4ca/0x708 [nfs] mount_fs+0x6b/0x152 INFO: Freed in rpcb_map_release+0x3f/0x44 [sunrpc] age=30 cpu=2 pid=29049 __slab_free+0x57/0x150 kfree+0x107/0x13a rpcb_map_release+0x3f/0x44 [sunrpc] rpc_release_calldata+0x12/0x14 [sunrpc] rpc_free_task+0x59/0x61 [sunrpc] rpc_final_put_task+0x82/0x8a [sunrpc] __rpc_execute+0x23c/0x24b [sunrpc] rpc_async_schedule+0x10/0x12 [sunrpc] process_one_work+0x230/0x41d worker_thread+0x133/0x217 kthread+0x7d/0x85 kernel_thread_helper+0x4/0x10 INFO: Slab 0xffffea00029aa470 objects=20 used=9 fp=0xffff8800be7830d8 flags=0x20000000004081 INFO: Object 0xffff8800be7830d8 @offset=4312 fp=0xffff8800be7827a8 Bytes b4 0xffff8800be7830c8: 87 a8 96 00 01 00 00 00 5a 5a 5a 5a 5a 5a 5a 5a .�......ZZZZZZZZ Object 0xffff8800be7830d8: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk Object 0xffff8800be7830e8: 6b 6b 6b 6b 01 08 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkk..kkkkkkkkkk Object 0xffff8800be7830f8: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk Object 0xffff8800be783108: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b a5 kkkkkkkkkkkkkkk� Redzone 0xffff8800be783118: bb bb bb bb bb bb bb bb ������������� Padding 0xffff8800be783258: 5a 5a 5a 5a 5a 5a 5a 5a ZZZZZZZZ Pid: 29049, comm: kworker/2:2 Not tainted 3.0.0-rc4+ #8 Call Trace: [<ffffffff811055c3>] print_trailer+0x131/0x13a [<ffffffff81105601>] object_err+0x35/0x3e [<ffffffff8110746f>] verify_mem_not_deleted+0x7a/0xb7 [<ffffffffa02851b5>] rpcb_getport_done+0x23/0x126 [sunrpc] [<ffffffffa027d0ba>] rpc_exit_task+0x3f/0x6d [sunrpc] [<ffffffffa027d4ab>] __rpc_execute+0x78/0x24b [sunrpc] [<ffffffffa027d6c0>] ? rpc_execute+0x42/0x42 [sunrpc] [<ffffffffa027d6d0>] rpc_async_schedule+0x10/0x12 [sunrpc] [<ffffffff810611b7>] process_one_work+0x230/0x41d [<ffffffff81061102>] ? process_one_work+0x17b/0x41d [<ffffffff81063613>] worker_thread+0x133/0x217 [<ffffffff810634e0>] ? manage_workers+0x191/0x191 [<ffffffff81066e10>] kthread+0x7d/0x85 [<ffffffff81485924>] kernel_thread_helper+0x4/0x10 [<ffffffff8147eb18>] ? retint_restore_args+0x13/0x13 [<ffffffff81066d93>] ? __init_kthread_worker+0x56/0x56 [<ffffffff81485920>] ? gs_change+0x13/0x13 Acked-by: Christoph Lameter <cl@linux.com> Signed-off-by: Ben Greear <greearb@candelatech.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-07-08 02:36:36 +08:00
mm: slub: introduce metadata_access_enable()/metadata_access_disable() It's ok for slub to access memory that marked by kasan as inaccessible (object's metadata). Kasan shouldn't print report in that case because these accesses are valid. Disabling instrumentation of slub.c code is not enough to achieve this because slub passes pointer to object's metadata into external functions like memchr_inv(). We don't want to disable instrumentation for memchr_inv() because this is quite generic function, and we don't want to miss bugs. metadata_access_enable/metadata_access_disable used to tell KASan where accesses to metadata starts/end, so we could temporarily disable KASan reports. Signed-off-by: Andrey Ryabinin <a.ryabinin@samsung.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Konstantin Serebryany <kcc@google.com> Cc: Dmitry Chernenkov <dmitryc@google.com> Signed-off-by: Andrey Konovalov <adech.fo@gmail.com> Cc: Yuri Gribov <tetra2005@gmail.com> Cc: Konstantin Khlebnikov <koct9i@gmail.com> Cc: Sasha Levin <sasha.levin@oracle.com> Cc: Christoph Lameter <cl@linux.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Ingo Molnar <mingo@elte.hu> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-14 06:39:38 +08:00
metadata_access_enable();
kasan, mm: reset tags when accessing metadata Kernel allocator code accesses metadata for slab objects, that may lie out-of-bounds of the object itself, or be accessed when an object is freed. Such accesses trigger tag faults and lead to false-positive reports with hardware tag-based KASAN. Software KASAN modes disable instrumentation for allocator code via KASAN_SANITIZE Makefile macro, and rely on kasan_enable/disable_current() annotations which are used to ignore KASAN reports. With hardware tag-based KASAN neither of those options are available, as it doesn't use compiler instrumetation, no tag faults are ignored, and MTE is disabled after the first one. Instead, reset tags when accessing metadata (currently only for SLUB). Link: https://lkml.kernel.org/r/a0f3cefbc49f34c843b664110842de4db28179d0.1606161801.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Signed-off-by: Vincenzo Frascino <vincenzo.frascino@arm.com> Acked-by: Marco Elver <elver@google.com> Reviewed-by: Alexander Potapenko <glider@google.com> Tested-by: Vincenzo Frascino <vincenzo.frascino@arm.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Branislav Rankov <Branislav.Rankov@arm.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Evgenii Stepanov <eugenis@google.com> Cc: Kevin Brodsky <kevin.brodsky@arm.com> Cc: Vasily Gorbik <gor@linux.ibm.com> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-23 04:02:17 +08:00
nr_entries = stack_trace_save(kasan_reset_tag(p->addrs),
TRACK_ADDRS_COUNT, 3);
mm: slub: introduce metadata_access_enable()/metadata_access_disable() It's ok for slub to access memory that marked by kasan as inaccessible (object's metadata). Kasan shouldn't print report in that case because these accesses are valid. Disabling instrumentation of slub.c code is not enough to achieve this because slub passes pointer to object's metadata into external functions like memchr_inv(). We don't want to disable instrumentation for memchr_inv() because this is quite generic function, and we don't want to miss bugs. metadata_access_enable/metadata_access_disable used to tell KASan where accesses to metadata starts/end, so we could temporarily disable KASan reports. Signed-off-by: Andrey Ryabinin <a.ryabinin@samsung.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Konstantin Serebryany <kcc@google.com> Cc: Dmitry Chernenkov <dmitryc@google.com> Signed-off-by: Andrey Konovalov <adech.fo@gmail.com> Cc: Yuri Gribov <tetra2005@gmail.com> Cc: Konstantin Khlebnikov <koct9i@gmail.com> Cc: Sasha Levin <sasha.levin@oracle.com> Cc: Christoph Lameter <cl@linux.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Ingo Molnar <mingo@elte.hu> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-14 06:39:38 +08:00
metadata_access_disable();
slub: Enable backtrace for create/delete points This patch attempts to grab a backtrace for the creation and deletion points of the slub object. When a fault is detected, we can then get a better idea of where the item was deleted. Example output from debugging some funky nfs/rpc behaviour: ============================================================================= BUG kmalloc-64: Object is on free-list ----------------------------------------------------------------------------- INFO: Allocated in rpcb_getport_async+0x39c/0x5a5 [sunrpc] age=381 cpu=3 pid=3750 __slab_alloc+0x348/0x3ba kmem_cache_alloc_trace+0x67/0xe7 rpcb_getport_async+0x39c/0x5a5 [sunrpc] call_bind+0x70/0x75 [sunrpc] __rpc_execute+0x78/0x24b [sunrpc] rpc_execute+0x3d/0x42 [sunrpc] rpc_run_task+0x79/0x81 [sunrpc] rpc_call_sync+0x3f/0x60 [sunrpc] rpc_ping+0x42/0x58 [sunrpc] rpc_create+0x4aa/0x527 [sunrpc] nfs_create_rpc_client+0xb1/0xf6 [nfs] nfs_init_client+0x3b/0x7d [nfs] nfs_get_client+0x453/0x5ab [nfs] nfs_create_server+0x10b/0x437 [nfs] nfs_fs_mount+0x4ca/0x708 [nfs] mount_fs+0x6b/0x152 INFO: Freed in rpcb_map_release+0x3f/0x44 [sunrpc] age=30 cpu=2 pid=29049 __slab_free+0x57/0x150 kfree+0x107/0x13a rpcb_map_release+0x3f/0x44 [sunrpc] rpc_release_calldata+0x12/0x14 [sunrpc] rpc_free_task+0x59/0x61 [sunrpc] rpc_final_put_task+0x82/0x8a [sunrpc] __rpc_execute+0x23c/0x24b [sunrpc] rpc_async_schedule+0x10/0x12 [sunrpc] process_one_work+0x230/0x41d worker_thread+0x133/0x217 kthread+0x7d/0x85 kernel_thread_helper+0x4/0x10 INFO: Slab 0xffffea00029aa470 objects=20 used=9 fp=0xffff8800be7830d8 flags=0x20000000004081 INFO: Object 0xffff8800be7830d8 @offset=4312 fp=0xffff8800be7827a8 Bytes b4 0xffff8800be7830c8: 87 a8 96 00 01 00 00 00 5a 5a 5a 5a 5a 5a 5a 5a .�......ZZZZZZZZ Object 0xffff8800be7830d8: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk Object 0xffff8800be7830e8: 6b 6b 6b 6b 01 08 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkk..kkkkkkkkkk Object 0xffff8800be7830f8: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk Object 0xffff8800be783108: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b a5 kkkkkkkkkkkkkkk� Redzone 0xffff8800be783118: bb bb bb bb bb bb bb bb ������������� Padding 0xffff8800be783258: 5a 5a 5a 5a 5a 5a 5a 5a ZZZZZZZZ Pid: 29049, comm: kworker/2:2 Not tainted 3.0.0-rc4+ #8 Call Trace: [<ffffffff811055c3>] print_trailer+0x131/0x13a [<ffffffff81105601>] object_err+0x35/0x3e [<ffffffff8110746f>] verify_mem_not_deleted+0x7a/0xb7 [<ffffffffa02851b5>] rpcb_getport_done+0x23/0x126 [sunrpc] [<ffffffffa027d0ba>] rpc_exit_task+0x3f/0x6d [sunrpc] [<ffffffffa027d4ab>] __rpc_execute+0x78/0x24b [sunrpc] [<ffffffffa027d6c0>] ? rpc_execute+0x42/0x42 [sunrpc] [<ffffffffa027d6d0>] rpc_async_schedule+0x10/0x12 [sunrpc] [<ffffffff810611b7>] process_one_work+0x230/0x41d [<ffffffff81061102>] ? process_one_work+0x17b/0x41d [<ffffffff81063613>] worker_thread+0x133/0x217 [<ffffffff810634e0>] ? manage_workers+0x191/0x191 [<ffffffff81066e10>] kthread+0x7d/0x85 [<ffffffff81485924>] kernel_thread_helper+0x4/0x10 [<ffffffff8147eb18>] ? retint_restore_args+0x13/0x13 [<ffffffff81066d93>] ? __init_kthread_worker+0x56/0x56 [<ffffffff81485920>] ? gs_change+0x13/0x13 Acked-by: Christoph Lameter <cl@linux.com> Signed-off-by: Ben Greear <greearb@candelatech.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-07-08 02:36:36 +08:00
mm/slub: Simplify stack trace retrieval Replace the indirection through struct stack_trace with an invocation of the storage array based interface. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Reviewed-by: Josh Poimboeuf <jpoimboe@redhat.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Andy Lutomirski <luto@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: linux-mm@kvack.org Cc: David Rientjes <rientjes@google.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Alexander Potapenko <glider@google.com> Cc: Alexey Dobriyan <adobriyan@gmail.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: kasan-dev@googlegroups.com Cc: Mike Rapoport <rppt@linux.vnet.ibm.com> Cc: Akinobu Mita <akinobu.mita@gmail.com> Cc: Christoph Hellwig <hch@lst.de> Cc: iommu@lists.linux-foundation.org Cc: Robin Murphy <robin.murphy@arm.com> Cc: Marek Szyprowski <m.szyprowski@samsung.com> Cc: Johannes Thumshirn <jthumshirn@suse.de> Cc: David Sterba <dsterba@suse.com> Cc: Chris Mason <clm@fb.com> Cc: Josef Bacik <josef@toxicpanda.com> Cc: linux-btrfs@vger.kernel.org Cc: dm-devel@redhat.com Cc: Mike Snitzer <snitzer@redhat.com> Cc: Alasdair Kergon <agk@redhat.com> Cc: Daniel Vetter <daniel@ffwll.ch> Cc: intel-gfx@lists.freedesktop.org Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com> Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> Cc: dri-devel@lists.freedesktop.org Cc: David Airlie <airlied@linux.ie> Cc: Jani Nikula <jani.nikula@linux.intel.com> Cc: Rodrigo Vivi <rodrigo.vivi@intel.com> Cc: Tom Zanussi <tom.zanussi@linux.intel.com> Cc: Miroslav Benes <mbenes@suse.cz> Cc: linux-arch@vger.kernel.org Link: https://lkml.kernel.org/r/20190425094801.771410441@linutronix.de
2019-04-25 17:45:00 +08:00
if (nr_entries < TRACK_ADDRS_COUNT)
p->addrs[nr_entries] = 0;
slub: Enable backtrace for create/delete points This patch attempts to grab a backtrace for the creation and deletion points of the slub object. When a fault is detected, we can then get a better idea of where the item was deleted. Example output from debugging some funky nfs/rpc behaviour: ============================================================================= BUG kmalloc-64: Object is on free-list ----------------------------------------------------------------------------- INFO: Allocated in rpcb_getport_async+0x39c/0x5a5 [sunrpc] age=381 cpu=3 pid=3750 __slab_alloc+0x348/0x3ba kmem_cache_alloc_trace+0x67/0xe7 rpcb_getport_async+0x39c/0x5a5 [sunrpc] call_bind+0x70/0x75 [sunrpc] __rpc_execute+0x78/0x24b [sunrpc] rpc_execute+0x3d/0x42 [sunrpc] rpc_run_task+0x79/0x81 [sunrpc] rpc_call_sync+0x3f/0x60 [sunrpc] rpc_ping+0x42/0x58 [sunrpc] rpc_create+0x4aa/0x527 [sunrpc] nfs_create_rpc_client+0xb1/0xf6 [nfs] nfs_init_client+0x3b/0x7d [nfs] nfs_get_client+0x453/0x5ab [nfs] nfs_create_server+0x10b/0x437 [nfs] nfs_fs_mount+0x4ca/0x708 [nfs] mount_fs+0x6b/0x152 INFO: Freed in rpcb_map_release+0x3f/0x44 [sunrpc] age=30 cpu=2 pid=29049 __slab_free+0x57/0x150 kfree+0x107/0x13a rpcb_map_release+0x3f/0x44 [sunrpc] rpc_release_calldata+0x12/0x14 [sunrpc] rpc_free_task+0x59/0x61 [sunrpc] rpc_final_put_task+0x82/0x8a [sunrpc] __rpc_execute+0x23c/0x24b [sunrpc] rpc_async_schedule+0x10/0x12 [sunrpc] process_one_work+0x230/0x41d worker_thread+0x133/0x217 kthread+0x7d/0x85 kernel_thread_helper+0x4/0x10 INFO: Slab 0xffffea00029aa470 objects=20 used=9 fp=0xffff8800be7830d8 flags=0x20000000004081 INFO: Object 0xffff8800be7830d8 @offset=4312 fp=0xffff8800be7827a8 Bytes b4 0xffff8800be7830c8: 87 a8 96 00 01 00 00 00 5a 5a 5a 5a 5a 5a 5a 5a .�......ZZZZZZZZ Object 0xffff8800be7830d8: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk Object 0xffff8800be7830e8: 6b 6b 6b 6b 01 08 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkk..kkkkkkkkkk Object 0xffff8800be7830f8: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk Object 0xffff8800be783108: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b a5 kkkkkkkkkkkkkkk� Redzone 0xffff8800be783118: bb bb bb bb bb bb bb bb ������������� Padding 0xffff8800be783258: 5a 5a 5a 5a 5a 5a 5a 5a ZZZZZZZZ Pid: 29049, comm: kworker/2:2 Not tainted 3.0.0-rc4+ #8 Call Trace: [<ffffffff811055c3>] print_trailer+0x131/0x13a [<ffffffff81105601>] object_err+0x35/0x3e [<ffffffff8110746f>] verify_mem_not_deleted+0x7a/0xb7 [<ffffffffa02851b5>] rpcb_getport_done+0x23/0x126 [sunrpc] [<ffffffffa027d0ba>] rpc_exit_task+0x3f/0x6d [sunrpc] [<ffffffffa027d4ab>] __rpc_execute+0x78/0x24b [sunrpc] [<ffffffffa027d6c0>] ? rpc_execute+0x42/0x42 [sunrpc] [<ffffffffa027d6d0>] rpc_async_schedule+0x10/0x12 [sunrpc] [<ffffffff810611b7>] process_one_work+0x230/0x41d [<ffffffff81061102>] ? process_one_work+0x17b/0x41d [<ffffffff81063613>] worker_thread+0x133/0x217 [<ffffffff810634e0>] ? manage_workers+0x191/0x191 [<ffffffff81066e10>] kthread+0x7d/0x85 [<ffffffff81485924>] kernel_thread_helper+0x4/0x10 [<ffffffff8147eb18>] ? retint_restore_args+0x13/0x13 [<ffffffff81066d93>] ? __init_kthread_worker+0x56/0x56 [<ffffffff81485920>] ? gs_change+0x13/0x13 Acked-by: Christoph Lameter <cl@linux.com> Signed-off-by: Ben Greear <greearb@candelatech.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-07-08 02:36:36 +08:00
#endif
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
p->addr = addr;
p->cpu = smp_processor_id();
p->pid = current->pid;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
p->when = jiffies;
} else {
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
memset(p, 0, sizeof(struct track));
}
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
static void init_tracking(struct kmem_cache *s, void *object)
{
if (!(s->flags & SLAB_STORE_USER))
return;
set_track(s, object, TRACK_FREE, 0UL);
set_track(s, object, TRACK_ALLOC, 0UL);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
mm/slub.c: use jitter-free reference while printing age When SLUB_DEBUG catches some issues, it prints all the required debug info. However, in a few cases where allocation and free of the object has happened in a very short time, 'age' might be misleading. See the example below: ============================================================================= BUG kmalloc-256 (Tainted: G W O ): Poison overwritten ----------------------------------------------------------------------------- ... INFO: Allocated in binder_transaction+0x4b0/0x2448 age=731 cpu=3 pid=5314 ... INFO: Freed in binder_free_transaction+0x2c/0x58 age=735 cpu=6 pid=2079 ... Object fffffff14956a870: 6b 6b 6b 6b 6b 6b 6b 6b 67 6b 6b 6b 6b 6b 6b a5 kkkkkkkkgkkkk In this case, object got freed later but 'age' shows otherwise. This could be because, while printing this info, we print allocation traces first and free traces thereafter. In between, if we get schedule out or jiffies increment, (jiffies - t->when) could become meaningless. Use the jitter free reference to calculate age. New output will exactly be same. 'age' is still staying with single jiffies ref in both prints. Change-Id: I0846565807a4229748649bbecb1ffb743d71fcd8 Link: http://lkml.kernel.org/r/1520492010-19389-1-git-send-email-cpandya@codeaurora.org Signed-off-by: Chintan Pandya <cpandya@codeaurora.org> Acked-by: Christoph Lameter <cl@linux.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-04-06 07:20:15 +08:00
static void print_track(const char *s, struct track *t, unsigned long pr_time)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
{
if (!t->addr)
return;
pr_err("%s in %pS age=%lu cpu=%u pid=%d\n",
mm/slub.c: use jitter-free reference while printing age When SLUB_DEBUG catches some issues, it prints all the required debug info. However, in a few cases where allocation and free of the object has happened in a very short time, 'age' might be misleading. See the example below: ============================================================================= BUG kmalloc-256 (Tainted: G W O ): Poison overwritten ----------------------------------------------------------------------------- ... INFO: Allocated in binder_transaction+0x4b0/0x2448 age=731 cpu=3 pid=5314 ... INFO: Freed in binder_free_transaction+0x2c/0x58 age=735 cpu=6 pid=2079 ... Object fffffff14956a870: 6b 6b 6b 6b 6b 6b 6b 6b 67 6b 6b 6b 6b 6b 6b a5 kkkkkkkkgkkkk In this case, object got freed later but 'age' shows otherwise. This could be because, while printing this info, we print allocation traces first and free traces thereafter. In between, if we get schedule out or jiffies increment, (jiffies - t->when) could become meaningless. Use the jitter free reference to calculate age. New output will exactly be same. 'age' is still staying with single jiffies ref in both prints. Change-Id: I0846565807a4229748649bbecb1ffb743d71fcd8 Link: http://lkml.kernel.org/r/1520492010-19389-1-git-send-email-cpandya@codeaurora.org Signed-off-by: Chintan Pandya <cpandya@codeaurora.org> Acked-by: Christoph Lameter <cl@linux.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-04-06 07:20:15 +08:00
s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
slub: Enable backtrace for create/delete points This patch attempts to grab a backtrace for the creation and deletion points of the slub object. When a fault is detected, we can then get a better idea of where the item was deleted. Example output from debugging some funky nfs/rpc behaviour: ============================================================================= BUG kmalloc-64: Object is on free-list ----------------------------------------------------------------------------- INFO: Allocated in rpcb_getport_async+0x39c/0x5a5 [sunrpc] age=381 cpu=3 pid=3750 __slab_alloc+0x348/0x3ba kmem_cache_alloc_trace+0x67/0xe7 rpcb_getport_async+0x39c/0x5a5 [sunrpc] call_bind+0x70/0x75 [sunrpc] __rpc_execute+0x78/0x24b [sunrpc] rpc_execute+0x3d/0x42 [sunrpc] rpc_run_task+0x79/0x81 [sunrpc] rpc_call_sync+0x3f/0x60 [sunrpc] rpc_ping+0x42/0x58 [sunrpc] rpc_create+0x4aa/0x527 [sunrpc] nfs_create_rpc_client+0xb1/0xf6 [nfs] nfs_init_client+0x3b/0x7d [nfs] nfs_get_client+0x453/0x5ab [nfs] nfs_create_server+0x10b/0x437 [nfs] nfs_fs_mount+0x4ca/0x708 [nfs] mount_fs+0x6b/0x152 INFO: Freed in rpcb_map_release+0x3f/0x44 [sunrpc] age=30 cpu=2 pid=29049 __slab_free+0x57/0x150 kfree+0x107/0x13a rpcb_map_release+0x3f/0x44 [sunrpc] rpc_release_calldata+0x12/0x14 [sunrpc] rpc_free_task+0x59/0x61 [sunrpc] rpc_final_put_task+0x82/0x8a [sunrpc] __rpc_execute+0x23c/0x24b [sunrpc] rpc_async_schedule+0x10/0x12 [sunrpc] process_one_work+0x230/0x41d worker_thread+0x133/0x217 kthread+0x7d/0x85 kernel_thread_helper+0x4/0x10 INFO: Slab 0xffffea00029aa470 objects=20 used=9 fp=0xffff8800be7830d8 flags=0x20000000004081 INFO: Object 0xffff8800be7830d8 @offset=4312 fp=0xffff8800be7827a8 Bytes b4 0xffff8800be7830c8: 87 a8 96 00 01 00 00 00 5a 5a 5a 5a 5a 5a 5a 5a .�......ZZZZZZZZ Object 0xffff8800be7830d8: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk Object 0xffff8800be7830e8: 6b 6b 6b 6b 01 08 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkk..kkkkkkkkkk Object 0xffff8800be7830f8: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk Object 0xffff8800be783108: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b a5 kkkkkkkkkkkkkkk� Redzone 0xffff8800be783118: bb bb bb bb bb bb bb bb ������������� Padding 0xffff8800be783258: 5a 5a 5a 5a 5a 5a 5a 5a ZZZZZZZZ Pid: 29049, comm: kworker/2:2 Not tainted 3.0.0-rc4+ #8 Call Trace: [<ffffffff811055c3>] print_trailer+0x131/0x13a [<ffffffff81105601>] object_err+0x35/0x3e [<ffffffff8110746f>] verify_mem_not_deleted+0x7a/0xb7 [<ffffffffa02851b5>] rpcb_getport_done+0x23/0x126 [sunrpc] [<ffffffffa027d0ba>] rpc_exit_task+0x3f/0x6d [sunrpc] [<ffffffffa027d4ab>] __rpc_execute+0x78/0x24b [sunrpc] [<ffffffffa027d6c0>] ? rpc_execute+0x42/0x42 [sunrpc] [<ffffffffa027d6d0>] rpc_async_schedule+0x10/0x12 [sunrpc] [<ffffffff810611b7>] process_one_work+0x230/0x41d [<ffffffff81061102>] ? process_one_work+0x17b/0x41d [<ffffffff81063613>] worker_thread+0x133/0x217 [<ffffffff810634e0>] ? manage_workers+0x191/0x191 [<ffffffff81066e10>] kthread+0x7d/0x85 [<ffffffff81485924>] kernel_thread_helper+0x4/0x10 [<ffffffff8147eb18>] ? retint_restore_args+0x13/0x13 [<ffffffff81066d93>] ? __init_kthread_worker+0x56/0x56 [<ffffffff81485920>] ? gs_change+0x13/0x13 Acked-by: Christoph Lameter <cl@linux.com> Signed-off-by: Ben Greear <greearb@candelatech.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-07-08 02:36:36 +08:00
#ifdef CONFIG_STACKTRACE
{
int i;
for (i = 0; i < TRACK_ADDRS_COUNT; i++)
if (t->addrs[i])
pr_err("\t%pS\n", (void *)t->addrs[i]);
slub: Enable backtrace for create/delete points This patch attempts to grab a backtrace for the creation and deletion points of the slub object. When a fault is detected, we can then get a better idea of where the item was deleted. Example output from debugging some funky nfs/rpc behaviour: ============================================================================= BUG kmalloc-64: Object is on free-list ----------------------------------------------------------------------------- INFO: Allocated in rpcb_getport_async+0x39c/0x5a5 [sunrpc] age=381 cpu=3 pid=3750 __slab_alloc+0x348/0x3ba kmem_cache_alloc_trace+0x67/0xe7 rpcb_getport_async+0x39c/0x5a5 [sunrpc] call_bind+0x70/0x75 [sunrpc] __rpc_execute+0x78/0x24b [sunrpc] rpc_execute+0x3d/0x42 [sunrpc] rpc_run_task+0x79/0x81 [sunrpc] rpc_call_sync+0x3f/0x60 [sunrpc] rpc_ping+0x42/0x58 [sunrpc] rpc_create+0x4aa/0x527 [sunrpc] nfs_create_rpc_client+0xb1/0xf6 [nfs] nfs_init_client+0x3b/0x7d [nfs] nfs_get_client+0x453/0x5ab [nfs] nfs_create_server+0x10b/0x437 [nfs] nfs_fs_mount+0x4ca/0x708 [nfs] mount_fs+0x6b/0x152 INFO: Freed in rpcb_map_release+0x3f/0x44 [sunrpc] age=30 cpu=2 pid=29049 __slab_free+0x57/0x150 kfree+0x107/0x13a rpcb_map_release+0x3f/0x44 [sunrpc] rpc_release_calldata+0x12/0x14 [sunrpc] rpc_free_task+0x59/0x61 [sunrpc] rpc_final_put_task+0x82/0x8a [sunrpc] __rpc_execute+0x23c/0x24b [sunrpc] rpc_async_schedule+0x10/0x12 [sunrpc] process_one_work+0x230/0x41d worker_thread+0x133/0x217 kthread+0x7d/0x85 kernel_thread_helper+0x4/0x10 INFO: Slab 0xffffea00029aa470 objects=20 used=9 fp=0xffff8800be7830d8 flags=0x20000000004081 INFO: Object 0xffff8800be7830d8 @offset=4312 fp=0xffff8800be7827a8 Bytes b4 0xffff8800be7830c8: 87 a8 96 00 01 00 00 00 5a 5a 5a 5a 5a 5a 5a 5a .�......ZZZZZZZZ Object 0xffff8800be7830d8: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk Object 0xffff8800be7830e8: 6b 6b 6b 6b 01 08 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkk..kkkkkkkkkk Object 0xffff8800be7830f8: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk Object 0xffff8800be783108: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b a5 kkkkkkkkkkkkkkk� Redzone 0xffff8800be783118: bb bb bb bb bb bb bb bb ������������� Padding 0xffff8800be783258: 5a 5a 5a 5a 5a 5a 5a 5a ZZZZZZZZ Pid: 29049, comm: kworker/2:2 Not tainted 3.0.0-rc4+ #8 Call Trace: [<ffffffff811055c3>] print_trailer+0x131/0x13a [<ffffffff81105601>] object_err+0x35/0x3e [<ffffffff8110746f>] verify_mem_not_deleted+0x7a/0xb7 [<ffffffffa02851b5>] rpcb_getport_done+0x23/0x126 [sunrpc] [<ffffffffa027d0ba>] rpc_exit_task+0x3f/0x6d [sunrpc] [<ffffffffa027d4ab>] __rpc_execute+0x78/0x24b [sunrpc] [<ffffffffa027d6c0>] ? rpc_execute+0x42/0x42 [sunrpc] [<ffffffffa027d6d0>] rpc_async_schedule+0x10/0x12 [sunrpc] [<ffffffff810611b7>] process_one_work+0x230/0x41d [<ffffffff81061102>] ? process_one_work+0x17b/0x41d [<ffffffff81063613>] worker_thread+0x133/0x217 [<ffffffff810634e0>] ? manage_workers+0x191/0x191 [<ffffffff81066e10>] kthread+0x7d/0x85 [<ffffffff81485924>] kernel_thread_helper+0x4/0x10 [<ffffffff8147eb18>] ? retint_restore_args+0x13/0x13 [<ffffffff81066d93>] ? __init_kthread_worker+0x56/0x56 [<ffffffff81485920>] ? gs_change+0x13/0x13 Acked-by: Christoph Lameter <cl@linux.com> Signed-off-by: Ben Greear <greearb@candelatech.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-07-08 02:36:36 +08:00
else
break;
}
#endif
}
mm, slab/slub: improve error reporting and overhead of cache_from_obj() cache_from_obj() was added by commit b9ce5ef49f00 ("sl[au]b: always get the cache from its page in kmem_cache_free()") to support kmemcg, where per-memcg cache can be different from the root one, so we can't use the kmem_cache pointer given to kmem_cache_free(). Prior to that commit, SLUB already had debugging check+warning that could be enabled to compare the given kmem_cache pointer to one referenced by the slab page where the object-to-be-freed resides. This check was moved to cache_from_obj(). Later the check was also enabled for SLAB_FREELIST_HARDENED configs by commit 598a0717a816 ("mm/slab: validate cache membership under freelist hardening"). These checks and warnings can be useful especially for the debugging, which can be improved. Commit 598a0717a816 changed the pr_err() with WARN_ON_ONCE() to WARN_ONCE() so only the first hit is now reported, others are silent. This patch changes it to WARN() so that all errors are reported. It's also useful to print SLUB allocation/free tracking info for the offending object, if tracking is enabled. Thus, export the SLUB print_tracking() function and provide an empty one for SLAB. For SLUB we can also benefit from the static key check in kmem_cache_debug_flags(), but we need to move this function to slab.h and declare the static key there. [1] https://lore.kernel.org/r/20200608230654.828134-18-guro@fb.com [vbabka@suse.cz: avoid bogus WARN()] Link: https://lore.kernel.org/r/20200623090213.GW5535@shao2-debian Link: http://lkml.kernel.org/r/b33e0fa7-cd28-4788-9e54-5927846329ef@suse.cz Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Acked-by: Kees Cook <keescook@chromium.org> Acked-by: Roman Gushchin <guro@fb.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Matthew Garrett <mjg59@google.com> Cc: Jann Horn <jannh@google.com> Cc: Vijayanand Jitta <vjitta@codeaurora.org> Cc: Vinayak Menon <vinmenon@codeaurora.org> Link: http://lkml.kernel.org/r/afeda7ac-748b-33d8-a905-56b708148ad5@suse.cz Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-07 14:19:05 +08:00
void print_tracking(struct kmem_cache *s, void *object)
{
mm/slub.c: use jitter-free reference while printing age When SLUB_DEBUG catches some issues, it prints all the required debug info. However, in a few cases where allocation and free of the object has happened in a very short time, 'age' might be misleading. See the example below: ============================================================================= BUG kmalloc-256 (Tainted: G W O ): Poison overwritten ----------------------------------------------------------------------------- ... INFO: Allocated in binder_transaction+0x4b0/0x2448 age=731 cpu=3 pid=5314 ... INFO: Freed in binder_free_transaction+0x2c/0x58 age=735 cpu=6 pid=2079 ... Object fffffff14956a870: 6b 6b 6b 6b 6b 6b 6b 6b 67 6b 6b 6b 6b 6b 6b a5 kkkkkkkkgkkkk In this case, object got freed later but 'age' shows otherwise. This could be because, while printing this info, we print allocation traces first and free traces thereafter. In between, if we get schedule out or jiffies increment, (jiffies - t->when) could become meaningless. Use the jitter free reference to calculate age. New output will exactly be same. 'age' is still staying with single jiffies ref in both prints. Change-Id: I0846565807a4229748649bbecb1ffb743d71fcd8 Link: http://lkml.kernel.org/r/1520492010-19389-1-git-send-email-cpandya@codeaurora.org Signed-off-by: Chintan Pandya <cpandya@codeaurora.org> Acked-by: Christoph Lameter <cl@linux.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-04-06 07:20:15 +08:00
unsigned long pr_time = jiffies;
if (!(s->flags & SLAB_STORE_USER))
return;
mm/slub.c: use jitter-free reference while printing age When SLUB_DEBUG catches some issues, it prints all the required debug info. However, in a few cases where allocation and free of the object has happened in a very short time, 'age' might be misleading. See the example below: ============================================================================= BUG kmalloc-256 (Tainted: G W O ): Poison overwritten ----------------------------------------------------------------------------- ... INFO: Allocated in binder_transaction+0x4b0/0x2448 age=731 cpu=3 pid=5314 ... INFO: Freed in binder_free_transaction+0x2c/0x58 age=735 cpu=6 pid=2079 ... Object fffffff14956a870: 6b 6b 6b 6b 6b 6b 6b 6b 67 6b 6b 6b 6b 6b 6b a5 kkkkkkkkgkkkk In this case, object got freed later but 'age' shows otherwise. This could be because, while printing this info, we print allocation traces first and free traces thereafter. In between, if we get schedule out or jiffies increment, (jiffies - t->when) could become meaningless. Use the jitter free reference to calculate age. New output will exactly be same. 'age' is still staying with single jiffies ref in both prints. Change-Id: I0846565807a4229748649bbecb1ffb743d71fcd8 Link: http://lkml.kernel.org/r/1520492010-19389-1-git-send-email-cpandya@codeaurora.org Signed-off-by: Chintan Pandya <cpandya@codeaurora.org> Acked-by: Christoph Lameter <cl@linux.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-04-06 07:20:15 +08:00
print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
}
static void print_page_info(struct page *page)
{
pr_err("Slab 0x%p objects=%u used=%u fp=0x%p flags=%#lx(%pGp)\n",
page, page->objects, page->inuse, page->freelist,
page->flags, &page->flags);
}
static void slab_bug(struct kmem_cache *s, char *fmt, ...)
{
struct va_format vaf;
va_list args;
va_start(args, fmt);
vaf.fmt = fmt;
vaf.va = &args;
pr_err("=============================================================================\n");
pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
pr_err("-----------------------------------------------------------------------------\n\n");
add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
va_end(args);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
static void slab_fix(struct kmem_cache *s, char *fmt, ...)
{
struct va_format vaf;
va_list args;
va_start(args, fmt);
vaf.fmt = fmt;
vaf.va = &args;
pr_err("FIX %s: %pV\n", s->name, &vaf);
va_end(args);
}
mm/slub.c: fix corrupted freechain in deactivate_slab() The slub_debug is able to fix the corrupted slab freelist/page. However, alloc_debug_processing() only checks the validity of current and next freepointer during allocation path. As a result, once some objects have their freepointers corrupted, deactivate_slab() may lead to page fault. Below is from a test kernel module when 'slub_debug=PUF,kmalloc-128 slub_nomerge'. The test kernel corrupts the freepointer of one free object on purpose. Unfortunately, deactivate_slab() does not detect it when iterating the freechain. BUG: unable to handle page fault for address: 00000000123456f8 #PF: supervisor read access in kernel mode #PF: error_code(0x0000) - not-present page PGD 0 P4D 0 Oops: 0000 [#1] SMP PTI ... ... RIP: 0010:deactivate_slab.isra.92+0xed/0x490 ... ... Call Trace: ___slab_alloc+0x536/0x570 __slab_alloc+0x17/0x30 __kmalloc+0x1d9/0x200 ext4_htree_store_dirent+0x30/0xf0 htree_dirblock_to_tree+0xcb/0x1c0 ext4_htree_fill_tree+0x1bc/0x2d0 ext4_readdir+0x54f/0x920 iterate_dir+0x88/0x190 __x64_sys_getdents+0xa6/0x140 do_syscall_64+0x49/0x170 entry_SYSCALL_64_after_hwframe+0x44/0xa9 Therefore, this patch adds extra consistency check in deactivate_slab(). Once an object's freepointer is corrupted, all following objects starting at this object are isolated. [akpm@linux-foundation.org: fix build with CONFIG_SLAB_DEBUG=n] Signed-off-by: Dongli Zhang <dongli.zhang@oracle.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Cc: Joe Jin <joe.jin@oracle.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Link: http://lkml.kernel.org/r/20200331031450.12182-1-dongli.zhang@oracle.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-02 12:45:47 +08:00
static bool freelist_corrupted(struct kmem_cache *s, struct page *page,
void **freelist, void *nextfree)
mm/slub.c: fix corrupted freechain in deactivate_slab() The slub_debug is able to fix the corrupted slab freelist/page. However, alloc_debug_processing() only checks the validity of current and next freepointer during allocation path. As a result, once some objects have their freepointers corrupted, deactivate_slab() may lead to page fault. Below is from a test kernel module when 'slub_debug=PUF,kmalloc-128 slub_nomerge'. The test kernel corrupts the freepointer of one free object on purpose. Unfortunately, deactivate_slab() does not detect it when iterating the freechain. BUG: unable to handle page fault for address: 00000000123456f8 #PF: supervisor read access in kernel mode #PF: error_code(0x0000) - not-present page PGD 0 P4D 0 Oops: 0000 [#1] SMP PTI ... ... RIP: 0010:deactivate_slab.isra.92+0xed/0x490 ... ... Call Trace: ___slab_alloc+0x536/0x570 __slab_alloc+0x17/0x30 __kmalloc+0x1d9/0x200 ext4_htree_store_dirent+0x30/0xf0 htree_dirblock_to_tree+0xcb/0x1c0 ext4_htree_fill_tree+0x1bc/0x2d0 ext4_readdir+0x54f/0x920 iterate_dir+0x88/0x190 __x64_sys_getdents+0xa6/0x140 do_syscall_64+0x49/0x170 entry_SYSCALL_64_after_hwframe+0x44/0xa9 Therefore, this patch adds extra consistency check in deactivate_slab(). Once an object's freepointer is corrupted, all following objects starting at this object are isolated. [akpm@linux-foundation.org: fix build with CONFIG_SLAB_DEBUG=n] Signed-off-by: Dongli Zhang <dongli.zhang@oracle.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Cc: Joe Jin <joe.jin@oracle.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Link: http://lkml.kernel.org/r/20200331031450.12182-1-dongli.zhang@oracle.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-02 12:45:47 +08:00
{
if ((s->flags & SLAB_CONSISTENCY_CHECKS) &&
!check_valid_pointer(s, page, nextfree) && freelist) {
object_err(s, page, *freelist, "Freechain corrupt");
*freelist = NULL;
mm/slub.c: fix corrupted freechain in deactivate_slab() The slub_debug is able to fix the corrupted slab freelist/page. However, alloc_debug_processing() only checks the validity of current and next freepointer during allocation path. As a result, once some objects have their freepointers corrupted, deactivate_slab() may lead to page fault. Below is from a test kernel module when 'slub_debug=PUF,kmalloc-128 slub_nomerge'. The test kernel corrupts the freepointer of one free object on purpose. Unfortunately, deactivate_slab() does not detect it when iterating the freechain. BUG: unable to handle page fault for address: 00000000123456f8 #PF: supervisor read access in kernel mode #PF: error_code(0x0000) - not-present page PGD 0 P4D 0 Oops: 0000 [#1] SMP PTI ... ... RIP: 0010:deactivate_slab.isra.92+0xed/0x490 ... ... Call Trace: ___slab_alloc+0x536/0x570 __slab_alloc+0x17/0x30 __kmalloc+0x1d9/0x200 ext4_htree_store_dirent+0x30/0xf0 htree_dirblock_to_tree+0xcb/0x1c0 ext4_htree_fill_tree+0x1bc/0x2d0 ext4_readdir+0x54f/0x920 iterate_dir+0x88/0x190 __x64_sys_getdents+0xa6/0x140 do_syscall_64+0x49/0x170 entry_SYSCALL_64_after_hwframe+0x44/0xa9 Therefore, this patch adds extra consistency check in deactivate_slab(). Once an object's freepointer is corrupted, all following objects starting at this object are isolated. [akpm@linux-foundation.org: fix build with CONFIG_SLAB_DEBUG=n] Signed-off-by: Dongli Zhang <dongli.zhang@oracle.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Cc: Joe Jin <joe.jin@oracle.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Link: http://lkml.kernel.org/r/20200331031450.12182-1-dongli.zhang@oracle.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-02 12:45:47 +08:00
slab_fix(s, "Isolate corrupted freechain");
return true;
}
return false;
}
static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
{
unsigned int off; /* Offset of last byte */
u8 *addr = page_address(page);
print_tracking(s, p);
print_page_info(page);
pr_err("Object 0x%p @offset=%tu fp=0x%p\n\n",
p, p - addr, get_freepointer(s, p));
if (s->flags & SLAB_RED_ZONE)
print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
s->red_left_pad);
else if (p > addr + 16)
print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
print_section(KERN_ERR, "Object ", p,
min_t(unsigned int, s->object_size, PAGE_SIZE));
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (s->flags & SLAB_RED_ZONE)
print_section(KERN_ERR, "Redzone ", p + s->object_size,
s->inuse - s->object_size);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
mm/slub: fix incorrect interpretation of s->offset In a couple of places in the slub memory allocator, the code uses "s->offset" as a check to see if the free pointer is put right after the object. That check is no longer true with commit 3202fa62fb43 ("slub: relocate freelist pointer to middle of object"). As a result, echoing "1" into the validate sysfs file, e.g. of dentry, may cause a bunch of "Freepointer corrupt" error reports like the following to appear with the system in panic afterwards. ============================================================================= BUG dentry(666:pmcd.service) (Tainted: G B): Freepointer corrupt ----------------------------------------------------------------------------- To fix it, use the check "s->offset == s->inuse" in the new helper function freeptr_outside_object() instead. Also add another helper function get_info_end() to return the end of info block (inuse + free pointer if not overlapping with object). Fixes: 3202fa62fb43 ("slub: relocate freelist pointer to middle of object") Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Acked-by: Rafael Aquini <aquini@redhat.com> Cc: Christoph Lameter <cl@linux.com> Cc: Vitaly Nikolenko <vnik@duasynt.com> Cc: Silvio Cesare <silvio.cesare@gmail.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Markus Elfring <Markus.Elfring@web.de> Cc: Changbin Du <changbin.du@gmail.com> Link: http://lkml.kernel.org/r/20200429135328.26976-1-longman@redhat.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-05-08 09:36:06 +08:00
off = get_info_end(s);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (s->flags & SLAB_STORE_USER)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
off += 2 * sizeof(struct track);
off += kasan_metadata_size(s);
if (off != size_from_object(s))
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/* Beginning of the filler is the free pointer */
print_section(KERN_ERR, "Padding ", p + off,
size_from_object(s) - off);
dump_stack();
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
void object_err(struct kmem_cache *s, struct page *page,
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
u8 *object, char *reason)
{
slab_bug(s, "%s", reason);
print_trailer(s, page, object);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
static __printf(3, 4) void slab_err(struct kmem_cache *s, struct page *page,
const char *fmt, ...)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
{
va_list args;
char buf[100];
va_start(args, fmt);
vsnprintf(buf, sizeof(buf), fmt, args);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
va_end(args);
slab_bug(s, "%s", buf);
print_page_info(page);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
dump_stack();
}
static void init_object(struct kmem_cache *s, void *object, u8 val)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
{
kasan, mm: reset tags when accessing metadata Kernel allocator code accesses metadata for slab objects, that may lie out-of-bounds of the object itself, or be accessed when an object is freed. Such accesses trigger tag faults and lead to false-positive reports with hardware tag-based KASAN. Software KASAN modes disable instrumentation for allocator code via KASAN_SANITIZE Makefile macro, and rely on kasan_enable/disable_current() annotations which are used to ignore KASAN reports. With hardware tag-based KASAN neither of those options are available, as it doesn't use compiler instrumetation, no tag faults are ignored, and MTE is disabled after the first one. Instead, reset tags when accessing metadata (currently only for SLUB). Link: https://lkml.kernel.org/r/a0f3cefbc49f34c843b664110842de4db28179d0.1606161801.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Signed-off-by: Vincenzo Frascino <vincenzo.frascino@arm.com> Acked-by: Marco Elver <elver@google.com> Reviewed-by: Alexander Potapenko <glider@google.com> Tested-by: Vincenzo Frascino <vincenzo.frascino@arm.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Branislav Rankov <Branislav.Rankov@arm.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Evgenii Stepanov <eugenis@google.com> Cc: Kevin Brodsky <kevin.brodsky@arm.com> Cc: Vasily Gorbik <gor@linux.ibm.com> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-23 04:02:17 +08:00
u8 *p = kasan_reset_tag(object);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (s->flags & SLAB_RED_ZONE)
memset(p - s->red_left_pad, val, s->red_left_pad);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (s->flags & __OBJECT_POISON) {
memset(p, POISON_FREE, s->object_size - 1);
p[s->object_size - 1] = POISON_END;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
if (s->flags & SLAB_RED_ZONE)
memset(p + s->object_size, val, s->inuse - s->object_size);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
void *from, void *to)
{
slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
memset(from, data, to - from);
}
static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
u8 *object, char *what,
u8 *start, unsigned int value, unsigned int bytes)
{
u8 *fault;
u8 *end;
mm: slub: print the offset of fault addresses With commit ad67b74d2469 ("printk: hash addresses printed with %p"), it is a little bit harder to match the fault addresses printed by check_bytes_and_report() or slab_pad_check() in the dump because the fault addresses may not show up in the dump. Print the offset of the fault addresses to make it easier to match the incorrect poison or padding values in the dump. Before: We have to search the "63" in the dump. If we want to get the offset of 63, we have to count it from the start of Object dump. ============================================================= BUG kmalloc-128 (Not tainted): Poison overwritten ------------------------------------------------------------- Disabling lock debugging due to kernel taint INFO: 0x00000000570da294-0x00000000570da294. First byte 0x63 instead of 0x6b ... INFO: Object 0x000000006ebb3b9e @offset=14208 fp=0x0000000065862488 Redzone 00000000a6abccff: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 00000000741c16f0: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 0000000061ad278f: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 000000000467c1bd: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 000000008812766b: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 000000003d9b8f25: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 0000000000d80c33: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 00000000867b0d90: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Object 000000006ebb3b9e: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 000000005ea59a9f: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 000000003ef8bddc: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 000000008190375d: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 000000006df7fb32: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 0000000069474eae: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 0000000008073b7d: 6b 6b 6b 6b 63 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 00000000b45ae74d: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b a5 After: We know the fault address is at @offset=1508, and the Object is at @offset=1408, so we know the fault address is at offset=100 within the object. ========================================================= BUG kmalloc-128 (Not tainted): Poison overwritten --------------------------------------------------------- Disabling lock debugging due to kernel taint INFO: 0x00000000638ec1d1-0x00000000638ec1d1 @offset=1508. First byte 0x63 instead of 0x6b ... INFO: Object 0x000000008171818d @offset=1408 fp=0x0000000066dae230 Redzone 00000000e2697ab6: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 0000000064b6a381: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 00000000e413a234: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 0000000004c1dfeb: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 000000009ad24d42: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 000000002a196a23: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 00000000a7b8468a: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 0000000088db6da3: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Object 000000008171818d: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 000000007c4035d4: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 000000004dd281a4: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 0000000079121dff: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 00000000756682a9: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 0000000053b7e541: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 0000000091f8d530: 6b 6b 6b 6b 63 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 000000009c76035c: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b a5 Link: http://lkml.kernel.org/r/20190925140807.20490-1-miles.chen@mediatek.com Signed-off-by: Miles Chen <miles.chen@mediatek.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Cc: Christoph Lameter <cl@linux.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: David Rientjes <rientjes@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:49:31 +08:00
u8 *addr = page_address(page);
mm: slub: introduce metadata_access_enable()/metadata_access_disable() It's ok for slub to access memory that marked by kasan as inaccessible (object's metadata). Kasan shouldn't print report in that case because these accesses are valid. Disabling instrumentation of slub.c code is not enough to achieve this because slub passes pointer to object's metadata into external functions like memchr_inv(). We don't want to disable instrumentation for memchr_inv() because this is quite generic function, and we don't want to miss bugs. metadata_access_enable/metadata_access_disable used to tell KASan where accesses to metadata starts/end, so we could temporarily disable KASan reports. Signed-off-by: Andrey Ryabinin <a.ryabinin@samsung.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Konstantin Serebryany <kcc@google.com> Cc: Dmitry Chernenkov <dmitryc@google.com> Signed-off-by: Andrey Konovalov <adech.fo@gmail.com> Cc: Yuri Gribov <tetra2005@gmail.com> Cc: Konstantin Khlebnikov <koct9i@gmail.com> Cc: Sasha Levin <sasha.levin@oracle.com> Cc: Christoph Lameter <cl@linux.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Ingo Molnar <mingo@elte.hu> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-14 06:39:38 +08:00
metadata_access_enable();
kasan, mm: reset tags when accessing metadata Kernel allocator code accesses metadata for slab objects, that may lie out-of-bounds of the object itself, or be accessed when an object is freed. Such accesses trigger tag faults and lead to false-positive reports with hardware tag-based KASAN. Software KASAN modes disable instrumentation for allocator code via KASAN_SANITIZE Makefile macro, and rely on kasan_enable/disable_current() annotations which are used to ignore KASAN reports. With hardware tag-based KASAN neither of those options are available, as it doesn't use compiler instrumetation, no tag faults are ignored, and MTE is disabled after the first one. Instead, reset tags when accessing metadata (currently only for SLUB). Link: https://lkml.kernel.org/r/a0f3cefbc49f34c843b664110842de4db28179d0.1606161801.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Signed-off-by: Vincenzo Frascino <vincenzo.frascino@arm.com> Acked-by: Marco Elver <elver@google.com> Reviewed-by: Alexander Potapenko <glider@google.com> Tested-by: Vincenzo Frascino <vincenzo.frascino@arm.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Branislav Rankov <Branislav.Rankov@arm.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Evgenii Stepanov <eugenis@google.com> Cc: Kevin Brodsky <kevin.brodsky@arm.com> Cc: Vasily Gorbik <gor@linux.ibm.com> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-23 04:02:17 +08:00
fault = memchr_inv(kasan_reset_tag(start), value, bytes);
mm: slub: introduce metadata_access_enable()/metadata_access_disable() It's ok for slub to access memory that marked by kasan as inaccessible (object's metadata). Kasan shouldn't print report in that case because these accesses are valid. Disabling instrumentation of slub.c code is not enough to achieve this because slub passes pointer to object's metadata into external functions like memchr_inv(). We don't want to disable instrumentation for memchr_inv() because this is quite generic function, and we don't want to miss bugs. metadata_access_enable/metadata_access_disable used to tell KASan where accesses to metadata starts/end, so we could temporarily disable KASan reports. Signed-off-by: Andrey Ryabinin <a.ryabinin@samsung.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Konstantin Serebryany <kcc@google.com> Cc: Dmitry Chernenkov <dmitryc@google.com> Signed-off-by: Andrey Konovalov <adech.fo@gmail.com> Cc: Yuri Gribov <tetra2005@gmail.com> Cc: Konstantin Khlebnikov <koct9i@gmail.com> Cc: Sasha Levin <sasha.levin@oracle.com> Cc: Christoph Lameter <cl@linux.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Ingo Molnar <mingo@elte.hu> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-14 06:39:38 +08:00
metadata_access_disable();
if (!fault)
return 1;
end = start + bytes;
while (end > fault && end[-1] == value)
end--;
slab_bug(s, "%s overwritten", what);
pr_err("0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
mm: slub: print the offset of fault addresses With commit ad67b74d2469 ("printk: hash addresses printed with %p"), it is a little bit harder to match the fault addresses printed by check_bytes_and_report() or slab_pad_check() in the dump because the fault addresses may not show up in the dump. Print the offset of the fault addresses to make it easier to match the incorrect poison or padding values in the dump. Before: We have to search the "63" in the dump. If we want to get the offset of 63, we have to count it from the start of Object dump. ============================================================= BUG kmalloc-128 (Not tainted): Poison overwritten ------------------------------------------------------------- Disabling lock debugging due to kernel taint INFO: 0x00000000570da294-0x00000000570da294. First byte 0x63 instead of 0x6b ... INFO: Object 0x000000006ebb3b9e @offset=14208 fp=0x0000000065862488 Redzone 00000000a6abccff: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 00000000741c16f0: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 0000000061ad278f: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 000000000467c1bd: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 000000008812766b: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 000000003d9b8f25: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 0000000000d80c33: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 00000000867b0d90: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Object 000000006ebb3b9e: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 000000005ea59a9f: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 000000003ef8bddc: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 000000008190375d: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 000000006df7fb32: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 0000000069474eae: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 0000000008073b7d: 6b 6b 6b 6b 63 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 00000000b45ae74d: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b a5 After: We know the fault address is at @offset=1508, and the Object is at @offset=1408, so we know the fault address is at offset=100 within the object. ========================================================= BUG kmalloc-128 (Not tainted): Poison overwritten --------------------------------------------------------- Disabling lock debugging due to kernel taint INFO: 0x00000000638ec1d1-0x00000000638ec1d1 @offset=1508. First byte 0x63 instead of 0x6b ... INFO: Object 0x000000008171818d @offset=1408 fp=0x0000000066dae230 Redzone 00000000e2697ab6: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 0000000064b6a381: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 00000000e413a234: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 0000000004c1dfeb: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 000000009ad24d42: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 000000002a196a23: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 00000000a7b8468a: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 0000000088db6da3: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Object 000000008171818d: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 000000007c4035d4: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 000000004dd281a4: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 0000000079121dff: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 00000000756682a9: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 0000000053b7e541: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 0000000091f8d530: 6b 6b 6b 6b 63 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 000000009c76035c: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b a5 Link: http://lkml.kernel.org/r/20190925140807.20490-1-miles.chen@mediatek.com Signed-off-by: Miles Chen <miles.chen@mediatek.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Cc: Christoph Lameter <cl@linux.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: David Rientjes <rientjes@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:49:31 +08:00
fault, end - 1, fault - addr,
fault[0], value);
print_trailer(s, page, object);
restore_bytes(s, what, value, fault, end);
return 0;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
/*
* Object layout:
*
* object address
* Bytes of the object to be managed.
* If the freepointer may overlay the object then the free
mm/slub: fix incorrect interpretation of s->offset In a couple of places in the slub memory allocator, the code uses "s->offset" as a check to see if the free pointer is put right after the object. That check is no longer true with commit 3202fa62fb43 ("slub: relocate freelist pointer to middle of object"). As a result, echoing "1" into the validate sysfs file, e.g. of dentry, may cause a bunch of "Freepointer corrupt" error reports like the following to appear with the system in panic afterwards. ============================================================================= BUG dentry(666:pmcd.service) (Tainted: G B): Freepointer corrupt ----------------------------------------------------------------------------- To fix it, use the check "s->offset == s->inuse" in the new helper function freeptr_outside_object() instead. Also add another helper function get_info_end() to return the end of info block (inuse + free pointer if not overlapping with object). Fixes: 3202fa62fb43 ("slub: relocate freelist pointer to middle of object") Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Acked-by: Rafael Aquini <aquini@redhat.com> Cc: Christoph Lameter <cl@linux.com> Cc: Vitaly Nikolenko <vnik@duasynt.com> Cc: Silvio Cesare <silvio.cesare@gmail.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Markus Elfring <Markus.Elfring@web.de> Cc: Changbin Du <changbin.du@gmail.com> Link: http://lkml.kernel.org/r/20200429135328.26976-1-longman@redhat.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-05-08 09:36:06 +08:00
* pointer is at the middle of the object.
*
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
* Poisoning uses 0x6b (POISON_FREE) and the last byte is
* 0xa5 (POISON_END)
*
* object + s->object_size
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
* Padding to reach word boundary. This is also used for Redzoning.
* Padding is extended by another word if Redzoning is enabled and
* object_size == inuse.
*
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
* We fill with 0xbb (RED_INACTIVE) for inactive objects and with
* 0xcc (RED_ACTIVE) for objects in use.
*
* object + s->inuse
* Meta data starts here.
*
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
* A. Free pointer (if we cannot overwrite object on free)
* B. Tracking data for SLAB_STORE_USER
* C. Padding to reach required alignment boundary or at minimum
* one word if debugging is on to be able to detect writes
* before the word boundary.
*
* Padding is done using 0x5a (POISON_INUSE)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*
* object + s->size
* Nothing is used beyond s->size.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*
* If slabcaches are merged then the object_size and inuse boundaries are mostly
* ignored. And therefore no slab options that rely on these boundaries
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
* may be used with merged slabcaches.
*/
static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
{
mm/slub: fix incorrect interpretation of s->offset In a couple of places in the slub memory allocator, the code uses "s->offset" as a check to see if the free pointer is put right after the object. That check is no longer true with commit 3202fa62fb43 ("slub: relocate freelist pointer to middle of object"). As a result, echoing "1" into the validate sysfs file, e.g. of dentry, may cause a bunch of "Freepointer corrupt" error reports like the following to appear with the system in panic afterwards. ============================================================================= BUG dentry(666:pmcd.service) (Tainted: G B): Freepointer corrupt ----------------------------------------------------------------------------- To fix it, use the check "s->offset == s->inuse" in the new helper function freeptr_outside_object() instead. Also add another helper function get_info_end() to return the end of info block (inuse + free pointer if not overlapping with object). Fixes: 3202fa62fb43 ("slub: relocate freelist pointer to middle of object") Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Acked-by: Rafael Aquini <aquini@redhat.com> Cc: Christoph Lameter <cl@linux.com> Cc: Vitaly Nikolenko <vnik@duasynt.com> Cc: Silvio Cesare <silvio.cesare@gmail.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Markus Elfring <Markus.Elfring@web.de> Cc: Changbin Du <changbin.du@gmail.com> Link: http://lkml.kernel.org/r/20200429135328.26976-1-longman@redhat.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-05-08 09:36:06 +08:00
unsigned long off = get_info_end(s); /* The end of info */
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (s->flags & SLAB_STORE_USER)
/* We also have user information there */
off += 2 * sizeof(struct track);
off += kasan_metadata_size(s);
if (size_from_object(s) == off)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
return 1;
return check_bytes_and_report(s, page, p, "Object padding",
p + off, POISON_INUSE, size_from_object(s) - off);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
/* Check the pad bytes at the end of a slab page */
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
static int slab_pad_check(struct kmem_cache *s, struct page *page)
{
u8 *start;
u8 *fault;
u8 *end;
u8 *pad;
int length;
int remainder;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (!(s->flags & SLAB_POISON))
return 1;
start = page_address(page);
length = page_size(page);
end = start + length;
remainder = length % s->size;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (!remainder)
return 1;
pad = end - remainder;
mm: slub: introduce metadata_access_enable()/metadata_access_disable() It's ok for slub to access memory that marked by kasan as inaccessible (object's metadata). Kasan shouldn't print report in that case because these accesses are valid. Disabling instrumentation of slub.c code is not enough to achieve this because slub passes pointer to object's metadata into external functions like memchr_inv(). We don't want to disable instrumentation for memchr_inv() because this is quite generic function, and we don't want to miss bugs. metadata_access_enable/metadata_access_disable used to tell KASan where accesses to metadata starts/end, so we could temporarily disable KASan reports. Signed-off-by: Andrey Ryabinin <a.ryabinin@samsung.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Konstantin Serebryany <kcc@google.com> Cc: Dmitry Chernenkov <dmitryc@google.com> Signed-off-by: Andrey Konovalov <adech.fo@gmail.com> Cc: Yuri Gribov <tetra2005@gmail.com> Cc: Konstantin Khlebnikov <koct9i@gmail.com> Cc: Sasha Levin <sasha.levin@oracle.com> Cc: Christoph Lameter <cl@linux.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Ingo Molnar <mingo@elte.hu> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-14 06:39:38 +08:00
metadata_access_enable();
kasan, mm: reset tags when accessing metadata Kernel allocator code accesses metadata for slab objects, that may lie out-of-bounds of the object itself, or be accessed when an object is freed. Such accesses trigger tag faults and lead to false-positive reports with hardware tag-based KASAN. Software KASAN modes disable instrumentation for allocator code via KASAN_SANITIZE Makefile macro, and rely on kasan_enable/disable_current() annotations which are used to ignore KASAN reports. With hardware tag-based KASAN neither of those options are available, as it doesn't use compiler instrumetation, no tag faults are ignored, and MTE is disabled after the first one. Instead, reset tags when accessing metadata (currently only for SLUB). Link: https://lkml.kernel.org/r/a0f3cefbc49f34c843b664110842de4db28179d0.1606161801.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Signed-off-by: Vincenzo Frascino <vincenzo.frascino@arm.com> Acked-by: Marco Elver <elver@google.com> Reviewed-by: Alexander Potapenko <glider@google.com> Tested-by: Vincenzo Frascino <vincenzo.frascino@arm.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Branislav Rankov <Branislav.Rankov@arm.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Evgenii Stepanov <eugenis@google.com> Cc: Kevin Brodsky <kevin.brodsky@arm.com> Cc: Vasily Gorbik <gor@linux.ibm.com> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-23 04:02:17 +08:00
fault = memchr_inv(kasan_reset_tag(pad), POISON_INUSE, remainder);
mm: slub: introduce metadata_access_enable()/metadata_access_disable() It's ok for slub to access memory that marked by kasan as inaccessible (object's metadata). Kasan shouldn't print report in that case because these accesses are valid. Disabling instrumentation of slub.c code is not enough to achieve this because slub passes pointer to object's metadata into external functions like memchr_inv(). We don't want to disable instrumentation for memchr_inv() because this is quite generic function, and we don't want to miss bugs. metadata_access_enable/metadata_access_disable used to tell KASan where accesses to metadata starts/end, so we could temporarily disable KASan reports. Signed-off-by: Andrey Ryabinin <a.ryabinin@samsung.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Konstantin Serebryany <kcc@google.com> Cc: Dmitry Chernenkov <dmitryc@google.com> Signed-off-by: Andrey Konovalov <adech.fo@gmail.com> Cc: Yuri Gribov <tetra2005@gmail.com> Cc: Konstantin Khlebnikov <koct9i@gmail.com> Cc: Sasha Levin <sasha.levin@oracle.com> Cc: Christoph Lameter <cl@linux.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Ingo Molnar <mingo@elte.hu> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-14 06:39:38 +08:00
metadata_access_disable();
if (!fault)
return 1;
while (end > fault && end[-1] == POISON_INUSE)
end--;
mm: slub: print the offset of fault addresses With commit ad67b74d2469 ("printk: hash addresses printed with %p"), it is a little bit harder to match the fault addresses printed by check_bytes_and_report() or slab_pad_check() in the dump because the fault addresses may not show up in the dump. Print the offset of the fault addresses to make it easier to match the incorrect poison or padding values in the dump. Before: We have to search the "63" in the dump. If we want to get the offset of 63, we have to count it from the start of Object dump. ============================================================= BUG kmalloc-128 (Not tainted): Poison overwritten ------------------------------------------------------------- Disabling lock debugging due to kernel taint INFO: 0x00000000570da294-0x00000000570da294. First byte 0x63 instead of 0x6b ... INFO: Object 0x000000006ebb3b9e @offset=14208 fp=0x0000000065862488 Redzone 00000000a6abccff: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 00000000741c16f0: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 0000000061ad278f: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 000000000467c1bd: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 000000008812766b: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 000000003d9b8f25: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 0000000000d80c33: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 00000000867b0d90: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Object 000000006ebb3b9e: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 000000005ea59a9f: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 000000003ef8bddc: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 000000008190375d: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 000000006df7fb32: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 0000000069474eae: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 0000000008073b7d: 6b 6b 6b 6b 63 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 00000000b45ae74d: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b a5 After: We know the fault address is at @offset=1508, and the Object is at @offset=1408, so we know the fault address is at offset=100 within the object. ========================================================= BUG kmalloc-128 (Not tainted): Poison overwritten --------------------------------------------------------- Disabling lock debugging due to kernel taint INFO: 0x00000000638ec1d1-0x00000000638ec1d1 @offset=1508. First byte 0x63 instead of 0x6b ... INFO: Object 0x000000008171818d @offset=1408 fp=0x0000000066dae230 Redzone 00000000e2697ab6: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 0000000064b6a381: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 00000000e413a234: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 0000000004c1dfeb: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 000000009ad24d42: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 000000002a196a23: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 00000000a7b8468a: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Redzone 0000000088db6da3: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb Object 000000008171818d: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 000000007c4035d4: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 000000004dd281a4: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 0000000079121dff: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 00000000756682a9: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 0000000053b7e541: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 0000000091f8d530: 6b 6b 6b 6b 63 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b Object 000000009c76035c: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b a5 Link: http://lkml.kernel.org/r/20190925140807.20490-1-miles.chen@mediatek.com Signed-off-by: Miles Chen <miles.chen@mediatek.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Cc: Christoph Lameter <cl@linux.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: David Rientjes <rientjes@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:49:31 +08:00
slab_err(s, page, "Padding overwritten. 0x%p-0x%p @offset=%tu",
fault, end - 1, fault - start);
print_section(KERN_ERR, "Padding ", pad, remainder);
restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
return 0;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
static int check_object(struct kmem_cache *s, struct page *page,
void *object, u8 val)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
{
u8 *p = object;
u8 *endobject = object + s->object_size;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (s->flags & SLAB_RED_ZONE) {
if (!check_bytes_and_report(s, page, object, "Redzone",
object - s->red_left_pad, val, s->red_left_pad))
return 0;
if (!check_bytes_and_report(s, page, object, "Redzone",
endobject, val, s->inuse - s->object_size))
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
return 0;
} else {
if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
check_bytes_and_report(s, page, p, "Alignment padding",
endobject, POISON_INUSE,
s->inuse - s->object_size);
}
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
if (s->flags & SLAB_POISON) {
if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
(!check_bytes_and_report(s, page, p, "Poison", p,
POISON_FREE, s->object_size - 1) ||
!check_bytes_and_report(s, page, p, "Poison",
p + s->object_size - 1, POISON_END, 1)))
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
return 0;
/*
* check_pad_bytes cleans up on its own.
*/
check_pad_bytes(s, page, p);
}
mm/slub: fix incorrect interpretation of s->offset In a couple of places in the slub memory allocator, the code uses "s->offset" as a check to see if the free pointer is put right after the object. That check is no longer true with commit 3202fa62fb43 ("slub: relocate freelist pointer to middle of object"). As a result, echoing "1" into the validate sysfs file, e.g. of dentry, may cause a bunch of "Freepointer corrupt" error reports like the following to appear with the system in panic afterwards. ============================================================================= BUG dentry(666:pmcd.service) (Tainted: G B): Freepointer corrupt ----------------------------------------------------------------------------- To fix it, use the check "s->offset == s->inuse" in the new helper function freeptr_outside_object() instead. Also add another helper function get_info_end() to return the end of info block (inuse + free pointer if not overlapping with object). Fixes: 3202fa62fb43 ("slub: relocate freelist pointer to middle of object") Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Acked-by: Rafael Aquini <aquini@redhat.com> Cc: Christoph Lameter <cl@linux.com> Cc: Vitaly Nikolenko <vnik@duasynt.com> Cc: Silvio Cesare <silvio.cesare@gmail.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Markus Elfring <Markus.Elfring@web.de> Cc: Changbin Du <changbin.du@gmail.com> Link: http://lkml.kernel.org/r/20200429135328.26976-1-longman@redhat.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-05-08 09:36:06 +08:00
if (!freeptr_outside_object(s) && val == SLUB_RED_ACTIVE)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/*
* Object and freepointer overlap. Cannot check
* freepointer while object is allocated.
*/
return 1;
/* Check free pointer validity */
if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
object_err(s, page, p, "Freepointer corrupt");
/*
* No choice but to zap it and thus lose the remainder
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
* of the free objects in this slab. May cause
* another error because the object count is now wrong.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*/
set_freepointer(s, p, NULL);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
return 0;
}
return 1;
}
static int check_slab(struct kmem_cache *s, struct page *page)
{
int maxobj;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
VM_BUG_ON(!irqs_disabled());
if (!PageSlab(page)) {
slab_err(s, page, "Not a valid slab page");
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
return 0;
}
maxobj = order_objects(compound_order(page), s->size);
if (page->objects > maxobj) {
slab_err(s, page, "objects %u > max %u",
page->objects, maxobj);
return 0;
}
if (page->inuse > page->objects) {
slab_err(s, page, "inuse %u > max %u",
page->inuse, page->objects);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
return 0;
}
/* Slab_pad_check fixes things up after itself */
slab_pad_check(s, page);
return 1;
}
/*
* Determine if a certain object on a page is on the freelist. Must hold the
* slab lock to guarantee that the chains are in a consistent state.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*/
static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
{
int nr = 0;
void *fp;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
void *object = NULL;
int max_objects;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
fp = page->freelist;
while (fp && nr <= page->objects) {
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (fp == search)
return 1;
if (!check_valid_pointer(s, page, fp)) {
if (object) {
object_err(s, page, object,
"Freechain corrupt");
set_freepointer(s, object, NULL);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
} else {
slab_err(s, page, "Freepointer corrupt");
page->freelist = NULL;
page->inuse = page->objects;
slab_fix(s, "Freelist cleared");
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
return 0;
}
break;
}
object = fp;
fp = get_freepointer(s, object);
nr++;
}
max_objects = order_objects(compound_order(page), s->size);
if (max_objects > MAX_OBJS_PER_PAGE)
max_objects = MAX_OBJS_PER_PAGE;
if (page->objects != max_objects) {
slab_err(s, page, "Wrong number of objects. Found %d but should be %d",
page->objects, max_objects);
page->objects = max_objects;
slab_fix(s, "Number of objects adjusted.");
}
if (page->inuse != page->objects - nr) {
slab_err(s, page, "Wrong object count. Counter is %d but counted were %d",
page->inuse, page->objects - nr);
page->inuse = page->objects - nr;
slab_fix(s, "Object count adjusted.");
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
return search == NULL;
}
static void trace(struct kmem_cache *s, struct page *page, void *object,
int alloc)
{
if (s->flags & SLAB_TRACE) {
pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
s->name,
alloc ? "alloc" : "free",
object, page->inuse,
page->freelist);
if (!alloc)
print_section(KERN_INFO, "Object ", (void *)object,
s->object_size);
dump_stack();
}
}
/*
* Tracking of fully allocated slabs for debugging purposes.
*/
static void add_full(struct kmem_cache *s,
struct kmem_cache_node *n, struct page *page)
{
if (!(s->flags & SLAB_STORE_USER))
return;
lockdep_assert_held(&n->list_lock);
list_add(&page->slab_list, &n->full);
}
static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
{
if (!(s->flags & SLAB_STORE_USER))
return;
lockdep_assert_held(&n->list_lock);
list_del(&page->slab_list);
}
/* Tracking of the number of slabs for debugging purposes */
static inline unsigned long slabs_node(struct kmem_cache *s, int node)
{
struct kmem_cache_node *n = get_node(s, node);
return atomic_long_read(&n->nr_slabs);
}
static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
{
return atomic_long_read(&n->nr_slabs);
}
static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
{
struct kmem_cache_node *n = get_node(s, node);
/*
* May be called early in order to allocate a slab for the
* kmem_cache_node structure. Solve the chicken-egg
* dilemma by deferring the increment of the count during
* bootstrap (see early_kmem_cache_node_alloc).
*/
if (likely(n)) {
atomic_long_inc(&n->nr_slabs);
atomic_long_add(objects, &n->total_objects);
}
}
static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
{
struct kmem_cache_node *n = get_node(s, node);
atomic_long_dec(&n->nr_slabs);
atomic_long_sub(objects, &n->total_objects);
}
/* Object debug checks for alloc/free paths */
static void setup_object_debug(struct kmem_cache *s, struct page *page,
void *object)
{
if (!kmem_cache_debug_flags(s, SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))
return;
init_object(s, object, SLUB_RED_INACTIVE);
init_tracking(s, object);
}
static
void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr)
{
if (!kmem_cache_debug_flags(s, SLAB_POISON))
return;
metadata_access_enable();
kasan, mm: reset tags when accessing metadata Kernel allocator code accesses metadata for slab objects, that may lie out-of-bounds of the object itself, or be accessed when an object is freed. Such accesses trigger tag faults and lead to false-positive reports with hardware tag-based KASAN. Software KASAN modes disable instrumentation for allocator code via KASAN_SANITIZE Makefile macro, and rely on kasan_enable/disable_current() annotations which are used to ignore KASAN reports. With hardware tag-based KASAN neither of those options are available, as it doesn't use compiler instrumetation, no tag faults are ignored, and MTE is disabled after the first one. Instead, reset tags when accessing metadata (currently only for SLUB). Link: https://lkml.kernel.org/r/a0f3cefbc49f34c843b664110842de4db28179d0.1606161801.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Signed-off-by: Vincenzo Frascino <vincenzo.frascino@arm.com> Acked-by: Marco Elver <elver@google.com> Reviewed-by: Alexander Potapenko <glider@google.com> Tested-by: Vincenzo Frascino <vincenzo.frascino@arm.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Branislav Rankov <Branislav.Rankov@arm.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Evgenii Stepanov <eugenis@google.com> Cc: Kevin Brodsky <kevin.brodsky@arm.com> Cc: Vasily Gorbik <gor@linux.ibm.com> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-23 04:02:17 +08:00
memset(kasan_reset_tag(addr), POISON_INUSE, page_size(page));
metadata_access_disable();
}
static inline int alloc_consistency_checks(struct kmem_cache *s,
struct page *page, void *object)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
{
if (!check_slab(s, page))
return 0;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (!check_valid_pointer(s, page, object)) {
object_err(s, page, object, "Freelist Pointer check fails");
return 0;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
if (!check_object(s, page, object, SLUB_RED_INACTIVE))
return 0;
return 1;
}
static noinline int alloc_debug_processing(struct kmem_cache *s,
struct page *page,
void *object, unsigned long addr)
{
if (s->flags & SLAB_CONSISTENCY_CHECKS) {
if (!alloc_consistency_checks(s, page, object))
goto bad;
}
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/* Success perform special debug activities for allocs */
if (s->flags & SLAB_STORE_USER)
set_track(s, object, TRACK_ALLOC, addr);
trace(s, page, object, 1);
init_object(s, object, SLUB_RED_ACTIVE);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
return 1;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
bad:
if (PageSlab(page)) {
/*
* If this is a slab page then lets do the best we can
* to avoid issues in the future. Marking all objects
* as used avoids touching the remaining objects.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*/
slab_fix(s, "Marking all objects used");
page->inuse = page->objects;
page->freelist = NULL;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
return 0;
}
static inline int free_consistency_checks(struct kmem_cache *s,
struct page *page, void *object, unsigned long addr)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
{
if (!check_valid_pointer(s, page, object)) {
slab_err(s, page, "Invalid object pointer 0x%p", object);
return 0;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
if (on_freelist(s, page, object)) {
object_err(s, page, object, "Object already free");
return 0;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
if (!check_object(s, page, object, SLUB_RED_ACTIVE))
return 0;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (unlikely(s != page->slab_cache)) {
if (!PageSlab(page)) {
slab_err(s, page, "Attempt to free object(0x%p) outside of slab",
object);
} else if (!page->slab_cache) {
pr_err("SLUB <none>: no slab for object 0x%p.\n",
object);
dump_stack();
} else
object_err(s, page, object,
"page slab pointer corrupt.");
return 0;
}
return 1;
}
/* Supports checking bulk free of a constructed freelist */
static noinline int free_debug_processing(
struct kmem_cache *s, struct page *page,
void *head, void *tail, int bulk_cnt,
unsigned long addr)
{
struct kmem_cache_node *n = get_node(s, page_to_nid(page));
void *object = head;
int cnt = 0;
treewide: Remove uninitialized_var() usage Using uninitialized_var() is dangerous as it papers over real bugs[1] (or can in the future), and suppresses unrelated compiler warnings (e.g. "unused variable"). If the compiler thinks it is uninitialized, either simply initialize the variable or make compiler changes. In preparation for removing[2] the[3] macro[4], remove all remaining needless uses with the following script: git grep '\buninitialized_var\b' | cut -d: -f1 | sort -u | \ xargs perl -pi -e \ 's/\buninitialized_var\(([^\)]+)\)/\1/g; s:\s*/\* (GCC be quiet|to make compiler happy) \*/$::g;' drivers/video/fbdev/riva/riva_hw.c was manually tweaked to avoid pathological white-space. No outstanding warnings were found building allmodconfig with GCC 9.3.0 for x86_64, i386, arm64, arm, powerpc, powerpc64le, s390x, mips, sparc64, alpha, and m68k. [1] https://lore.kernel.org/lkml/20200603174714.192027-1-glider@google.com/ [2] https://lore.kernel.org/lkml/CA+55aFw+Vbj0i=1TGqCR5vQkCzWJ0QxK6CernOU6eedsudAixw@mail.gmail.com/ [3] https://lore.kernel.org/lkml/CA+55aFwgbgqhbp1fkxvRKEpzyR5J8n1vKT1VZdz9knmPuXhOeg@mail.gmail.com/ [4] https://lore.kernel.org/lkml/CA+55aFz2500WfbKXAx8s67wrm9=yVJu65TpLgN_ybYNv0VEOKA@mail.gmail.com/ Reviewed-by: Leon Romanovsky <leonro@mellanox.com> # drivers/infiniband and mlx4/mlx5 Acked-by: Jason Gunthorpe <jgg@mellanox.com> # IB Acked-by: Kalle Valo <kvalo@codeaurora.org> # wireless drivers Reviewed-by: Chao Yu <yuchao0@huawei.com> # erofs Signed-off-by: Kees Cook <keescook@chromium.org>
2020-06-04 04:09:38 +08:00
unsigned long flags;
int ret = 0;
spin_lock_irqsave(&n->list_lock, flags);
slab_lock(page);
if (s->flags & SLAB_CONSISTENCY_CHECKS) {
if (!check_slab(s, page))
goto out;
}
next_object:
cnt++;
if (s->flags & SLAB_CONSISTENCY_CHECKS) {
if (!free_consistency_checks(s, page, object, addr))
goto out;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
if (s->flags & SLAB_STORE_USER)
set_track(s, object, TRACK_FREE, addr);
trace(s, page, object, 0);
/* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
init_object(s, object, SLUB_RED_INACTIVE);
/* Reached end of constructed freelist yet? */
if (object != tail) {
object = get_freepointer(s, object);
goto next_object;
}
ret = 1;
out:
if (cnt != bulk_cnt)
slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
bulk_cnt, cnt);
slab_unlock(page);
slub: drop lock at the end of free_debug_processing This series takes the suggestion of Christoph Lameter and only focuses on optimizing the slow path where the debug processing runs. The two main optimizations in this series are letting the consistency checks be skipped and relaxing the cmpxchg restrictions when we are not doing consistency checks. With hackbench -g 20 -l 1000 averaged over 100 runs: Before slub_debug=P mean 15.607 variance .086 stdev .294 After slub_debug=P mean 10.836 variance .155 stdev .394 This still isn't as fast as what is in grsecurity unfortunately so there's still work to be done. Profiling ___slab_alloc shows that 25-50% of time is spent in deactivate_slab. I haven't looked too closely to see if this is something that can be optimized. My plan for now is to focus on getting all of this merged (if appropriate) before digging in to another task. This patch (of 4): Currently, free_debug_processing has a comment "Keep node_lock to preserve integrity until the object is actually freed". In actuallity, the lock is dropped immediately in __slab_free. Rather than wait until __slab_free and potentially throw off the unlikely marking, just drop the lock in __slab_free. This also lets free_debug_processing take its own copy of the spinlock flags rather than trying to share the ones from __slab_free. Since there is no use for the node afterwards, change the return type of free_debug_processing to return an int like alloc_debug_processing. Credit to Mathias Krause for the original work which inspired this series [akpm@linux-foundation.org: fix build] Signed-off-by: Laura Abbott <labbott@fedoraproject.org> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <js1304@gmail.com> Cc: Kees Cook <keescook@chromium.org> Cc: Mathias Krause <minipli@googlemail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-03-16 05:54:59 +08:00
spin_unlock_irqrestore(&n->list_lock, flags);
if (!ret)
slab_fix(s, "Object at 0x%p not freed", object);
return ret;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
2020-08-07 14:18:35 +08:00
/*
* Parse a block of slub_debug options. Blocks are delimited by ';'
*
* @str: start of block
* @flags: returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified
* @slabs: return start of list of slabs, or NULL when there's no list
* @init: assume this is initial parsing and not per-kmem-create parsing
*
* returns the start of next block if there's any, or NULL
*/
static char *
parse_slub_debug_flags(char *str, slab_flags_t *flags, char **slabs, bool init)
{
2020-08-07 14:18:35 +08:00
bool higher_order_disable = false;
2020-08-07 14:18:35 +08:00
/* Skip any completely empty blocks */
while (*str && *str == ';')
str++;
if (*str == ',') {
/*
* No options but restriction on slabs. This means full
* debugging for slabs matching a pattern.
*/
2020-08-07 14:18:35 +08:00
*flags = DEBUG_DEFAULT_FLAGS;
goto check_slabs;
2020-08-07 14:18:35 +08:00
}
*flags = 0;
2020-08-07 14:18:35 +08:00
/* Determine which debug features should be switched on */
for (; *str && *str != ',' && *str != ';'; str++) {
switch (tolower(*str)) {
2020-08-07 14:18:35 +08:00
case '-':
*flags = 0;
break;
case 'f':
2020-08-07 14:18:35 +08:00
*flags |= SLAB_CONSISTENCY_CHECKS;
break;
case 'z':
2020-08-07 14:18:35 +08:00
*flags |= SLAB_RED_ZONE;
break;
case 'p':
2020-08-07 14:18:35 +08:00
*flags |= SLAB_POISON;
break;
case 'u':
2020-08-07 14:18:35 +08:00
*flags |= SLAB_STORE_USER;
break;
case 't':
2020-08-07 14:18:35 +08:00
*flags |= SLAB_TRACE;
break;
case 'a':
2020-08-07 14:18:35 +08:00
*flags |= SLAB_FAILSLAB;
break;
case 'o':
/*
* Avoid enabling debugging on caches if its minimum
* order would increase as a result.
*/
2020-08-07 14:18:35 +08:00
higher_order_disable = true;
break;
default:
2020-08-07 14:18:35 +08:00
if (init)
pr_err("slub_debug option '%c' unknown. skipped\n", *str);
}
}
check_slabs:
if (*str == ',')
2020-08-07 14:18:35 +08:00
*slabs = ++str;
else
*slabs = NULL;
/* Skip over the slab list */
while (*str && *str != ';')
str++;
/* Skip any completely empty blocks */
while (*str && *str == ';')
str++;
if (init && higher_order_disable)
disable_higher_order_debug = 1;
if (*str)
return str;
else
return NULL;
}
static int __init setup_slub_debug(char *str)
{
slab_flags_t flags;
char *saved_str;
char *slab_list;
bool global_slub_debug_changed = false;
bool slab_list_specified = false;
slub_debug = DEBUG_DEFAULT_FLAGS;
if (*str++ != '=' || !*str)
/*
* No options specified. Switch on full debugging.
*/
goto out;
saved_str = str;
while (str) {
str = parse_slub_debug_flags(str, &flags, &slab_list, true);
if (!slab_list) {
slub_debug = flags;
global_slub_debug_changed = true;
} else {
slab_list_specified = true;
}
}
/*
* For backwards compatibility, a single list of flags with list of
* slabs means debugging is only enabled for those slabs, so the global
* slub_debug should be 0. We can extended that to multiple lists as
* long as there is no option specifying flags without a slab list.
*/
if (slab_list_specified) {
if (!global_slub_debug_changed)
slub_debug = 0;
slub_debug_string = saved_str;
}
out:
if (slub_debug != 0 || slub_debug_string)
static_branch_enable(&slub_debug_enabled);
mm: security: introduce init_on_alloc=1 and init_on_free=1 boot options Patch series "add init_on_alloc/init_on_free boot options", v10. Provide init_on_alloc and init_on_free boot options. These are aimed at preventing possible information leaks and making the control-flow bugs that depend on uninitialized values more deterministic. Enabling either of the options guarantees that the memory returned by the page allocator and SL[AU]B is initialized with zeroes. SLOB allocator isn't supported at the moment, as its emulation of kmem caches complicates handling of SLAB_TYPESAFE_BY_RCU caches correctly. Enabling init_on_free also guarantees that pages and heap objects are initialized right after they're freed, so it won't be possible to access stale data by using a dangling pointer. As suggested by Michal Hocko, right now we don't let the heap users to disable initialization for certain allocations. There's not enough evidence that doing so can speed up real-life cases, and introducing ways to opt-out may result in things going out of control. This patch (of 2): The new options are needed to prevent possible information leaks and make control-flow bugs that depend on uninitialized values more deterministic. This is expected to be on-by-default on Android and Chrome OS. And it gives the opportunity for anyone else to use it under distros too via the boot args. (The init_on_free feature is regularly requested by folks where memory forensics is included in their threat models.) init_on_alloc=1 makes the kernel initialize newly allocated pages and heap objects with zeroes. Initialization is done at allocation time at the places where checks for __GFP_ZERO are performed. init_on_free=1 makes the kernel initialize freed pages and heap objects with zeroes upon their deletion. This helps to ensure sensitive data doesn't leak via use-after-free accesses. Both init_on_alloc=1 and init_on_free=1 guarantee that the allocator returns zeroed memory. The two exceptions are slab caches with constructors and SLAB_TYPESAFE_BY_RCU flag. Those are never zero-initialized to preserve their semantics. Both init_on_alloc and init_on_free default to zero, but those defaults can be overridden with CONFIG_INIT_ON_ALLOC_DEFAULT_ON and CONFIG_INIT_ON_FREE_DEFAULT_ON. If either SLUB poisoning or page poisoning is enabled, those options take precedence over init_on_alloc and init_on_free: initialization is only applied to unpoisoned allocations. Slowdown for the new features compared to init_on_free=0, init_on_alloc=0: hackbench, init_on_free=1: +7.62% sys time (st.err 0.74%) hackbench, init_on_alloc=1: +7.75% sys time (st.err 2.14%) Linux build with -j12, init_on_free=1: +8.38% wall time (st.err 0.39%) Linux build with -j12, init_on_free=1: +24.42% sys time (st.err 0.52%) Linux build with -j12, init_on_alloc=1: -0.13% wall time (st.err 0.42%) Linux build with -j12, init_on_alloc=1: +0.57% sys time (st.err 0.40%) The slowdown for init_on_free=0, init_on_alloc=0 compared to the baseline is within the standard error. The new features are also going to pave the way for hardware memory tagging (e.g. arm64's MTE), which will require both on_alloc and on_free hooks to set the tags for heap objects. With MTE, tagging will have the same cost as memory initialization. Although init_on_free is rather costly, there are paranoid use-cases where in-memory data lifetime is desired to be minimized. There are various arguments for/against the realism of the associated threat models, but given that we'll need the infrastructure for MTE anyway, and there are people who want wipe-on-free behavior no matter what the performance cost, it seems reasonable to include it in this series. [glider@google.com: v8] Link: http://lkml.kernel.org/r/20190626121943.131390-2-glider@google.com [glider@google.com: v9] Link: http://lkml.kernel.org/r/20190627130316.254309-2-glider@google.com [glider@google.com: v10] Link: http://lkml.kernel.org/r/20190628093131.199499-2-glider@google.com Link: http://lkml.kernel.org/r/20190617151050.92663-2-glider@google.com Signed-off-by: Alexander Potapenko <glider@google.com> Acked-by: Kees Cook <keescook@chromium.org> Acked-by: Michal Hocko <mhocko@suse.cz> [page and dmapool parts Acked-by: James Morris <jamorris@linux.microsoft.com>] Cc: Christoph Lameter <cl@linux.com> Cc: Masahiro Yamada <yamada.masahiro@socionext.com> Cc: "Serge E. Hallyn" <serge@hallyn.com> Cc: Nick Desaulniers <ndesaulniers@google.com> Cc: Kostya Serebryany <kcc@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Sandeep Patil <sspatil@android.com> Cc: Laura Abbott <labbott@redhat.com> Cc: Randy Dunlap <rdunlap@infradead.org> Cc: Jann Horn <jannh@google.com> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Marco Elver <elver@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 11:59:19 +08:00
if ((static_branch_unlikely(&init_on_alloc) ||
static_branch_unlikely(&init_on_free)) &&
(slub_debug & SLAB_POISON))
pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
return 1;
}
__setup("slub_debug", setup_slub_debug);
/*
* kmem_cache_flags - apply debugging options to the cache
* @object_size: the size of an object without meta data
* @flags: flags to set
* @name: name of the cache
*
* Debug option(s) are applied to @flags. In addition to the debug
* option(s), if a slab name (or multiple) is specified i.e.
* slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
* then only the select slabs will receive the debug option(s).
*/
slab_flags_t kmem_cache_flags(unsigned int object_size,
slab_flags_t flags, const char *name)
{
char *iter;
size_t len;
2020-08-07 14:18:35 +08:00
char *next_block;
slab_flags_t block_flags;
slab_flags_t slub_debug_local = slub_debug;
/*
* If the slab cache is for debugging (e.g. kmemleak) then
* don't store user (stack trace) information by default,
* but let the user enable it via the command line below.
*/
if (flags & SLAB_NOLEAKTRACE)
slub_debug_local &= ~SLAB_STORE_USER;
len = strlen(name);
2020-08-07 14:18:35 +08:00
next_block = slub_debug_string;
/* Go through all blocks of debug options, see if any matches our slab's name */
while (next_block) {
next_block = parse_slub_debug_flags(next_block, &block_flags, &iter, false);
if (!iter)
continue;
/* Found a block that has a slab list, search it */
while (*iter) {
char *end, *glob;
size_t cmplen;
end = strchrnul(iter, ',');
if (next_block && next_block < end)
end = next_block - 1;
glob = strnchr(iter, end - iter, '*');
if (glob)
cmplen = glob - iter;
else
cmplen = max_t(size_t, len, (end - iter));
2020-08-07 14:18:35 +08:00
if (!strncmp(name, iter, cmplen)) {
flags |= block_flags;
return flags;
}
2020-08-07 14:18:35 +08:00
if (!*end || *end == ';')
break;
iter = end + 1;
}
}
return flags | slub_debug_local;
}
#else /* !CONFIG_SLUB_DEBUG */
static inline void setup_object_debug(struct kmem_cache *s,
struct page *page, void *object) {}
static inline
void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr) {}
static inline int alloc_debug_processing(struct kmem_cache *s,
struct page *page, void *object, unsigned long addr) { return 0; }
slub: drop lock at the end of free_debug_processing This series takes the suggestion of Christoph Lameter and only focuses on optimizing the slow path where the debug processing runs. The two main optimizations in this series are letting the consistency checks be skipped and relaxing the cmpxchg restrictions when we are not doing consistency checks. With hackbench -g 20 -l 1000 averaged over 100 runs: Before slub_debug=P mean 15.607 variance .086 stdev .294 After slub_debug=P mean 10.836 variance .155 stdev .394 This still isn't as fast as what is in grsecurity unfortunately so there's still work to be done. Profiling ___slab_alloc shows that 25-50% of time is spent in deactivate_slab. I haven't looked too closely to see if this is something that can be optimized. My plan for now is to focus on getting all of this merged (if appropriate) before digging in to another task. This patch (of 4): Currently, free_debug_processing has a comment "Keep node_lock to preserve integrity until the object is actually freed". In actuallity, the lock is dropped immediately in __slab_free. Rather than wait until __slab_free and potentially throw off the unlikely marking, just drop the lock in __slab_free. This also lets free_debug_processing take its own copy of the spinlock flags rather than trying to share the ones from __slab_free. Since there is no use for the node afterwards, change the return type of free_debug_processing to return an int like alloc_debug_processing. Credit to Mathias Krause for the original work which inspired this series [akpm@linux-foundation.org: fix build] Signed-off-by: Laura Abbott <labbott@fedoraproject.org> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <js1304@gmail.com> Cc: Kees Cook <keescook@chromium.org> Cc: Mathias Krause <minipli@googlemail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-03-16 05:54:59 +08:00
static inline int free_debug_processing(
struct kmem_cache *s, struct page *page,
void *head, void *tail, int bulk_cnt,
slub: drop lock at the end of free_debug_processing This series takes the suggestion of Christoph Lameter and only focuses on optimizing the slow path where the debug processing runs. The two main optimizations in this series are letting the consistency checks be skipped and relaxing the cmpxchg restrictions when we are not doing consistency checks. With hackbench -g 20 -l 1000 averaged over 100 runs: Before slub_debug=P mean 15.607 variance .086 stdev .294 After slub_debug=P mean 10.836 variance .155 stdev .394 This still isn't as fast as what is in grsecurity unfortunately so there's still work to be done. Profiling ___slab_alloc shows that 25-50% of time is spent in deactivate_slab. I haven't looked too closely to see if this is something that can be optimized. My plan for now is to focus on getting all of this merged (if appropriate) before digging in to another task. This patch (of 4): Currently, free_debug_processing has a comment "Keep node_lock to preserve integrity until the object is actually freed". In actuallity, the lock is dropped immediately in __slab_free. Rather than wait until __slab_free and potentially throw off the unlikely marking, just drop the lock in __slab_free. This also lets free_debug_processing take its own copy of the spinlock flags rather than trying to share the ones from __slab_free. Since there is no use for the node afterwards, change the return type of free_debug_processing to return an int like alloc_debug_processing. Credit to Mathias Krause for the original work which inspired this series [akpm@linux-foundation.org: fix build] Signed-off-by: Laura Abbott <labbott@fedoraproject.org> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <js1304@gmail.com> Cc: Kees Cook <keescook@chromium.org> Cc: Mathias Krause <minipli@googlemail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-03-16 05:54:59 +08:00
unsigned long addr) { return 0; }
static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
{ return 1; }
static inline int check_object(struct kmem_cache *s, struct page *page,
void *object, u8 val) { return 1; }
static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
struct page *page) {}
static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
struct page *page) {}
slab_flags_t kmem_cache_flags(unsigned int object_size,
slab_flags_t flags, const char *name)
{
return flags;
}
#define slub_debug 0
#define disable_higher_order_debug 0
static inline unsigned long slabs_node(struct kmem_cache *s, int node)
{ return 0; }
static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
{ return 0; }
static inline void inc_slabs_node(struct kmem_cache *s, int node,
int objects) {}
static inline void dec_slabs_node(struct kmem_cache *s, int node,
int objects) {}
mm/slub.c: fix corrupted freechain in deactivate_slab() The slub_debug is able to fix the corrupted slab freelist/page. However, alloc_debug_processing() only checks the validity of current and next freepointer during allocation path. As a result, once some objects have their freepointers corrupted, deactivate_slab() may lead to page fault. Below is from a test kernel module when 'slub_debug=PUF,kmalloc-128 slub_nomerge'. The test kernel corrupts the freepointer of one free object on purpose. Unfortunately, deactivate_slab() does not detect it when iterating the freechain. BUG: unable to handle page fault for address: 00000000123456f8 #PF: supervisor read access in kernel mode #PF: error_code(0x0000) - not-present page PGD 0 P4D 0 Oops: 0000 [#1] SMP PTI ... ... RIP: 0010:deactivate_slab.isra.92+0xed/0x490 ... ... Call Trace: ___slab_alloc+0x536/0x570 __slab_alloc+0x17/0x30 __kmalloc+0x1d9/0x200 ext4_htree_store_dirent+0x30/0xf0 htree_dirblock_to_tree+0xcb/0x1c0 ext4_htree_fill_tree+0x1bc/0x2d0 ext4_readdir+0x54f/0x920 iterate_dir+0x88/0x190 __x64_sys_getdents+0xa6/0x140 do_syscall_64+0x49/0x170 entry_SYSCALL_64_after_hwframe+0x44/0xa9 Therefore, this patch adds extra consistency check in deactivate_slab(). Once an object's freepointer is corrupted, all following objects starting at this object are isolated. [akpm@linux-foundation.org: fix build with CONFIG_SLAB_DEBUG=n] Signed-off-by: Dongli Zhang <dongli.zhang@oracle.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Cc: Joe Jin <joe.jin@oracle.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Link: http://lkml.kernel.org/r/20200331031450.12182-1-dongli.zhang@oracle.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-02 12:45:47 +08:00
static bool freelist_corrupted(struct kmem_cache *s, struct page *page,
void **freelist, void *nextfree)
mm/slub.c: fix corrupted freechain in deactivate_slab() The slub_debug is able to fix the corrupted slab freelist/page. However, alloc_debug_processing() only checks the validity of current and next freepointer during allocation path. As a result, once some objects have their freepointers corrupted, deactivate_slab() may lead to page fault. Below is from a test kernel module when 'slub_debug=PUF,kmalloc-128 slub_nomerge'. The test kernel corrupts the freepointer of one free object on purpose. Unfortunately, deactivate_slab() does not detect it when iterating the freechain. BUG: unable to handle page fault for address: 00000000123456f8 #PF: supervisor read access in kernel mode #PF: error_code(0x0000) - not-present page PGD 0 P4D 0 Oops: 0000 [#1] SMP PTI ... ... RIP: 0010:deactivate_slab.isra.92+0xed/0x490 ... ... Call Trace: ___slab_alloc+0x536/0x570 __slab_alloc+0x17/0x30 __kmalloc+0x1d9/0x200 ext4_htree_store_dirent+0x30/0xf0 htree_dirblock_to_tree+0xcb/0x1c0 ext4_htree_fill_tree+0x1bc/0x2d0 ext4_readdir+0x54f/0x920 iterate_dir+0x88/0x190 __x64_sys_getdents+0xa6/0x140 do_syscall_64+0x49/0x170 entry_SYSCALL_64_after_hwframe+0x44/0xa9 Therefore, this patch adds extra consistency check in deactivate_slab(). Once an object's freepointer is corrupted, all following objects starting at this object are isolated. [akpm@linux-foundation.org: fix build with CONFIG_SLAB_DEBUG=n] Signed-off-by: Dongli Zhang <dongli.zhang@oracle.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Cc: Joe Jin <joe.jin@oracle.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Link: http://lkml.kernel.org/r/20200331031450.12182-1-dongli.zhang@oracle.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-02 12:45:47 +08:00
{
return false;
}
#endif /* CONFIG_SLUB_DEBUG */
/*
* Hooks for other subsystems that check memory allocations. In a typical
* production configuration these hooks all should produce no code at all.
*/
kasan, mm: change hooks signatures Patch series "kasan: add software tag-based mode for arm64", v13. This patchset adds a new software tag-based mode to KASAN [1]. (Initially this mode was called KHWASAN, but it got renamed, see the naming rationale at the end of this section). The plan is to implement HWASan [2] for the kernel with the incentive, that it's going to have comparable to KASAN performance, but in the same time consume much less memory, trading that off for somewhat imprecise bug detection and being supported only for arm64. The underlying ideas of the approach used by software tag-based KASAN are: 1. By using the Top Byte Ignore (TBI) arm64 CPU feature, we can store pointer tags in the top byte of each kernel pointer. 2. Using shadow memory, we can store memory tags for each chunk of kernel memory. 3. On each memory allocation, we can generate a random tag, embed it into the returned pointer and set the memory tags that correspond to this chunk of memory to the same value. 4. By using compiler instrumentation, before each memory access we can add a check that the pointer tag matches the tag of the memory that is being accessed. 5. On a tag mismatch we report an error. With this patchset the existing KASAN mode gets renamed to generic KASAN, with the word "generic" meaning that the implementation can be supported by any architecture as it is purely software. The new mode this patchset adds is called software tag-based KASAN. The word "tag-based" refers to the fact that this mode uses tags embedded into the top byte of kernel pointers and the TBI arm64 CPU feature that allows to dereference such pointers. The word "software" here means that shadow memory manipulation and tag checking on pointer dereference is done in software. As it is the only tag-based implementation right now, "software tag-based" KASAN is sometimes referred to as simply "tag-based" in this patchset. A potential expansion of this mode is a hardware tag-based mode, which would use hardware memory tagging support (announced by Arm [3]) instead of compiler instrumentation and manual shadow memory manipulation. Same as generic KASAN, software tag-based KASAN is strictly a debugging feature. [1] https://www.kernel.org/doc/html/latest/dev-tools/kasan.html [2] http://clang.llvm.org/docs/HardwareAssistedAddressSanitizerDesign.html [3] https://community.arm.com/processors/b/blog/posts/arm-a-profile-architecture-2018-developments-armv85a ====== Rationale On mobile devices generic KASAN's memory usage is significant problem. One of the main reasons to have tag-based KASAN is to be able to perform a similar set of checks as the generic one does, but with lower memory requirements. Comment from Vishwath Mohan <vishwath@google.com>: I don't have data on-hand, but anecdotally both ASAN and KASAN have proven problematic to enable for environments that don't tolerate the increased memory pressure well. This includes (a) Low-memory form factors - Wear, TV, Things, lower-tier phones like Go, (c) Connected components like Pixel's visual core [1]. These are both places I'd love to have a low(er) memory footprint option at my disposal. Comment from Evgenii Stepanov <eugenis@google.com>: Looking at a live Android device under load, slab (according to /proc/meminfo) + kernel stack take 8-10% available RAM (~350MB). KASAN's overhead of 2x - 3x on top of it is not insignificant. Not having this overhead enables near-production use - ex. running KASAN/KHWASAN kernel on a personal, daily-use device to catch bugs that do not reproduce in test configuration. These are the ones that often cost the most engineering time to track down. CPU overhead is bad, but generally tolerable. RAM is critical, in our experience. Once it gets low enough, OOM-killer makes your life miserable. [1] https://www.blog.google/products/pixel/pixel-visual-core-image-processing-and-machine-learning-pixel-2/ ====== Technical details Software tag-based KASAN mode is implemented in a very similar way to the generic one. This patchset essentially does the following: 1. TCR_TBI1 is set to enable Top Byte Ignore. 2. Shadow memory is used (with a different scale, 1:16, so each shadow byte corresponds to 16 bytes of kernel memory) to store memory tags. 3. All slab objects are aligned to shadow scale, which is 16 bytes. 4. All pointers returned from the slab allocator are tagged with a random tag and the corresponding shadow memory is poisoned with the same value. 5. Compiler instrumentation is used to insert tag checks. Either by calling callbacks or by inlining them (CONFIG_KASAN_OUTLINE and CONFIG_KASAN_INLINE flags are reused). 6. When a tag mismatch is detected in callback instrumentation mode KASAN simply prints a bug report. In case of inline instrumentation, clang inserts a brk instruction, and KASAN has it's own brk handler, which reports the bug. 7. The memory in between slab objects is marked with a reserved tag, and acts as a redzone. 8. When a slab object is freed it's marked with a reserved tag. Bug detection is imprecise for two reasons: 1. We won't catch some small out-of-bounds accesses, that fall into the same shadow cell, as the last byte of a slab object. 2. We only have 1 byte to store tags, which means we have a 1/256 probability of a tag match for an incorrect access (actually even slightly less due to reserved tag values). Despite that there's a particular type of bugs that tag-based KASAN can detect compared to generic KASAN: use-after-free after the object has been allocated by someone else. ====== Testing Some kernel developers voiced a concern that changing the top byte of kernel pointers may lead to subtle bugs that are difficult to discover. To address this concern deliberate testing has been performed. It doesn't seem feasible to do some kind of static checking to find potential issues with pointer tagging, so a dynamic approach was taken. All pointer comparisons/subtractions have been instrumented in an LLVM compiler pass and a kernel module that would print a bug report whenever two pointers with different tags are being compared/subtracted (ignoring comparisons with NULL pointers and with pointers obtained by casting an error code to a pointer type) has been used. Then the kernel has been booted in QEMU and on an Odroid C2 board and syzkaller has been run. This yielded the following results. The two places that look interesting are: is_vmalloc_addr in include/linux/mm.h is_kernel_rodata in mm/util.c Here we compare a pointer with some fixed untagged values to make sure that the pointer lies in a particular part of the kernel address space. Since tag-based KASAN doesn't add tags to pointers that belong to rodata or vmalloc regions, this should work as is. To make sure debug checks to those two functions that check that the result doesn't change whether we operate on pointers with or without untagging has been added. A few other cases that don't look that interesting: Comparing pointers to achieve unique sorting order of pointee objects (e.g. sorting locks addresses before performing a double lock): tty_ldisc_lock_pair_timeout in drivers/tty/tty_ldisc.c pipe_double_lock in fs/pipe.c unix_state_double_lock in net/unix/af_unix.c lock_two_nondirectories in fs/inode.c mutex_lock_double in kernel/events/core.c ep_cmp_ffd in fs/eventpoll.c fsnotify_compare_groups fs/notify/mark.c Nothing needs to be done here, since the tags embedded into pointers don't change, so the sorting order would still be unique. Checks that a pointer belongs to some particular allocation: is_sibling_entry in lib/radix-tree.c object_is_on_stack in include/linux/sched/task_stack.h Nothing needs to be done here either, since two pointers can only belong to the same allocation if they have the same tag. Overall, since the kernel boots and works, there are no critical bugs. As for the rest, the traditional kernel testing way (use until fails) is the only one that looks feasible. Another point here is that tag-based KASAN is available under a separate config option that needs to be deliberately enabled. Even though it might be used in a "near-production" environment to find bugs that are not found during fuzzing or running tests, it is still a debug tool. ====== Benchmarks The following numbers were collected on Odroid C2 board. Both generic and tag-based KASAN were used in inline instrumentation mode. Boot time [1]: * ~1.7 sec for clean kernel * ~5.0 sec for generic KASAN * ~5.0 sec for tag-based KASAN Network performance [2]: * 8.33 Gbits/sec for clean kernel * 3.17 Gbits/sec for generic KASAN * 2.85 Gbits/sec for tag-based KASAN Slab memory usage after boot [3]: * ~40 kb for clean kernel * ~105 kb (~260% overhead) for generic KASAN * ~47 kb (~20% overhead) for tag-based KASAN KASAN memory overhead consists of three main parts: 1. Increased slab memory usage due to redzones. 2. Shadow memory (the whole reserved once during boot). 3. Quaratine (grows gradually until some preset limit; the more the limit, the more the chance to detect a use-after-free). Comparing tag-based vs generic KASAN for each of these points: 1. 20% vs 260% overhead. 2. 1/16th vs 1/8th of physical memory. 3. Tag-based KASAN doesn't require quarantine. [1] Time before the ext4 driver is initialized. [2] Measured as `iperf -s & iperf -c 127.0.0.1 -t 30`. [3] Measured as `cat /proc/meminfo | grep Slab`. ====== Some notes A few notes: 1. The patchset can be found here: https://github.com/xairy/kasan-prototype/tree/khwasan 2. Building requires a recent Clang version (7.0.0 or later). 3. Stack instrumentation is not supported yet and will be added later. This patch (of 25): Tag-based KASAN changes the value of the top byte of pointers returned from the kernel allocation functions (such as kmalloc). This patch updates KASAN hooks signatures and their usage in SLAB and SLUB code to reflect that. Link: http://lkml.kernel.org/r/aec2b5e3973781ff8a6bb6760f8543643202c451.1544099024.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Reviewed-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Reviewed-by: Dmitry Vyukov <dvyukov@google.com> Cc: Christoph Lameter <cl@linux.com> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 16:29:37 +08:00
static inline void *kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
{
ptr = kasan_kmalloc_large(ptr, size, flags);
/* As ptr might get tagged, call kmemleak hook after KASAN. */
kmemleak_alloc(ptr, size, 1, flags);
return ptr;
}
static __always_inline void kfree_hook(void *x)
{
kmemleak_free(x);
kasan_kfree_large(x);
}
kasan, mm: integrate slab init_on_free with HW_TAGS This change uses the previously added memory initialization feature of HW_TAGS KASAN routines for slab memory when init_on_free is enabled. With this change, memory initialization memset() is no longer called when both HW_TAGS KASAN and init_on_free are enabled. Instead, memory is initialized in KASAN runtime. For SLUB, the memory initialization memset() is moved into slab_free_hook() that currently directly follows the initialization loop. A new argument is added to slab_free_hook() that indicates whether to initialize the memory or not. To avoid discrepancies with which memory gets initialized that can be caused by future changes, both KASAN hook and initialization memset() are put together and a warning comment is added. Combining setting allocation tags with memory initialization improves HW_TAGS KASAN performance when init_on_free is enabled. Link: https://lkml.kernel.org/r/190fd15c1886654afdec0d19ebebd5ade665b601.1615296150.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Reviewed-by: Marco Elver <elver@google.com> Cc: Alexander Potapenko <glider@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Branislav Rankov <Branislav.Rankov@arm.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christoph Lameter <cl@linux.com> Cc: David Rientjes <rientjes@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Evgenii Stepanov <eugenis@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kevin Brodsky <kevin.brodsky@arm.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Collingbourne <pcc@google.com> Cc: Vincenzo Frascino <vincenzo.frascino@arm.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-30 14:00:09 +08:00
static __always_inline bool slab_free_hook(struct kmem_cache *s,
void *x, bool init)
{
kmemleak_free_recursive(x, s->flags);
/*
* Trouble is that we may no longer disable interrupts in the fast path
* So in order to make the debug calls that expect irqs to be
* disabled we need to disable interrupts temporarily.
*/
#ifdef CONFIG_LOCKDEP
{
unsigned long flags;
local_irq_save(flags);
debug_check_no_locks_freed(x, s->object_size);
local_irq_restore(flags);
}
#endif
if (!(s->flags & SLAB_DEBUG_OBJECTS))
debug_check_no_obj_freed(x, s->object_size);
mm: slub: add kernel address sanitizer support for slub allocator With this patch kasan will be able to catch bugs in memory allocated by slub. Initially all objects in newly allocated slab page, marked as redzone. Later, when allocation of slub object happens, requested by caller number of bytes marked as accessible, and the rest of the object (including slub's metadata) marked as redzone (inaccessible). We also mark object as accessible if ksize was called for this object. There is some places in kernel where ksize function is called to inquire size of really allocated area. Such callers could validly access whole allocated memory, so it should be marked as accessible. Code in slub.c and slab_common.c files could validly access to object's metadata, so instrumentation for this files are disabled. Signed-off-by: Andrey Ryabinin <a.ryabinin@samsung.com> Signed-off-by: Dmitry Chernenkov <dmitryc@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Konstantin Serebryany <kcc@google.com> Signed-off-by: Andrey Konovalov <adech.fo@gmail.com> Cc: Yuri Gribov <tetra2005@gmail.com> Cc: Konstantin Khlebnikov <koct9i@gmail.com> Cc: Sasha Levin <sasha.levin@oracle.com> Cc: Christoph Lameter <cl@linux.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Ingo Molnar <mingo@elte.hu> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-14 06:39:42 +08:00
/* Use KCSAN to help debug racy use-after-free. */
if (!(s->flags & SLAB_TYPESAFE_BY_RCU))
__kcsan_check_access(x, s->object_size,
KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT);
kasan, mm: integrate slab init_on_free with HW_TAGS This change uses the previously added memory initialization feature of HW_TAGS KASAN routines for slab memory when init_on_free is enabled. With this change, memory initialization memset() is no longer called when both HW_TAGS KASAN and init_on_free are enabled. Instead, memory is initialized in KASAN runtime. For SLUB, the memory initialization memset() is moved into slab_free_hook() that currently directly follows the initialization loop. A new argument is added to slab_free_hook() that indicates whether to initialize the memory or not. To avoid discrepancies with which memory gets initialized that can be caused by future changes, both KASAN hook and initialization memset() are put together and a warning comment is added. Combining setting allocation tags with memory initialization improves HW_TAGS KASAN performance when init_on_free is enabled. Link: https://lkml.kernel.org/r/190fd15c1886654afdec0d19ebebd5ade665b601.1615296150.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Reviewed-by: Marco Elver <elver@google.com> Cc: Alexander Potapenko <glider@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Branislav Rankov <Branislav.Rankov@arm.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christoph Lameter <cl@linux.com> Cc: David Rientjes <rientjes@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Evgenii Stepanov <eugenis@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kevin Brodsky <kevin.brodsky@arm.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Collingbourne <pcc@google.com> Cc: Vincenzo Frascino <vincenzo.frascino@arm.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-30 14:00:09 +08:00
/*
* As memory initialization might be integrated into KASAN,
* kasan_slab_free and initialization memset's must be
* kept together to avoid discrepancies in behavior.
*
* The initialization memset's clear the object and the metadata,
* but don't touch the SLAB redzone.
*/
if (init) {
int rsize;
if (!kasan_has_integrated_init())
memset(kasan_reset_tag(x), 0, s->object_size);
rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad : 0;
memset((char *)kasan_reset_tag(x) + s->inuse, 0,
s->size - s->inuse - rsize);
}
/* KASAN might put x into memory quarantine, delaying its reuse. */
return kasan_slab_free(s, x, init);
}
kasan, slub: fix handling of kasan_slab_free hook The kasan_slab_free hook's return value denotes whether the reuse of a slab object must be delayed (e.g. when the object is put into memory qurantine). The current way SLUB handles this hook is by ignoring its return value and hardcoding checks similar (but not exactly the same) to the ones performed in kasan_slab_free, which is prone to making mistakes. The main difference between the hardcoded checks and the ones in kasan_slab_free is whether we want to perform a free in case when an invalid-free or a double-free was detected (we don't). This patch changes the way SLUB handles this by: 1. taking into account the return value of kasan_slab_free for each of the objects, that are being freed; 2. reconstructing the freelist of objects to exclude the ones, whose reuse must be delayed. [andreyknvl@google.com: eliminate unnecessary branch in slab_free] Link: http://lkml.kernel.org/r/a62759a2545fddf69b0c034547212ca1eb1b3ce2.1520359686.git.andreyknvl@google.com Link: http://lkml.kernel.org/r/083f58501e54731203801d899632d76175868e97.1519400992.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Acked-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Alexander Potapenko <glider@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Kostya Serebryany <kcc@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-04-11 07:30:31 +08:00
static inline bool slab_free_freelist_hook(struct kmem_cache *s,
void **head, void **tail)
{
mm: security: introduce init_on_alloc=1 and init_on_free=1 boot options Patch series "add init_on_alloc/init_on_free boot options", v10. Provide init_on_alloc and init_on_free boot options. These are aimed at preventing possible information leaks and making the control-flow bugs that depend on uninitialized values more deterministic. Enabling either of the options guarantees that the memory returned by the page allocator and SL[AU]B is initialized with zeroes. SLOB allocator isn't supported at the moment, as its emulation of kmem caches complicates handling of SLAB_TYPESAFE_BY_RCU caches correctly. Enabling init_on_free also guarantees that pages and heap objects are initialized right after they're freed, so it won't be possible to access stale data by using a dangling pointer. As suggested by Michal Hocko, right now we don't let the heap users to disable initialization for certain allocations. There's not enough evidence that doing so can speed up real-life cases, and introducing ways to opt-out may result in things going out of control. This patch (of 2): The new options are needed to prevent possible information leaks and make control-flow bugs that depend on uninitialized values more deterministic. This is expected to be on-by-default on Android and Chrome OS. And it gives the opportunity for anyone else to use it under distros too via the boot args. (The init_on_free feature is regularly requested by folks where memory forensics is included in their threat models.) init_on_alloc=1 makes the kernel initialize newly allocated pages and heap objects with zeroes. Initialization is done at allocation time at the places where checks for __GFP_ZERO are performed. init_on_free=1 makes the kernel initialize freed pages and heap objects with zeroes upon their deletion. This helps to ensure sensitive data doesn't leak via use-after-free accesses. Both init_on_alloc=1 and init_on_free=1 guarantee that the allocator returns zeroed memory. The two exceptions are slab caches with constructors and SLAB_TYPESAFE_BY_RCU flag. Those are never zero-initialized to preserve their semantics. Both init_on_alloc and init_on_free default to zero, but those defaults can be overridden with CONFIG_INIT_ON_ALLOC_DEFAULT_ON and CONFIG_INIT_ON_FREE_DEFAULT_ON. If either SLUB poisoning or page poisoning is enabled, those options take precedence over init_on_alloc and init_on_free: initialization is only applied to unpoisoned allocations. Slowdown for the new features compared to init_on_free=0, init_on_alloc=0: hackbench, init_on_free=1: +7.62% sys time (st.err 0.74%) hackbench, init_on_alloc=1: +7.75% sys time (st.err 2.14%) Linux build with -j12, init_on_free=1: +8.38% wall time (st.err 0.39%) Linux build with -j12, init_on_free=1: +24.42% sys time (st.err 0.52%) Linux build with -j12, init_on_alloc=1: -0.13% wall time (st.err 0.42%) Linux build with -j12, init_on_alloc=1: +0.57% sys time (st.err 0.40%) The slowdown for init_on_free=0, init_on_alloc=0 compared to the baseline is within the standard error. The new features are also going to pave the way for hardware memory tagging (e.g. arm64's MTE), which will require both on_alloc and on_free hooks to set the tags for heap objects. With MTE, tagging will have the same cost as memory initialization. Although init_on_free is rather costly, there are paranoid use-cases where in-memory data lifetime is desired to be minimized. There are various arguments for/against the realism of the associated threat models, but given that we'll need the infrastructure for MTE anyway, and there are people who want wipe-on-free behavior no matter what the performance cost, it seems reasonable to include it in this series. [glider@google.com: v8] Link: http://lkml.kernel.org/r/20190626121943.131390-2-glider@google.com [glider@google.com: v9] Link: http://lkml.kernel.org/r/20190627130316.254309-2-glider@google.com [glider@google.com: v10] Link: http://lkml.kernel.org/r/20190628093131.199499-2-glider@google.com Link: http://lkml.kernel.org/r/20190617151050.92663-2-glider@google.com Signed-off-by: Alexander Potapenko <glider@google.com> Acked-by: Kees Cook <keescook@chromium.org> Acked-by: Michal Hocko <mhocko@suse.cz> [page and dmapool parts Acked-by: James Morris <jamorris@linux.microsoft.com>] Cc: Christoph Lameter <cl@linux.com> Cc: Masahiro Yamada <yamada.masahiro@socionext.com> Cc: "Serge E. Hallyn" <serge@hallyn.com> Cc: Nick Desaulniers <ndesaulniers@google.com> Cc: Kostya Serebryany <kcc@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Sandeep Patil <sspatil@android.com> Cc: Laura Abbott <labbott@redhat.com> Cc: Randy Dunlap <rdunlap@infradead.org> Cc: Jann Horn <jannh@google.com> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Marco Elver <elver@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 11:59:19 +08:00
void *object;
void *next = *head;
void *old_tail = *tail ? *tail : *head;
mm, kfence: insert KFENCE hooks for SLUB Inserts KFENCE hooks into the SLUB allocator. To pass the originally requested size to KFENCE, add an argument 'orig_size' to slab_alloc*(). The additional argument is required to preserve the requested original size for kmalloc() allocations, which uses size classes (e.g. an allocation of 272 bytes will return an object of size 512). Therefore, kmem_cache::size does not represent the kmalloc-caller's requested size, and we must introduce the argument 'orig_size' to propagate the originally requested size to KFENCE. Without the originally requested size, we would not be able to detect out-of-bounds accesses for objects placed at the end of a KFENCE object page if that object is not equal to the kmalloc-size class it was bucketed into. When KFENCE is disabled, there is no additional overhead, since slab_alloc*() functions are __always_inline. Link: https://lkml.kernel.org/r/20201103175841.3495947-6-elver@google.com Signed-off-by: Marco Elver <elver@google.com> Signed-off-by: Alexander Potapenko <glider@google.com> Reviewed-by: Dmitry Vyukov <dvyukov@google.com> Reviewed-by: Jann Horn <jannh@google.com> Co-developed-by: Marco Elver <elver@google.com> Cc: Andrey Konovalov <andreyknvl@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christopher Lameter <cl@linux.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: David Rientjes <rientjes@google.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Hillf Danton <hdanton@sina.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Joern Engel <joern@purestorage.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kees Cook <keescook@chromium.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Paul E. McKenney <paulmck@kernel.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: SeongJae Park <sjpark@amazon.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-26 09:19:16 +08:00
if (is_kfence_address(next)) {
kasan, mm: integrate slab init_on_free with HW_TAGS This change uses the previously added memory initialization feature of HW_TAGS KASAN routines for slab memory when init_on_free is enabled. With this change, memory initialization memset() is no longer called when both HW_TAGS KASAN and init_on_free are enabled. Instead, memory is initialized in KASAN runtime. For SLUB, the memory initialization memset() is moved into slab_free_hook() that currently directly follows the initialization loop. A new argument is added to slab_free_hook() that indicates whether to initialize the memory or not. To avoid discrepancies with which memory gets initialized that can be caused by future changes, both KASAN hook and initialization memset() are put together and a warning comment is added. Combining setting allocation tags with memory initialization improves HW_TAGS KASAN performance when init_on_free is enabled. Link: https://lkml.kernel.org/r/190fd15c1886654afdec0d19ebebd5ade665b601.1615296150.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Reviewed-by: Marco Elver <elver@google.com> Cc: Alexander Potapenko <glider@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Branislav Rankov <Branislav.Rankov@arm.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christoph Lameter <cl@linux.com> Cc: David Rientjes <rientjes@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Evgenii Stepanov <eugenis@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kevin Brodsky <kevin.brodsky@arm.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Collingbourne <pcc@google.com> Cc: Vincenzo Frascino <vincenzo.frascino@arm.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-30 14:00:09 +08:00
slab_free_hook(s, next, false);
mm, kfence: insert KFENCE hooks for SLUB Inserts KFENCE hooks into the SLUB allocator. To pass the originally requested size to KFENCE, add an argument 'orig_size' to slab_alloc*(). The additional argument is required to preserve the requested original size for kmalloc() allocations, which uses size classes (e.g. an allocation of 272 bytes will return an object of size 512). Therefore, kmem_cache::size does not represent the kmalloc-caller's requested size, and we must introduce the argument 'orig_size' to propagate the originally requested size to KFENCE. Without the originally requested size, we would not be able to detect out-of-bounds accesses for objects placed at the end of a KFENCE object page if that object is not equal to the kmalloc-size class it was bucketed into. When KFENCE is disabled, there is no additional overhead, since slab_alloc*() functions are __always_inline. Link: https://lkml.kernel.org/r/20201103175841.3495947-6-elver@google.com Signed-off-by: Marco Elver <elver@google.com> Signed-off-by: Alexander Potapenko <glider@google.com> Reviewed-by: Dmitry Vyukov <dvyukov@google.com> Reviewed-by: Jann Horn <jannh@google.com> Co-developed-by: Marco Elver <elver@google.com> Cc: Andrey Konovalov <andreyknvl@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christopher Lameter <cl@linux.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: David Rientjes <rientjes@google.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Hillf Danton <hdanton@sina.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Joern Engel <joern@purestorage.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kees Cook <keescook@chromium.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Paul E. McKenney <paulmck@kernel.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: SeongJae Park <sjpark@amazon.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-26 09:19:16 +08:00
return true;
}
mm: slub: really fix slab walking for init_on_free Commit 1b7e816fc80e ("mm: slub: Fix slab walking for init_on_free") fixed one problem with the slab walking but missed a key detail: When walking the list, the head and tail pointers need to be updated since we end up reversing the list as a result. Without doing this, bulk free is broken. One way this is exposed is a NULL pointer with slub_debug=F: ============================================================================= BUG skbuff_head_cache (Tainted: G T): Object already free ----------------------------------------------------------------------------- INFO: Slab 0x000000000d2d2f8f objects=16 used=3 fp=0x0000000064309071 flags=0x3fff00000000201 BUG: kernel NULL pointer dereference, address: 0000000000000000 Oops: 0000 [#1] PREEMPT SMP PTI RIP: 0010:print_trailer+0x70/0x1d5 Call Trace: <IRQ> free_debug_processing.cold.37+0xc9/0x149 __slab_free+0x22a/0x3d0 kmem_cache_free_bulk+0x415/0x420 __kfree_skb_flush+0x30/0x40 net_rx_action+0x2dd/0x480 __do_softirq+0xf0/0x246 irq_exit+0x93/0xb0 do_IRQ+0xa0/0x110 common_interrupt+0xf/0xf </IRQ> Given we're now almost identical to the existing debugging code which correctly walks the list, combine with that. Link: https://lkml.kernel.org/r/20191104170303.GA50361@gandi.net Link: http://lkml.kernel.org/r/20191106222208.26815-1-labbott@redhat.com Fixes: 1b7e816fc80e ("mm: slub: Fix slab walking for init_on_free") Signed-off-by: Laura Abbott <labbott@redhat.com> Reported-by: Thibaut Sautereau <thibaut.sautereau@clip-os.org> Acked-by: David Rientjes <rientjes@google.com> Tested-by: Alexander Potapenko <glider@google.com> Acked-by: Alexander Potapenko <glider@google.com> Cc: Kees Cook <keescook@chromium.org> Cc: "David S. Miller" <davem@davemloft.net> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: <clipos@ssi.gouv.fr> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: <stable@vger.kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-11-16 09:34:50 +08:00
/* Head and tail of the reconstructed freelist */
*head = NULL;
*tail = NULL;
mm: slub: Fix slab walking for init_on_free To properly clear the slab on free with slab_want_init_on_free, we walk the list of free objects using get_freepointer/set_freepointer. The value we get from get_freepointer may not be valid. This isn't an issue since an actual value will get written later but this means there's a chance of triggering a bug if we use this value with set_freepointer: kernel BUG at mm/slub.c:306! invalid opcode: 0000 [#1] PREEMPT PTI CPU: 0 PID: 0 Comm: swapper Not tainted 5.2.0-05754-g6471384a #4 RIP: 0010:kfree+0x58a/0x5c0 Code: 48 83 05 78 37 51 02 01 0f 0b 48 83 05 7e 37 51 02 01 48 83 05 7e 37 51 02 01 48 83 05 7e 37 51 02 01 48 83 05 d6 37 51 02 01 <0f> 0b 48 83 05 d4 37 51 02 01 48 83 05 d4 37 51 02 01 48 83 05 d4 RSP: 0000:ffffffff82603d90 EFLAGS: 00010002 RAX: ffff8c3976c04320 RBX: ffff8c3976c04300 RCX: 0000000000000000 RDX: ffff8c3976c04300 RSI: 0000000000000000 RDI: ffff8c3976c04320 RBP: ffffffff82603db8 R08: 0000000000000000 R09: 0000000000000000 R10: ffff8c3976c04320 R11: ffffffff8289e1e0 R12: ffffd52cc8db0100 R13: ffff8c3976c01a00 R14: ffffffff810f10d4 R15: ffff8c3976c04300 FS: 0000000000000000(0000) GS:ffffffff8266b000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: ffff8c397ffff000 CR3: 0000000125020000 CR4: 00000000000406b0 Call Trace: apply_wqattrs_prepare+0x154/0x280 apply_workqueue_attrs_locked+0x4e/0xe0 apply_workqueue_attrs+0x36/0x60 alloc_workqueue+0x25a/0x6d0 workqueue_init_early+0x246/0x348 start_kernel+0x3c7/0x7ec x86_64_start_reservations+0x40/0x49 x86_64_start_kernel+0xda/0xe4 secondary_startup_64+0xb6/0xc0 Modules linked in: ---[ end trace f67eb9af4d8d492b ]--- Fix this by ensuring the value we set with set_freepointer is either NULL or another value in the chain. Reported-by: kernel test robot <rong.a.chen@intel.com> Signed-off-by: Laura Abbott <labbott@redhat.com> Fixes: 6471384af2a6 ("mm: security: introduce init_on_alloc=1 and init_on_free=1 boot options") Reviewed-by: Kees Cook <keescook@chromium.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-08-01 03:32:40 +08:00
mm: slub: really fix slab walking for init_on_free Commit 1b7e816fc80e ("mm: slub: Fix slab walking for init_on_free") fixed one problem with the slab walking but missed a key detail: When walking the list, the head and tail pointers need to be updated since we end up reversing the list as a result. Without doing this, bulk free is broken. One way this is exposed is a NULL pointer with slub_debug=F: ============================================================================= BUG skbuff_head_cache (Tainted: G T): Object already free ----------------------------------------------------------------------------- INFO: Slab 0x000000000d2d2f8f objects=16 used=3 fp=0x0000000064309071 flags=0x3fff00000000201 BUG: kernel NULL pointer dereference, address: 0000000000000000 Oops: 0000 [#1] PREEMPT SMP PTI RIP: 0010:print_trailer+0x70/0x1d5 Call Trace: <IRQ> free_debug_processing.cold.37+0xc9/0x149 __slab_free+0x22a/0x3d0 kmem_cache_free_bulk+0x415/0x420 __kfree_skb_flush+0x30/0x40 net_rx_action+0x2dd/0x480 __do_softirq+0xf0/0x246 irq_exit+0x93/0xb0 do_IRQ+0xa0/0x110 common_interrupt+0xf/0xf </IRQ> Given we're now almost identical to the existing debugging code which correctly walks the list, combine with that. Link: https://lkml.kernel.org/r/20191104170303.GA50361@gandi.net Link: http://lkml.kernel.org/r/20191106222208.26815-1-labbott@redhat.com Fixes: 1b7e816fc80e ("mm: slub: Fix slab walking for init_on_free") Signed-off-by: Laura Abbott <labbott@redhat.com> Reported-by: Thibaut Sautereau <thibaut.sautereau@clip-os.org> Acked-by: David Rientjes <rientjes@google.com> Tested-by: Alexander Potapenko <glider@google.com> Acked-by: Alexander Potapenko <glider@google.com> Cc: Kees Cook <keescook@chromium.org> Cc: "David S. Miller" <davem@davemloft.net> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: <clipos@ssi.gouv.fr> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: <stable@vger.kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-11-16 09:34:50 +08:00
do {
object = next;
next = get_freepointer(s, object);
kasan, slub: fix handling of kasan_slab_free hook The kasan_slab_free hook's return value denotes whether the reuse of a slab object must be delayed (e.g. when the object is put into memory qurantine). The current way SLUB handles this hook is by ignoring its return value and hardcoding checks similar (but not exactly the same) to the ones performed in kasan_slab_free, which is prone to making mistakes. The main difference between the hardcoded checks and the ones in kasan_slab_free is whether we want to perform a free in case when an invalid-free or a double-free was detected (we don't). This patch changes the way SLUB handles this by: 1. taking into account the return value of kasan_slab_free for each of the objects, that are being freed; 2. reconstructing the freelist of objects to exclude the ones, whose reuse must be delayed. [andreyknvl@google.com: eliminate unnecessary branch in slab_free] Link: http://lkml.kernel.org/r/a62759a2545fddf69b0c034547212ca1eb1b3ce2.1520359686.git.andreyknvl@google.com Link: http://lkml.kernel.org/r/083f58501e54731203801d899632d76175868e97.1519400992.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Acked-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Alexander Potapenko <glider@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Kostya Serebryany <kcc@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-04-11 07:30:31 +08:00
/* If object's reuse doesn't have to be delayed */
kasan, mm: integrate slab init_on_free with HW_TAGS This change uses the previously added memory initialization feature of HW_TAGS KASAN routines for slab memory when init_on_free is enabled. With this change, memory initialization memset() is no longer called when both HW_TAGS KASAN and init_on_free are enabled. Instead, memory is initialized in KASAN runtime. For SLUB, the memory initialization memset() is moved into slab_free_hook() that currently directly follows the initialization loop. A new argument is added to slab_free_hook() that indicates whether to initialize the memory or not. To avoid discrepancies with which memory gets initialized that can be caused by future changes, both KASAN hook and initialization memset() are put together and a warning comment is added. Combining setting allocation tags with memory initialization improves HW_TAGS KASAN performance when init_on_free is enabled. Link: https://lkml.kernel.org/r/190fd15c1886654afdec0d19ebebd5ade665b601.1615296150.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Reviewed-by: Marco Elver <elver@google.com> Cc: Alexander Potapenko <glider@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Branislav Rankov <Branislav.Rankov@arm.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christoph Lameter <cl@linux.com> Cc: David Rientjes <rientjes@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Evgenii Stepanov <eugenis@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kevin Brodsky <kevin.brodsky@arm.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Collingbourne <pcc@google.com> Cc: Vincenzo Frascino <vincenzo.frascino@arm.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-30 14:00:09 +08:00
if (!slab_free_hook(s, object, slab_want_init_on_free(s))) {
kasan, slub: fix handling of kasan_slab_free hook The kasan_slab_free hook's return value denotes whether the reuse of a slab object must be delayed (e.g. when the object is put into memory qurantine). The current way SLUB handles this hook is by ignoring its return value and hardcoding checks similar (but not exactly the same) to the ones performed in kasan_slab_free, which is prone to making mistakes. The main difference between the hardcoded checks and the ones in kasan_slab_free is whether we want to perform a free in case when an invalid-free or a double-free was detected (we don't). This patch changes the way SLUB handles this by: 1. taking into account the return value of kasan_slab_free for each of the objects, that are being freed; 2. reconstructing the freelist of objects to exclude the ones, whose reuse must be delayed. [andreyknvl@google.com: eliminate unnecessary branch in slab_free] Link: http://lkml.kernel.org/r/a62759a2545fddf69b0c034547212ca1eb1b3ce2.1520359686.git.andreyknvl@google.com Link: http://lkml.kernel.org/r/083f58501e54731203801d899632d76175868e97.1519400992.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Acked-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Alexander Potapenko <glider@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Kostya Serebryany <kcc@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-04-11 07:30:31 +08:00
/* Move object to the new freelist */
set_freepointer(s, object, *head);
*head = object;
if (!*tail)
*tail = object;
}
} while (object != old_tail);
if (*head == *tail)
*tail = NULL;
return *head != NULL;
}
static void *setup_object(struct kmem_cache *s, struct page *page,
void *object)
{
setup_object_debug(s, page, object);
object = kasan_init_slab_obj(s, object);
if (unlikely(s->ctor)) {
kasan_unpoison_object_data(s, object);
s->ctor(object);
kasan_poison_object_data(s, object);
}
return object;
}
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/*
* Slab allocation and freeing
*/
static inline struct page *alloc_slab_page(struct kmem_cache *s,
gfp_t flags, int node, struct kmem_cache_order_objects oo)
{
struct page *page;
unsigned int order = oo_order(oo);
if (node == NUMA_NO_NODE)
page = alloc_pages(flags, order);
else
mm: rename alloc_pages_exact_node() to __alloc_pages_node() alloc_pages_exact_node() was introduced in commit 6484eb3e2a81 ("page allocator: do not check NUMA node ID when the caller knows the node is valid") as an optimized variant of alloc_pages_node(), that doesn't fallback to current node for nid == NUMA_NO_NODE. Unfortunately the name of the function can easily suggest that the allocation is restricted to the given node and fails otherwise. In truth, the node is only preferred, unless __GFP_THISNODE is passed among the gfp flags. The misleading name has lead to mistakes in the past, see for example commits 5265047ac301 ("mm, thp: really limit transparent hugepage allocation to local node") and b360edb43f8e ("mm, mempolicy: migrate_to_node should only migrate to node"). Another issue with the name is that there's a family of alloc_pages_exact*() functions where 'exact' means exact size (instead of page order), which leads to more confusion. To prevent further mistakes, this patch effectively renames alloc_pages_exact_node() to __alloc_pages_node() to better convey that it's an optimized variant of alloc_pages_node() not intended for general usage. Both functions get described in comments. It has been also considered to really provide a convenience function for allocations restricted to a node, but the major opinion seems to be that __GFP_THISNODE already provides that functionality and we shouldn't duplicate the API needlessly. The number of users would be small anyway. Existing callers of alloc_pages_exact_node() are simply converted to call __alloc_pages_node(), with the exception of sba_alloc_coherent() which open-codes the check for NUMA_NO_NODE, so it is converted to use alloc_pages_node() instead. This means it no longer performs some VM_BUG_ON checks, and since the current check for nid in alloc_pages_node() uses a 'nid < 0' comparison (which includes NUMA_NO_NODE), it may hide wrong values which would be previously exposed. Both differences will be rectified by the next patch. To sum up, this patch makes no functional changes, except temporarily hiding potentially buggy callers. Restricting the checks in alloc_pages_node() is left for the next patch which can in turn expose more existing buggy callers. Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Acked-by: Johannes Weiner <hannes@cmpxchg.org> Acked-by: Robin Holt <robinmholt@gmail.com> Acked-by: Michal Hocko <mhocko@suse.com> Acked-by: Christoph Lameter <cl@linux.com> Acked-by: Michael Ellerman <mpe@ellerman.id.au> Cc: Mel Gorman <mgorman@suse.de> Cc: David Rientjes <rientjes@google.com> Cc: Greg Thelen <gthelen@google.com> Cc: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com> Cc: Tony Luck <tony.luck@intel.com> Cc: Fenghua Yu <fenghua.yu@intel.com> Cc: Arnd Bergmann <arnd@arndb.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: Gleb Natapov <gleb@kernel.org> Cc: Paolo Bonzini <pbonzini@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Ingo Molnar <mingo@redhat.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Cliff Whickman <cpw@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-09 06:03:50 +08:00
page = __alloc_pages_node(node, flags, order);
return page;
}
mm: SLUB freelist randomization Implements freelist randomization for the SLUB allocator. It was previous implemented for the SLAB allocator. Both use the same configuration option (CONFIG_SLAB_FREELIST_RANDOM). The list is randomized during initialization of a new set of pages. The order on different freelist sizes is pre-computed at boot for performance. Each kmem_cache has its own randomized freelist. This security feature reduces the predictability of the kernel SLUB allocator against heap overflows rendering attacks much less stable. For example these attacks exploit the predictability of the heap: - Linux Kernel CAN SLUB overflow (https://goo.gl/oMNWkU) - Exploiting Linux Kernel Heap corruptions (http://goo.gl/EXLn95) Performance results: slab_test impact is between 3% to 4% on average for 100000 attempts without smp. It is a very focused testing, kernbench show the overall impact on the system is way lower. Before: Single thread testing ===================== 1. Kmalloc: Repeatedly allocate then free test 100000 times kmalloc(8) -> 49 cycles kfree -> 77 cycles 100000 times kmalloc(16) -> 51 cycles kfree -> 79 cycles 100000 times kmalloc(32) -> 53 cycles kfree -> 83 cycles 100000 times kmalloc(64) -> 62 cycles kfree -> 90 cycles 100000 times kmalloc(128) -> 81 cycles kfree -> 97 cycles 100000 times kmalloc(256) -> 98 cycles kfree -> 121 cycles 100000 times kmalloc(512) -> 95 cycles kfree -> 122 cycles 100000 times kmalloc(1024) -> 96 cycles kfree -> 126 cycles 100000 times kmalloc(2048) -> 115 cycles kfree -> 140 cycles 100000 times kmalloc(4096) -> 149 cycles kfree -> 171 cycles 2. Kmalloc: alloc/free test 100000 times kmalloc(8)/kfree -> 70 cycles 100000 times kmalloc(16)/kfree -> 70 cycles 100000 times kmalloc(32)/kfree -> 70 cycles 100000 times kmalloc(64)/kfree -> 70 cycles 100000 times kmalloc(128)/kfree -> 70 cycles 100000 times kmalloc(256)/kfree -> 69 cycles 100000 times kmalloc(512)/kfree -> 70 cycles 100000 times kmalloc(1024)/kfree -> 73 cycles 100000 times kmalloc(2048)/kfree -> 72 cycles 100000 times kmalloc(4096)/kfree -> 71 cycles After: Single thread testing ===================== 1. Kmalloc: Repeatedly allocate then free test 100000 times kmalloc(8) -> 57 cycles kfree -> 78 cycles 100000 times kmalloc(16) -> 61 cycles kfree -> 81 cycles 100000 times kmalloc(32) -> 76 cycles kfree -> 93 cycles 100000 times kmalloc(64) -> 83 cycles kfree -> 94 cycles 100000 times kmalloc(128) -> 106 cycles kfree -> 107 cycles 100000 times kmalloc(256) -> 118 cycles kfree -> 117 cycles 100000 times kmalloc(512) -> 114 cycles kfree -> 116 cycles 100000 times kmalloc(1024) -> 115 cycles kfree -> 118 cycles 100000 times kmalloc(2048) -> 147 cycles kfree -> 131 cycles 100000 times kmalloc(4096) -> 214 cycles kfree -> 161 cycles 2. Kmalloc: alloc/free test 100000 times kmalloc(8)/kfree -> 66 cycles 100000 times kmalloc(16)/kfree -> 66 cycles 100000 times kmalloc(32)/kfree -> 66 cycles 100000 times kmalloc(64)/kfree -> 66 cycles 100000 times kmalloc(128)/kfree -> 65 cycles 100000 times kmalloc(256)/kfree -> 67 cycles 100000 times kmalloc(512)/kfree -> 67 cycles 100000 times kmalloc(1024)/kfree -> 64 cycles 100000 times kmalloc(2048)/kfree -> 67 cycles 100000 times kmalloc(4096)/kfree -> 67 cycles Kernbench, before: Average Optimal load -j 12 Run (std deviation): Elapsed Time 101.873 (1.16069) User Time 1045.22 (1.60447) System Time 88.969 (0.559195) Percent CPU 1112.9 (13.8279) Context Switches 189140 (2282.15) Sleeps 99008.6 (768.091) After: Average Optimal load -j 12 Run (std deviation): Elapsed Time 102.47 (0.562732) User Time 1045.3 (1.34263) System Time 88.311 (0.342554) Percent CPU 1105.8 (6.49444) Context Switches 189081 (2355.78) Sleeps 99231.5 (800.358) Link: http://lkml.kernel.org/r/1464295031-26375-3-git-send-email-thgarnie@google.com Signed-off-by: Thomas Garnier <thgarnie@google.com> Reviewed-by: Kees Cook <keescook@chromium.org> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-07-27 06:21:59 +08:00
#ifdef CONFIG_SLAB_FREELIST_RANDOM
/* Pre-initialize the random sequence cache */
static int init_cache_random_seq(struct kmem_cache *s)
{
unsigned int count = oo_objects(s->oo);
mm: SLUB freelist randomization Implements freelist randomization for the SLUB allocator. It was previous implemented for the SLAB allocator. Both use the same configuration option (CONFIG_SLAB_FREELIST_RANDOM). The list is randomized during initialization of a new set of pages. The order on different freelist sizes is pre-computed at boot for performance. Each kmem_cache has its own randomized freelist. This security feature reduces the predictability of the kernel SLUB allocator against heap overflows rendering attacks much less stable. For example these attacks exploit the predictability of the heap: - Linux Kernel CAN SLUB overflow (https://goo.gl/oMNWkU) - Exploiting Linux Kernel Heap corruptions (http://goo.gl/EXLn95) Performance results: slab_test impact is between 3% to 4% on average for 100000 attempts without smp. It is a very focused testing, kernbench show the overall impact on the system is way lower. Before: Single thread testing ===================== 1. Kmalloc: Repeatedly allocate then free test 100000 times kmalloc(8) -> 49 cycles kfree -> 77 cycles 100000 times kmalloc(16) -> 51 cycles kfree -> 79 cycles 100000 times kmalloc(32) -> 53 cycles kfree -> 83 cycles 100000 times kmalloc(64) -> 62 cycles kfree -> 90 cycles 100000 times kmalloc(128) -> 81 cycles kfree -> 97 cycles 100000 times kmalloc(256) -> 98 cycles kfree -> 121 cycles 100000 times kmalloc(512) -> 95 cycles kfree -> 122 cycles 100000 times kmalloc(1024) -> 96 cycles kfree -> 126 cycles 100000 times kmalloc(2048) -> 115 cycles kfree -> 140 cycles 100000 times kmalloc(4096) -> 149 cycles kfree -> 171 cycles 2. Kmalloc: alloc/free test 100000 times kmalloc(8)/kfree -> 70 cycles 100000 times kmalloc(16)/kfree -> 70 cycles 100000 times kmalloc(32)/kfree -> 70 cycles 100000 times kmalloc(64)/kfree -> 70 cycles 100000 times kmalloc(128)/kfree -> 70 cycles 100000 times kmalloc(256)/kfree -> 69 cycles 100000 times kmalloc(512)/kfree -> 70 cycles 100000 times kmalloc(1024)/kfree -> 73 cycles 100000 times kmalloc(2048)/kfree -> 72 cycles 100000 times kmalloc(4096)/kfree -> 71 cycles After: Single thread testing ===================== 1. Kmalloc: Repeatedly allocate then free test 100000 times kmalloc(8) -> 57 cycles kfree -> 78 cycles 100000 times kmalloc(16) -> 61 cycles kfree -> 81 cycles 100000 times kmalloc(32) -> 76 cycles kfree -> 93 cycles 100000 times kmalloc(64) -> 83 cycles kfree -> 94 cycles 100000 times kmalloc(128) -> 106 cycles kfree -> 107 cycles 100000 times kmalloc(256) -> 118 cycles kfree -> 117 cycles 100000 times kmalloc(512) -> 114 cycles kfree -> 116 cycles 100000 times kmalloc(1024) -> 115 cycles kfree -> 118 cycles 100000 times kmalloc(2048) -> 147 cycles kfree -> 131 cycles 100000 times kmalloc(4096) -> 214 cycles kfree -> 161 cycles 2. Kmalloc: alloc/free test 100000 times kmalloc(8)/kfree -> 66 cycles 100000 times kmalloc(16)/kfree -> 66 cycles 100000 times kmalloc(32)/kfree -> 66 cycles 100000 times kmalloc(64)/kfree -> 66 cycles 100000 times kmalloc(128)/kfree -> 65 cycles 100000 times kmalloc(256)/kfree -> 67 cycles 100000 times kmalloc(512)/kfree -> 67 cycles 100000 times kmalloc(1024)/kfree -> 64 cycles 100000 times kmalloc(2048)/kfree -> 67 cycles 100000 times kmalloc(4096)/kfree -> 67 cycles Kernbench, before: Average Optimal load -j 12 Run (std deviation): Elapsed Time 101.873 (1.16069) User Time 1045.22 (1.60447) System Time 88.969 (0.559195) Percent CPU 1112.9 (13.8279) Context Switches 189140 (2282.15) Sleeps 99008.6 (768.091) After: Average Optimal load -j 12 Run (std deviation): Elapsed Time 102.47 (0.562732) User Time 1045.3 (1.34263) System Time 88.311 (0.342554) Percent CPU 1105.8 (6.49444) Context Switches 189081 (2355.78) Sleeps 99231.5 (800.358) Link: http://lkml.kernel.org/r/1464295031-26375-3-git-send-email-thgarnie@google.com Signed-off-by: Thomas Garnier <thgarnie@google.com> Reviewed-by: Kees Cook <keescook@chromium.org> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-07-27 06:21:59 +08:00
int err;
/* Bailout if already initialised */
if (s->random_seq)
return 0;
mm: SLUB freelist randomization Implements freelist randomization for the SLUB allocator. It was previous implemented for the SLAB allocator. Both use the same configuration option (CONFIG_SLAB_FREELIST_RANDOM). The list is randomized during initialization of a new set of pages. The order on different freelist sizes is pre-computed at boot for performance. Each kmem_cache has its own randomized freelist. This security feature reduces the predictability of the kernel SLUB allocator against heap overflows rendering attacks much less stable. For example these attacks exploit the predictability of the heap: - Linux Kernel CAN SLUB overflow (https://goo.gl/oMNWkU) - Exploiting Linux Kernel Heap corruptions (http://goo.gl/EXLn95) Performance results: slab_test impact is between 3% to 4% on average for 100000 attempts without smp. It is a very focused testing, kernbench show the overall impact on the system is way lower. Before: Single thread testing ===================== 1. Kmalloc: Repeatedly allocate then free test 100000 times kmalloc(8) -> 49 cycles kfree -> 77 cycles 100000 times kmalloc(16) -> 51 cycles kfree -> 79 cycles 100000 times kmalloc(32) -> 53 cycles kfree -> 83 cycles 100000 times kmalloc(64) -> 62 cycles kfree -> 90 cycles 100000 times kmalloc(128) -> 81 cycles kfree -> 97 cycles 100000 times kmalloc(256) -> 98 cycles kfree -> 121 cycles 100000 times kmalloc(512) -> 95 cycles kfree -> 122 cycles 100000 times kmalloc(1024) -> 96 cycles kfree -> 126 cycles 100000 times kmalloc(2048) -> 115 cycles kfree -> 140 cycles 100000 times kmalloc(4096) -> 149 cycles kfree -> 171 cycles 2. Kmalloc: alloc/free test 100000 times kmalloc(8)/kfree -> 70 cycles 100000 times kmalloc(16)/kfree -> 70 cycles 100000 times kmalloc(32)/kfree -> 70 cycles 100000 times kmalloc(64)/kfree -> 70 cycles 100000 times kmalloc(128)/kfree -> 70 cycles 100000 times kmalloc(256)/kfree -> 69 cycles 100000 times kmalloc(512)/kfree -> 70 cycles 100000 times kmalloc(1024)/kfree -> 73 cycles 100000 times kmalloc(2048)/kfree -> 72 cycles 100000 times kmalloc(4096)/kfree -> 71 cycles After: Single thread testing ===================== 1. Kmalloc: Repeatedly allocate then free test 100000 times kmalloc(8) -> 57 cycles kfree -> 78 cycles 100000 times kmalloc(16) -> 61 cycles kfree -> 81 cycles 100000 times kmalloc(32) -> 76 cycles kfree -> 93 cycles 100000 times kmalloc(64) -> 83 cycles kfree -> 94 cycles 100000 times kmalloc(128) -> 106 cycles kfree -> 107 cycles 100000 times kmalloc(256) -> 118 cycles kfree -> 117 cycles 100000 times kmalloc(512) -> 114 cycles kfree -> 116 cycles 100000 times kmalloc(1024) -> 115 cycles kfree -> 118 cycles 100000 times kmalloc(2048) -> 147 cycles kfree -> 131 cycles 100000 times kmalloc(4096) -> 214 cycles kfree -> 161 cycles 2. Kmalloc: alloc/free test 100000 times kmalloc(8)/kfree -> 66 cycles 100000 times kmalloc(16)/kfree -> 66 cycles 100000 times kmalloc(32)/kfree -> 66 cycles 100000 times kmalloc(64)/kfree -> 66 cycles 100000 times kmalloc(128)/kfree -> 65 cycles 100000 times kmalloc(256)/kfree -> 67 cycles 100000 times kmalloc(512)/kfree -> 67 cycles 100000 times kmalloc(1024)/kfree -> 64 cycles 100000 times kmalloc(2048)/kfree -> 67 cycles 100000 times kmalloc(4096)/kfree -> 67 cycles Kernbench, before: Average Optimal load -j 12 Run (std deviation): Elapsed Time 101.873 (1.16069) User Time 1045.22 (1.60447) System Time 88.969 (0.559195) Percent CPU 1112.9 (13.8279) Context Switches 189140 (2282.15) Sleeps 99008.6 (768.091) After: Average Optimal load -j 12 Run (std deviation): Elapsed Time 102.47 (0.562732) User Time 1045.3 (1.34263) System Time 88.311 (0.342554) Percent CPU 1105.8 (6.49444) Context Switches 189081 (2355.78) Sleeps 99231.5 (800.358) Link: http://lkml.kernel.org/r/1464295031-26375-3-git-send-email-thgarnie@google.com Signed-off-by: Thomas Garnier <thgarnie@google.com> Reviewed-by: Kees Cook <keescook@chromium.org> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-07-27 06:21:59 +08:00
err = cache_random_seq_create(s, count, GFP_KERNEL);
if (err) {
pr_err("SLUB: Unable to initialize free list for %s\n",
s->name);
return err;
}
/* Transform to an offset on the set of pages */
if (s->random_seq) {
unsigned int i;
mm: SLUB freelist randomization Implements freelist randomization for the SLUB allocator. It was previous implemented for the SLAB allocator. Both use the same configuration option (CONFIG_SLAB_FREELIST_RANDOM). The list is randomized during initialization of a new set of pages. The order on different freelist sizes is pre-computed at boot for performance. Each kmem_cache has its own randomized freelist. This security feature reduces the predictability of the kernel SLUB allocator against heap overflows rendering attacks much less stable. For example these attacks exploit the predictability of the heap: - Linux Kernel CAN SLUB overflow (https://goo.gl/oMNWkU) - Exploiting Linux Kernel Heap corruptions (http://goo.gl/EXLn95) Performance results: slab_test impact is between 3% to 4% on average for 100000 attempts without smp. It is a very focused testing, kernbench show the overall impact on the system is way lower. Before: Single thread testing ===================== 1. Kmalloc: Repeatedly allocate then free test 100000 times kmalloc(8) -> 49 cycles kfree -> 77 cycles 100000 times kmalloc(16) -> 51 cycles kfree -> 79 cycles 100000 times kmalloc(32) -> 53 cycles kfree -> 83 cycles 100000 times kmalloc(64) -> 62 cycles kfree -> 90 cycles 100000 times kmalloc(128) -> 81 cycles kfree -> 97 cycles 100000 times kmalloc(256) -> 98 cycles kfree -> 121 cycles 100000 times kmalloc(512) -> 95 cycles kfree -> 122 cycles 100000 times kmalloc(1024) -> 96 cycles kfree -> 126 cycles 100000 times kmalloc(2048) -> 115 cycles kfree -> 140 cycles 100000 times kmalloc(4096) -> 149 cycles kfree -> 171 cycles 2. Kmalloc: alloc/free test 100000 times kmalloc(8)/kfree -> 70 cycles 100000 times kmalloc(16)/kfree -> 70 cycles 100000 times kmalloc(32)/kfree -> 70 cycles 100000 times kmalloc(64)/kfree -> 70 cycles 100000 times kmalloc(128)/kfree -> 70 cycles 100000 times kmalloc(256)/kfree -> 69 cycles 100000 times kmalloc(512)/kfree -> 70 cycles 100000 times kmalloc(1024)/kfree -> 73 cycles 100000 times kmalloc(2048)/kfree -> 72 cycles 100000 times kmalloc(4096)/kfree -> 71 cycles After: Single thread testing ===================== 1. Kmalloc: Repeatedly allocate then free test 100000 times kmalloc(8) -> 57 cycles kfree -> 78 cycles 100000 times kmalloc(16) -> 61 cycles kfree -> 81 cycles 100000 times kmalloc(32) -> 76 cycles kfree -> 93 cycles 100000 times kmalloc(64) -> 83 cycles kfree -> 94 cycles 100000 times kmalloc(128) -> 106 cycles kfree -> 107 cycles 100000 times kmalloc(256) -> 118 cycles kfree -> 117 cycles 100000 times kmalloc(512) -> 114 cycles kfree -> 116 cycles 100000 times kmalloc(1024) -> 115 cycles kfree -> 118 cycles 100000 times kmalloc(2048) -> 147 cycles kfree -> 131 cycles 100000 times kmalloc(4096) -> 214 cycles kfree -> 161 cycles 2. Kmalloc: alloc/free test 100000 times kmalloc(8)/kfree -> 66 cycles 100000 times kmalloc(16)/kfree -> 66 cycles 100000 times kmalloc(32)/kfree -> 66 cycles 100000 times kmalloc(64)/kfree -> 66 cycles 100000 times kmalloc(128)/kfree -> 65 cycles 100000 times kmalloc(256)/kfree -> 67 cycles 100000 times kmalloc(512)/kfree -> 67 cycles 100000 times kmalloc(1024)/kfree -> 64 cycles 100000 times kmalloc(2048)/kfree -> 67 cycles 100000 times kmalloc(4096)/kfree -> 67 cycles Kernbench, before: Average Optimal load -j 12 Run (std deviation): Elapsed Time 101.873 (1.16069) User Time 1045.22 (1.60447) System Time 88.969 (0.559195) Percent CPU 1112.9 (13.8279) Context Switches 189140 (2282.15) Sleeps 99008.6 (768.091) After: Average Optimal load -j 12 Run (std deviation): Elapsed Time 102.47 (0.562732) User Time 1045.3 (1.34263) System Time 88.311 (0.342554) Percent CPU 1105.8 (6.49444) Context Switches 189081 (2355.78) Sleeps 99231.5 (800.358) Link: http://lkml.kernel.org/r/1464295031-26375-3-git-send-email-thgarnie@google.com Signed-off-by: Thomas Garnier <thgarnie@google.com> Reviewed-by: Kees Cook <keescook@chromium.org> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-07-27 06:21:59 +08:00
for (i = 0; i < count; i++)
s->random_seq[i] *= s->size;
}
return 0;
}
/* Initialize each random sequence freelist per cache */
static void __init init_freelist_randomization(void)
{
struct kmem_cache *s;
mutex_lock(&slab_mutex);
list_for_each_entry(s, &slab_caches, list)
init_cache_random_seq(s);
mutex_unlock(&slab_mutex);
}
/* Get the next entry on the pre-computed freelist randomized */
static void *next_freelist_entry(struct kmem_cache *s, struct page *page,
unsigned long *pos, void *start,
unsigned long page_limit,
unsigned long freelist_count)
{
unsigned int idx;
/*
* If the target page allocation failed, the number of objects on the
* page might be smaller than the usual size defined by the cache.
*/
do {
idx = s->random_seq[*pos];
*pos += 1;
if (*pos >= freelist_count)
*pos = 0;
} while (unlikely(idx >= page_limit));
return (char *)start + idx;
}
/* Shuffle the single linked freelist based on a random pre-computed sequence */
static bool shuffle_freelist(struct kmem_cache *s, struct page *page)
{
void *start;
void *cur;
void *next;
unsigned long idx, pos, page_limit, freelist_count;
if (page->objects < 2 || !s->random_seq)
return false;
freelist_count = oo_objects(s->oo);
pos = get_random_int() % freelist_count;
page_limit = page->objects * s->size;
start = fixup_red_left(s, page_address(page));
/* First entry is used as the base of the freelist */
cur = next_freelist_entry(s, page, &pos, start, page_limit,
freelist_count);
cur = setup_object(s, page, cur);
mm: SLUB freelist randomization Implements freelist randomization for the SLUB allocator. It was previous implemented for the SLAB allocator. Both use the same configuration option (CONFIG_SLAB_FREELIST_RANDOM). The list is randomized during initialization of a new set of pages. The order on different freelist sizes is pre-computed at boot for performance. Each kmem_cache has its own randomized freelist. This security feature reduces the predictability of the kernel SLUB allocator against heap overflows rendering attacks much less stable. For example these attacks exploit the predictability of the heap: - Linux Kernel CAN SLUB overflow (https://goo.gl/oMNWkU) - Exploiting Linux Kernel Heap corruptions (http://goo.gl/EXLn95) Performance results: slab_test impact is between 3% to 4% on average for 100000 attempts without smp. It is a very focused testing, kernbench show the overall impact on the system is way lower. Before: Single thread testing ===================== 1. Kmalloc: Repeatedly allocate then free test 100000 times kmalloc(8) -> 49 cycles kfree -> 77 cycles 100000 times kmalloc(16) -> 51 cycles kfree -> 79 cycles 100000 times kmalloc(32) -> 53 cycles kfree -> 83 cycles 100000 times kmalloc(64) -> 62 cycles kfree -> 90 cycles 100000 times kmalloc(128) -> 81 cycles kfree -> 97 cycles 100000 times kmalloc(256) -> 98 cycles kfree -> 121 cycles 100000 times kmalloc(512) -> 95 cycles kfree -> 122 cycles 100000 times kmalloc(1024) -> 96 cycles kfree -> 126 cycles 100000 times kmalloc(2048) -> 115 cycles kfree -> 140 cycles 100000 times kmalloc(4096) -> 149 cycles kfree -> 171 cycles 2. Kmalloc: alloc/free test 100000 times kmalloc(8)/kfree -> 70 cycles 100000 times kmalloc(16)/kfree -> 70 cycles 100000 times kmalloc(32)/kfree -> 70 cycles 100000 times kmalloc(64)/kfree -> 70 cycles 100000 times kmalloc(128)/kfree -> 70 cycles 100000 times kmalloc(256)/kfree -> 69 cycles 100000 times kmalloc(512)/kfree -> 70 cycles 100000 times kmalloc(1024)/kfree -> 73 cycles 100000 times kmalloc(2048)/kfree -> 72 cycles 100000 times kmalloc(4096)/kfree -> 71 cycles After: Single thread testing ===================== 1. Kmalloc: Repeatedly allocate then free test 100000 times kmalloc(8) -> 57 cycles kfree -> 78 cycles 100000 times kmalloc(16) -> 61 cycles kfree -> 81 cycles 100000 times kmalloc(32) -> 76 cycles kfree -> 93 cycles 100000 times kmalloc(64) -> 83 cycles kfree -> 94 cycles 100000 times kmalloc(128) -> 106 cycles kfree -> 107 cycles 100000 times kmalloc(256) -> 118 cycles kfree -> 117 cycles 100000 times kmalloc(512) -> 114 cycles kfree -> 116 cycles 100000 times kmalloc(1024) -> 115 cycles kfree -> 118 cycles 100000 times kmalloc(2048) -> 147 cycles kfree -> 131 cycles 100000 times kmalloc(4096) -> 214 cycles kfree -> 161 cycles 2. Kmalloc: alloc/free test 100000 times kmalloc(8)/kfree -> 66 cycles 100000 times kmalloc(16)/kfree -> 66 cycles 100000 times kmalloc(32)/kfree -> 66 cycles 100000 times kmalloc(64)/kfree -> 66 cycles 100000 times kmalloc(128)/kfree -> 65 cycles 100000 times kmalloc(256)/kfree -> 67 cycles 100000 times kmalloc(512)/kfree -> 67 cycles 100000 times kmalloc(1024)/kfree -> 64 cycles 100000 times kmalloc(2048)/kfree -> 67 cycles 100000 times kmalloc(4096)/kfree -> 67 cycles Kernbench, before: Average Optimal load -j 12 Run (std deviation): Elapsed Time 101.873 (1.16069) User Time 1045.22 (1.60447) System Time 88.969 (0.559195) Percent CPU 1112.9 (13.8279) Context Switches 189140 (2282.15) Sleeps 99008.6 (768.091) After: Average Optimal load -j 12 Run (std deviation): Elapsed Time 102.47 (0.562732) User Time 1045.3 (1.34263) System Time 88.311 (0.342554) Percent CPU 1105.8 (6.49444) Context Switches 189081 (2355.78) Sleeps 99231.5 (800.358) Link: http://lkml.kernel.org/r/1464295031-26375-3-git-send-email-thgarnie@google.com Signed-off-by: Thomas Garnier <thgarnie@google.com> Reviewed-by: Kees Cook <keescook@chromium.org> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-07-27 06:21:59 +08:00
page->freelist = cur;
for (idx = 1; idx < page->objects; idx++) {
next = next_freelist_entry(s, page, &pos, start, page_limit,
freelist_count);
next = setup_object(s, page, next);
mm: SLUB freelist randomization Implements freelist randomization for the SLUB allocator. It was previous implemented for the SLAB allocator. Both use the same configuration option (CONFIG_SLAB_FREELIST_RANDOM). The list is randomized during initialization of a new set of pages. The order on different freelist sizes is pre-computed at boot for performance. Each kmem_cache has its own randomized freelist. This security feature reduces the predictability of the kernel SLUB allocator against heap overflows rendering attacks much less stable. For example these attacks exploit the predictability of the heap: - Linux Kernel CAN SLUB overflow (https://goo.gl/oMNWkU) - Exploiting Linux Kernel Heap corruptions (http://goo.gl/EXLn95) Performance results: slab_test impact is between 3% to 4% on average for 100000 attempts without smp. It is a very focused testing, kernbench show the overall impact on the system is way lower. Before: Single thread testing ===================== 1. Kmalloc: Repeatedly allocate then free test 100000 times kmalloc(8) -> 49 cycles kfree -> 77 cycles 100000 times kmalloc(16) -> 51 cycles kfree -> 79 cycles 100000 times kmalloc(32) -> 53 cycles kfree -> 83 cycles 100000 times kmalloc(64) -> 62 cycles kfree -> 90 cycles 100000 times kmalloc(128) -> 81 cycles kfree -> 97 cycles 100000 times kmalloc(256) -> 98 cycles kfree -> 121 cycles 100000 times kmalloc(512) -> 95 cycles kfree -> 122 cycles 100000 times kmalloc(1024) -> 96 cycles kfree -> 126 cycles 100000 times kmalloc(2048) -> 115 cycles kfree -> 140 cycles 100000 times kmalloc(4096) -> 149 cycles kfree -> 171 cycles 2. Kmalloc: alloc/free test 100000 times kmalloc(8)/kfree -> 70 cycles 100000 times kmalloc(16)/kfree -> 70 cycles 100000 times kmalloc(32)/kfree -> 70 cycles 100000 times kmalloc(64)/kfree -> 70 cycles 100000 times kmalloc(128)/kfree -> 70 cycles 100000 times kmalloc(256)/kfree -> 69 cycles 100000 times kmalloc(512)/kfree -> 70 cycles 100000 times kmalloc(1024)/kfree -> 73 cycles 100000 times kmalloc(2048)/kfree -> 72 cycles 100000 times kmalloc(4096)/kfree -> 71 cycles After: Single thread testing ===================== 1. Kmalloc: Repeatedly allocate then free test 100000 times kmalloc(8) -> 57 cycles kfree -> 78 cycles 100000 times kmalloc(16) -> 61 cycles kfree -> 81 cycles 100000 times kmalloc(32) -> 76 cycles kfree -> 93 cycles 100000 times kmalloc(64) -> 83 cycles kfree -> 94 cycles 100000 times kmalloc(128) -> 106 cycles kfree -> 107 cycles 100000 times kmalloc(256) -> 118 cycles kfree -> 117 cycles 100000 times kmalloc(512) -> 114 cycles kfree -> 116 cycles 100000 times kmalloc(1024) -> 115 cycles kfree -> 118 cycles 100000 times kmalloc(2048) -> 147 cycles kfree -> 131 cycles 100000 times kmalloc(4096) -> 214 cycles kfree -> 161 cycles 2. Kmalloc: alloc/free test 100000 times kmalloc(8)/kfree -> 66 cycles 100000 times kmalloc(16)/kfree -> 66 cycles 100000 times kmalloc(32)/kfree -> 66 cycles 100000 times kmalloc(64)/kfree -> 66 cycles 100000 times kmalloc(128)/kfree -> 65 cycles 100000 times kmalloc(256)/kfree -> 67 cycles 100000 times kmalloc(512)/kfree -> 67 cycles 100000 times kmalloc(1024)/kfree -> 64 cycles 100000 times kmalloc(2048)/kfree -> 67 cycles 100000 times kmalloc(4096)/kfree -> 67 cycles Kernbench, before: Average Optimal load -j 12 Run (std deviation): Elapsed Time 101.873 (1.16069) User Time 1045.22 (1.60447) System Time 88.969 (0.559195) Percent CPU 1112.9 (13.8279) Context Switches 189140 (2282.15) Sleeps 99008.6 (768.091) After: Average Optimal load -j 12 Run (std deviation): Elapsed Time 102.47 (0.562732) User Time 1045.3 (1.34263) System Time 88.311 (0.342554) Percent CPU 1105.8 (6.49444) Context Switches 189081 (2355.78) Sleeps 99231.5 (800.358) Link: http://lkml.kernel.org/r/1464295031-26375-3-git-send-email-thgarnie@google.com Signed-off-by: Thomas Garnier <thgarnie@google.com> Reviewed-by: Kees Cook <keescook@chromium.org> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-07-27 06:21:59 +08:00
set_freepointer(s, cur, next);
cur = next;
}
set_freepointer(s, cur, NULL);
return true;
}
#else
static inline int init_cache_random_seq(struct kmem_cache *s)
{
return 0;
}
static inline void init_freelist_randomization(void) { }
static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page)
{
return false;
}
#endif /* CONFIG_SLAB_FREELIST_RANDOM */
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
{
struct page *page;
struct kmem_cache_order_objects oo = s->oo;
gfp_t alloc_gfp;
void *start, *p, *next;
int idx;
mm: SLUB freelist randomization Implements freelist randomization for the SLUB allocator. It was previous implemented for the SLAB allocator. Both use the same configuration option (CONFIG_SLAB_FREELIST_RANDOM). The list is randomized during initialization of a new set of pages. The order on different freelist sizes is pre-computed at boot for performance. Each kmem_cache has its own randomized freelist. This security feature reduces the predictability of the kernel SLUB allocator against heap overflows rendering attacks much less stable. For example these attacks exploit the predictability of the heap: - Linux Kernel CAN SLUB overflow (https://goo.gl/oMNWkU) - Exploiting Linux Kernel Heap corruptions (http://goo.gl/EXLn95) Performance results: slab_test impact is between 3% to 4% on average for 100000 attempts without smp. It is a very focused testing, kernbench show the overall impact on the system is way lower. Before: Single thread testing ===================== 1. Kmalloc: Repeatedly allocate then free test 100000 times kmalloc(8) -> 49 cycles kfree -> 77 cycles 100000 times kmalloc(16) -> 51 cycles kfree -> 79 cycles 100000 times kmalloc(32) -> 53 cycles kfree -> 83 cycles 100000 times kmalloc(64) -> 62 cycles kfree -> 90 cycles 100000 times kmalloc(128) -> 81 cycles kfree -> 97 cycles 100000 times kmalloc(256) -> 98 cycles kfree -> 121 cycles 100000 times kmalloc(512) -> 95 cycles kfree -> 122 cycles 100000 times kmalloc(1024) -> 96 cycles kfree -> 126 cycles 100000 times kmalloc(2048) -> 115 cycles kfree -> 140 cycles 100000 times kmalloc(4096) -> 149 cycles kfree -> 171 cycles 2. Kmalloc: alloc/free test 100000 times kmalloc(8)/kfree -> 70 cycles 100000 times kmalloc(16)/kfree -> 70 cycles 100000 times kmalloc(32)/kfree -> 70 cycles 100000 times kmalloc(64)/kfree -> 70 cycles 100000 times kmalloc(128)/kfree -> 70 cycles 100000 times kmalloc(256)/kfree -> 69 cycles 100000 times kmalloc(512)/kfree -> 70 cycles 100000 times kmalloc(1024)/kfree -> 73 cycles 100000 times kmalloc(2048)/kfree -> 72 cycles 100000 times kmalloc(4096)/kfree -> 71 cycles After: Single thread testing ===================== 1. Kmalloc: Repeatedly allocate then free test 100000 times kmalloc(8) -> 57 cycles kfree -> 78 cycles 100000 times kmalloc(16) -> 61 cycles kfree -> 81 cycles 100000 times kmalloc(32) -> 76 cycles kfree -> 93 cycles 100000 times kmalloc(64) -> 83 cycles kfree -> 94 cycles 100000 times kmalloc(128) -> 106 cycles kfree -> 107 cycles 100000 times kmalloc(256) -> 118 cycles kfree -> 117 cycles 100000 times kmalloc(512) -> 114 cycles kfree -> 116 cycles 100000 times kmalloc(1024) -> 115 cycles kfree -> 118 cycles 100000 times kmalloc(2048) -> 147 cycles kfree -> 131 cycles 100000 times kmalloc(4096) -> 214 cycles kfree -> 161 cycles 2. Kmalloc: alloc/free test 100000 times kmalloc(8)/kfree -> 66 cycles 100000 times kmalloc(16)/kfree -> 66 cycles 100000 times kmalloc(32)/kfree -> 66 cycles 100000 times kmalloc(64)/kfree -> 66 cycles 100000 times kmalloc(128)/kfree -> 65 cycles 100000 times kmalloc(256)/kfree -> 67 cycles 100000 times kmalloc(512)/kfree -> 67 cycles 100000 times kmalloc(1024)/kfree -> 64 cycles 100000 times kmalloc(2048)/kfree -> 67 cycles 100000 times kmalloc(4096)/kfree -> 67 cycles Kernbench, before: Average Optimal load -j 12 Run (std deviation): Elapsed Time 101.873 (1.16069) User Time 1045.22 (1.60447) System Time 88.969 (0.559195) Percent CPU 1112.9 (13.8279) Context Switches 189140 (2282.15) Sleeps 99008.6 (768.091) After: Average Optimal load -j 12 Run (std deviation): Elapsed Time 102.47 (0.562732) User Time 1045.3 (1.34263) System Time 88.311 (0.342554) Percent CPU 1105.8 (6.49444) Context Switches 189081 (2355.78) Sleeps 99231.5 (800.358) Link: http://lkml.kernel.org/r/1464295031-26375-3-git-send-email-thgarnie@google.com Signed-off-by: Thomas Garnier <thgarnie@google.com> Reviewed-by: Kees Cook <keescook@chromium.org> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-07-27 06:21:59 +08:00
bool shuffle;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
flags &= gfp_allowed_mask;
mm, page_alloc: distinguish between being unable to sleep, unwilling to sleep and avoiding waking kswapd __GFP_WAIT has been used to identify atomic context in callers that hold spinlocks or are in interrupts. They are expected to be high priority and have access one of two watermarks lower than "min" which can be referred to as the "atomic reserve". __GFP_HIGH users get access to the first lower watermark and can be called the "high priority reserve". Over time, callers had a requirement to not block when fallback options were available. Some have abused __GFP_WAIT leading to a situation where an optimisitic allocation with a fallback option can access atomic reserves. This patch uses __GFP_ATOMIC to identify callers that are truely atomic, cannot sleep and have no alternative. High priority users continue to use __GFP_HIGH. __GFP_DIRECT_RECLAIM identifies callers that can sleep and are willing to enter direct reclaim. __GFP_KSWAPD_RECLAIM to identify callers that want to wake kswapd for background reclaim. __GFP_WAIT is redefined as a caller that is willing to enter direct reclaim and wake kswapd for background reclaim. This patch then converts a number of sites o __GFP_ATOMIC is used by callers that are high priority and have memory pools for those requests. GFP_ATOMIC uses this flag. o Callers that have a limited mempool to guarantee forward progress clear __GFP_DIRECT_RECLAIM but keep __GFP_KSWAPD_RECLAIM. bio allocations fall into this category where kswapd will still be woken but atomic reserves are not used as there is a one-entry mempool to guarantee progress. o Callers that are checking if they are non-blocking should use the helper gfpflags_allow_blocking() where possible. This is because checking for __GFP_WAIT as was done historically now can trigger false positives. Some exceptions like dm-crypt.c exist where the code intent is clearer if __GFP_DIRECT_RECLAIM is used instead of the helper due to flag manipulations. o Callers that built their own GFP flags instead of starting with GFP_KERNEL and friends now also need to specify __GFP_KSWAPD_RECLAIM. The first key hazard to watch out for is callers that removed __GFP_WAIT and was depending on access to atomic reserves for inconspicuous reasons. In some cases it may be appropriate for them to use __GFP_HIGH. The second key hazard is callers that assembled their own combination of GFP flags instead of starting with something like GFP_KERNEL. They may now wish to specify __GFP_KSWAPD_RECLAIM. It's almost certainly harmless if it's missed in most cases as other activity will wake kswapd. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Acked-by: Vlastimil Babka <vbabka@suse.cz> Acked-by: Michal Hocko <mhocko@suse.com> Acked-by: Johannes Weiner <hannes@cmpxchg.org> Cc: Christoph Lameter <cl@linux.com> Cc: David Rientjes <rientjes@google.com> Cc: Vitaly Wool <vitalywool@gmail.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-07 08:28:21 +08:00
if (gfpflags_allow_blocking(flags))
local_irq_enable();
flags |= s->allocflags;
/*
* Let the initial higher-order allocation fail under memory pressure
* so we fall-back to the minimum order allocation.
*/
alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
mm, page_alloc: distinguish between being unable to sleep, unwilling to sleep and avoiding waking kswapd __GFP_WAIT has been used to identify atomic context in callers that hold spinlocks or are in interrupts. They are expected to be high priority and have access one of two watermarks lower than "min" which can be referred to as the "atomic reserve". __GFP_HIGH users get access to the first lower watermark and can be called the "high priority reserve". Over time, callers had a requirement to not block when fallback options were available. Some have abused __GFP_WAIT leading to a situation where an optimisitic allocation with a fallback option can access atomic reserves. This patch uses __GFP_ATOMIC to identify callers that are truely atomic, cannot sleep and have no alternative. High priority users continue to use __GFP_HIGH. __GFP_DIRECT_RECLAIM identifies callers that can sleep and are willing to enter direct reclaim. __GFP_KSWAPD_RECLAIM to identify callers that want to wake kswapd for background reclaim. __GFP_WAIT is redefined as a caller that is willing to enter direct reclaim and wake kswapd for background reclaim. This patch then converts a number of sites o __GFP_ATOMIC is used by callers that are high priority and have memory pools for those requests. GFP_ATOMIC uses this flag. o Callers that have a limited mempool to guarantee forward progress clear __GFP_DIRECT_RECLAIM but keep __GFP_KSWAPD_RECLAIM. bio allocations fall into this category where kswapd will still be woken but atomic reserves are not used as there is a one-entry mempool to guarantee progress. o Callers that are checking if they are non-blocking should use the helper gfpflags_allow_blocking() where possible. This is because checking for __GFP_WAIT as was done historically now can trigger false positives. Some exceptions like dm-crypt.c exist where the code intent is clearer if __GFP_DIRECT_RECLAIM is used instead of the helper due to flag manipulations. o Callers that built their own GFP flags instead of starting with GFP_KERNEL and friends now also need to specify __GFP_KSWAPD_RECLAIM. The first key hazard to watch out for is callers that removed __GFP_WAIT and was depending on access to atomic reserves for inconspicuous reasons. In some cases it may be appropriate for them to use __GFP_HIGH. The second key hazard is callers that assembled their own combination of GFP flags instead of starting with something like GFP_KERNEL. They may now wish to specify __GFP_KSWAPD_RECLAIM. It's almost certainly harmless if it's missed in most cases as other activity will wake kswapd. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Acked-by: Vlastimil Babka <vbabka@suse.cz> Acked-by: Michal Hocko <mhocko@suse.com> Acked-by: Johannes Weiner <hannes@cmpxchg.org> Cc: Christoph Lameter <cl@linux.com> Cc: David Rientjes <rientjes@google.com> Cc: Vitaly Wool <vitalywool@gmail.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-07 08:28:21 +08:00
if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
mm: thp: set THP defrag by default to madvise and add a stall-free defrag option THP defrag is enabled by default to direct reclaim/compact but not wake kswapd in the event of a THP allocation failure. The problem is that THP allocation requests potentially enter reclaim/compaction. This potentially incurs a severe stall that is not guaranteed to be offset by reduced TLB misses. While there has been considerable effort to reduce the impact of reclaim/compaction, it is still a high cost and workloads that should fit in memory fail to do so. Specifically, a simple anon/file streaming workload will enter direct reclaim on NUMA at least even though the working set size is 80% of RAM. It's been years and it's time to throw in the towel. First, this patch defines THP defrag as follows; madvise: A failed allocation will direct reclaim/compact if the application requests it never: Neither reclaim/compact nor wake kswapd defer: A failed allocation will wake kswapd/kcompactd always: A failed allocation will direct reclaim/compact (historical behaviour) khugepaged defrag will enter direct/reclaim but not wake kswapd. Next it sets the default defrag option to be "madvise" to only enter direct reclaim/compaction for applications that specifically requested it. Lastly, it removes a check from the page allocator slowpath that is related to __GFP_THISNODE to allow "defer" to work. The callers that really cares are slub/slab and they are updated accordingly. The slab one may be surprising because it also corrects a comment as kswapd was never woken up by that path. This means that a THP fault will no longer stall for most applications by default and the ideal for most users that get THP if they are immediately available. There are still options for users that prefer a stall at startup of a new application by either restoring historical behaviour with "always" or pick a half-way point with "defer" where kswapd does some of the work in the background and wakes kcompactd if necessary. THP defrag for khugepaged remains enabled and will enter direct/reclaim but no wakeup kswapd or kcompactd. After this patch a THP allocation failure will quickly fallback and rely on khugepaged to recover the situation at some time in the future. In some cases, this will reduce THP usage but the benefit of THP is hard to measure and not a universal win where as a stall to reclaim/compaction is definitely measurable and can be painful. The first test for this is using "usemem" to read a large file and write a large anonymous mapping (to avoid the zero page) multiple times. The total size of the mappings is 80% of RAM and the benchmark simply measures how long it takes to complete. It uses multiple threads to see if that is a factor. On UMA, the performance is almost identical so is not reported but on NUMA, we see this usemem 4.4.0 4.4.0 kcompactd-v1r1 nodefrag-v1r3 Amean System-1 102.86 ( 0.00%) 46.81 ( 54.50%) Amean System-4 37.85 ( 0.00%) 34.02 ( 10.12%) Amean System-7 48.12 ( 0.00%) 46.89 ( 2.56%) Amean System-12 51.98 ( 0.00%) 56.96 ( -9.57%) Amean System-21 80.16 ( 0.00%) 79.05 ( 1.39%) Amean System-30 110.71 ( 0.00%) 107.17 ( 3.20%) Amean System-48 127.98 ( 0.00%) 124.83 ( 2.46%) Amean Elapsd-1 185.84 ( 0.00%) 105.51 ( 43.23%) Amean Elapsd-4 26.19 ( 0.00%) 25.58 ( 2.33%) Amean Elapsd-7 21.65 ( 0.00%) 21.62 ( 0.16%) Amean Elapsd-12 18.58 ( 0.00%) 17.94 ( 3.43%) Amean Elapsd-21 17.53 ( 0.00%) 16.60 ( 5.33%) Amean Elapsd-30 17.45 ( 0.00%) 17.13 ( 1.84%) Amean Elapsd-48 15.40 ( 0.00%) 15.27 ( 0.82%) For a single thread, the benchmark completes 43.23% faster with this patch applied with smaller benefits as the thread increases. Similar, notice the large reduction in most cases in system CPU usage. The overall CPU time is 4.4.0 4.4.0 kcompactd-v1r1 nodefrag-v1r3 User 10357.65 10438.33 System 3988.88 3543.94 Elapsed 2203.01 1634.41 Which is substantial. Now, the reclaim figures 4.4.0 4.4.0 kcompactd-v1r1nodefrag-v1r3 Minor Faults 128458477 278352931 Major Faults 2174976 225 Swap Ins 16904701 0 Swap Outs 17359627 0 Allocation stalls 43611 0 DMA allocs 0 0 DMA32 allocs 19832646 19448017 Normal allocs 614488453 580941839 Movable allocs 0 0 Direct pages scanned 24163800 0 Kswapd pages scanned 0 0 Kswapd pages reclaimed 0 0 Direct pages reclaimed 20691346 0 Compaction stalls 42263 0 Compaction success 938 0 Compaction failures 41325 0 This patch eliminates almost all swapping and direct reclaim activity. There is still overhead but it's from NUMA balancing which does not identify that it's pointless trying to do anything with this workload. I also tried the thpscale benchmark which forces a corner case where compaction can be used heavily and measures the latency of whether base or huge pages were used thpscale Fault Latencies 4.4.0 4.4.0 kcompactd-v1r1 nodefrag-v1r3 Amean fault-base-1 5288.84 ( 0.00%) 2817.12 ( 46.73%) Amean fault-base-3 6365.53 ( 0.00%) 3499.11 ( 45.03%) Amean fault-base-5 6526.19 ( 0.00%) 4363.06 ( 33.15%) Amean fault-base-7 7142.25 ( 0.00%) 4858.08 ( 31.98%) Amean fault-base-12 13827.64 ( 0.00%) 10292.11 ( 25.57%) Amean fault-base-18 18235.07 ( 0.00%) 13788.84 ( 24.38%) Amean fault-base-24 21597.80 ( 0.00%) 24388.03 (-12.92%) Amean fault-base-30 26754.15 ( 0.00%) 19700.55 ( 26.36%) Amean fault-base-32 26784.94 ( 0.00%) 19513.57 ( 27.15%) Amean fault-huge-1 4223.96 ( 0.00%) 2178.57 ( 48.42%) Amean fault-huge-3 2194.77 ( 0.00%) 2149.74 ( 2.05%) Amean fault-huge-5 2569.60 ( 0.00%) 2346.95 ( 8.66%) Amean fault-huge-7 3612.69 ( 0.00%) 2997.70 ( 17.02%) Amean fault-huge-12 3301.75 ( 0.00%) 6727.02 (-103.74%) Amean fault-huge-18 6696.47 ( 0.00%) 6685.72 ( 0.16%) Amean fault-huge-24 8000.72 ( 0.00%) 9311.43 (-16.38%) Amean fault-huge-30 13305.55 ( 0.00%) 9750.45 ( 26.72%) Amean fault-huge-32 9981.71 ( 0.00%) 10316.06 ( -3.35%) The average time to fault pages is substantially reduced in the majority of caseds but with the obvious caveat that fewer THPs are actually used in this adverse workload 4.4.0 4.4.0 kcompactd-v1r1 nodefrag-v1r3 Percentage huge-1 0.71 ( 0.00%) 14.04 (1865.22%) Percentage huge-3 10.77 ( 0.00%) 33.05 (206.85%) Percentage huge-5 60.39 ( 0.00%) 38.51 (-36.23%) Percentage huge-7 45.97 ( 0.00%) 34.57 (-24.79%) Percentage huge-12 68.12 ( 0.00%) 40.07 (-41.17%) Percentage huge-18 64.93 ( 0.00%) 47.82 (-26.35%) Percentage huge-24 62.69 ( 0.00%) 44.23 (-29.44%) Percentage huge-30 43.49 ( 0.00%) 55.38 ( 27.34%) Percentage huge-32 50.72 ( 0.00%) 51.90 ( 2.35%) 4.4.0 4.4.0 kcompactd-v1r1nodefrag-v1r3 Minor Faults 37429143 47564000 Major Faults 1916 1558 Swap Ins 1466 1079 Swap Outs 2936863 149626 Allocation stalls 62510 3 DMA allocs 0 0 DMA32 allocs 6566458 6401314 Normal allocs 216361697 216538171 Movable allocs 0 0 Direct pages scanned 25977580 17998 Kswapd pages scanned 0 3638931 Kswapd pages reclaimed 0 207236 Direct pages reclaimed 8833714 88 Compaction stalls 103349 5 Compaction success 270 4 Compaction failures 103079 1 Note again that while this does swap as it's an aggressive workload, the direct relcim activity and allocation stalls is substantially reduced. There is some kswapd activity but ftrace showed that the kswapd activity was due to normal wakeups from 4K pages being allocated. Compaction-related stalls and activity are almost eliminated. I also tried the stutter benchmark. For this, I do not have figures for NUMA but it's something that does impact UMA so I'll report what is available stutter 4.4.0 4.4.0 kcompactd-v1r1 nodefrag-v1r3 Min mmap 7.3571 ( 0.00%) 7.3438 ( 0.18%) 1st-qrtle mmap 7.5278 ( 0.00%) 17.9200 (-138.05%) 2nd-qrtle mmap 7.6818 ( 0.00%) 21.6055 (-181.25%) 3rd-qrtle mmap 11.0889 ( 0.00%) 21.8881 (-97.39%) Max-90% mmap 27.8978 ( 0.00%) 22.1632 ( 20.56%) Max-93% mmap 28.3202 ( 0.00%) 22.3044 ( 21.24%) Max-95% mmap 28.5600 ( 0.00%) 22.4580 ( 21.37%) Max-99% mmap 29.6032 ( 0.00%) 25.5216 ( 13.79%) Max mmap 4109.7289 ( 0.00%) 4813.9832 (-17.14%) Mean mmap 12.4474 ( 0.00%) 19.3027 (-55.07%) This benchmark is trying to fault an anonymous mapping while there is a heavy IO load -- a scenario that desktop users used to complain about frequently. This shows a mix because the ideal case of mapping with THP is not hit as often. However, note that 99% of the mappings complete 13.79% faster. The CPU usage here is particularly interesting 4.4.0 4.4.0 kcompactd-v1r1nodefrag-v1r3 User 67.50 0.99 System 1327.88 91.30 Elapsed 2079.00 2128.98 And once again we look at the reclaim figures 4.4.0 4.4.0 kcompactd-v1r1nodefrag-v1r3 Minor Faults 335241922 1314582827 Major Faults 715 819 Swap Ins 0 0 Swap Outs 0 0 Allocation stalls 532723 0 DMA allocs 0 0 DMA32 allocs 1822364341 1177950222 Normal allocs 1815640808 1517844854 Movable allocs 0 0 Direct pages scanned 21892772 0 Kswapd pages scanned 20015890 41879484 Kswapd pages reclaimed 19961986 41822072 Direct pages reclaimed 21892741 0 Compaction stalls 1065755 0 Compaction success 514 0 Compaction failures 1065241 0 Allocation stalls and all direct reclaim activity is eliminated as well as compaction-related stalls. THP gives impressive gains in some cases but only if they are quickly available. We're not going to reach the point where they are completely free so lets take the costs out of the fast paths finally and defer the cost to kswapd, kcompactd and khugepaged where it belongs. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Acked-by: Rik van Riel <riel@redhat.com> Acked-by: Johannes Weiner <hannes@cmpxchg.org> Acked-by: Vlastimil Babka <vbabka@suse.cz> Cc: Andrea Arcangeli <aarcange@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-03-18 05:19:23 +08:00
alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
page = alloc_slab_page(s, alloc_gfp, node, oo);
if (unlikely(!page)) {
oo = s->min;
alloc_gfp = flags;
/*
* Allocation may have failed due to fragmentation.
* Try a lower order alloc if possible
*/
page = alloc_slab_page(s, alloc_gfp, node, oo);
if (unlikely(!page))
goto out;
stat(s, ORDER_FALLBACK);
}
page->objects = oo_objects(oo);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
mm: memcg/slab: pre-allocate obj_cgroups for slab caches with SLAB_ACCOUNT In general it's unknown in advance if a slab page will contain accounted objects or not. In order to avoid memory waste, an obj_cgroup vector is allocated dynamically when a need to account of a new object arises. Such approach is memory efficient, but requires an expensive cmpxchg() to set up the memcg/objcgs pointer, because an allocation can race with a different allocation on another cpu. But in some common cases it's known for sure that a slab page will contain accounted objects: if the page belongs to a slab cache with a SLAB_ACCOUNT flag set. It includes such popular objects like vm_area_struct, anon_vma, task_struct, etc. In such cases we can pre-allocate the objcgs vector and simple assign it to the page without any atomic operations, because at this early stage the page is not visible to anyone else. A very simplistic benchmark (allocating 10000000 64-bytes objects in a row) shows ~15% win. In the real life it seems that most workloads are not very sensitive to the speed of (accounted) slab allocations. [guro@fb.com: open-code set_page_objcgs() and add some comments, by Johannes] Link: https://lkml.kernel.org/r/20201113001926.GA2934489@carbon.dhcp.thefacebook.com [akpm@linux-foundation.org: fix it for mm-slub-call-account_slab_page-after-slab-page-initialization-fix.patch] Link: https://lkml.kernel.org/r/20201110195753.530157-2-guro@fb.com Signed-off-by: Roman Gushchin <guro@fb.com> Acked-by: Johannes Weiner <hannes@cmpxchg.org> Reviewed-by: Shakeel Butt <shakeelb@google.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Christoph Lameter <cl@linux.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-25 04:03:11 +08:00
account_slab_page(page, oo_order(oo), s, flags);
page->slab_cache = s;
__SetPageSlab(page);
mm: make page pfmemalloc check more robust Commit c48a11c7ad26 ("netvm: propagate page->pfmemalloc to skb") added checks for page->pfmemalloc to __skb_fill_page_desc(): if (page->pfmemalloc && !page->mapping) skb->pfmemalloc = true; It assumes page->mapping == NULL implies that page->pfmemalloc can be trusted. However, __delete_from_page_cache() can set set page->mapping to NULL and leave page->index value alone. Due to being in union, a non-zero page->index will be interpreted as true page->pfmemalloc. So the assumption is invalid if the networking code can see such a page. And it seems it can. We have encountered this with a NFS over loopback setup when such a page is attached to a new skbuf. There is no copying going on in this case so the page confuses __skb_fill_page_desc which interprets the index as pfmemalloc flag and the network stack drops packets that have been allocated using the reserves unless they are to be queued on sockets handling the swapping which is the case here and that leads to hangs when the nfs client waits for a response from the server which has been dropped and thus never arrive. The struct page is already heavily packed so rather than finding another hole to put it in, let's do a trick instead. We can reuse the index again but define it to an impossible value (-1UL). This is the page index so it should never see the value that large. Replace all direct users of page->pfmemalloc by page_is_pfmemalloc which will hide this nastiness from unspoiled eyes. The information will get lost if somebody wants to use page->index obviously but that was the case before and the original code expected that the information should be persisted somewhere else if that is really needed (e.g. what SLAB and SLUB do). [akpm@linux-foundation.org: fix blooper in slub] Fixes: c48a11c7ad26 ("netvm: propagate page->pfmemalloc to skb") Signed-off-by: Michal Hocko <mhocko@suse.com> Debugged-by: Vlastimil Babka <vbabka@suse.com> Debugged-by: Jiri Bohac <jbohac@suse.com> Cc: Eric Dumazet <eric.dumazet@gmail.com> Cc: David Miller <davem@davemloft.net> Acked-by: Mel Gorman <mgorman@suse.de> Cc: <stable@vger.kernel.org> [3.6+] Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-08-22 05:11:51 +08:00
if (page_is_pfmemalloc(page))
mm: sl[au]b: add knowledge of PFMEMALLOC reserve pages When a user or administrator requires swap for their application, they create a swap partition and file, format it with mkswap and activate it with swapon. Swap over the network is considered as an option in diskless systems. The two likely scenarios are when blade servers are used as part of a cluster where the form factor or maintenance costs do not allow the use of disks and thin clients. The Linux Terminal Server Project recommends the use of the Network Block Device (NBD) for swap according to the manual at https://sourceforge.net/projects/ltsp/files/Docs-Admin-Guide/LTSPManual.pdf/download There is also documentation and tutorials on how to setup swap over NBD at places like https://help.ubuntu.com/community/UbuntuLTSP/EnableNBDSWAP The nbd-client also documents the use of NBD as swap. Despite this, the fact is that a machine using NBD for swap can deadlock within minutes if swap is used intensively. This patch series addresses the problem. The core issue is that network block devices do not use mempools like normal block devices do. As the host cannot control where they receive packets from, they cannot reliably work out in advance how much memory they might need. Some years ago, Peter Zijlstra developed a series of patches that supported swap over an NFS that at least one distribution is carrying within their kernels. This patch series borrows very heavily from Peter's work to support swapping over NBD as a pre-requisite to supporting swap-over-NFS. The bulk of the complexity is concerned with preserving memory that is allocated from the PFMEMALLOC reserves for use by the network layer which is needed for both NBD and NFS. Patch 1 adds knowledge of the PFMEMALLOC reserves to SLAB and SLUB to preserve access to pages allocated under low memory situations to callers that are freeing memory. Patch 2 optimises the SLUB fast path to avoid pfmemalloc checks Patch 3 introduces __GFP_MEMALLOC to allow access to the PFMEMALLOC reserves without setting PFMEMALLOC. Patch 4 opens the possibility for softirqs to use PFMEMALLOC reserves for later use by network packet processing. Patch 5 only sets page->pfmemalloc when ALLOC_NO_WATERMARKS was required Patch 6 ignores memory policies when ALLOC_NO_WATERMARKS is set. Patches 7-12 allows network processing to use PFMEMALLOC reserves when the socket has been marked as being used by the VM to clean pages. If packets are received and stored in pages that were allocated under low-memory situations and are unrelated to the VM, the packets are dropped. Patch 11 reintroduces __skb_alloc_page which the networking folk may object to but is needed in some cases to propogate pfmemalloc from a newly allocated page to an skb. If there is a strong objection, this patch can be dropped with the impact being that swap-over-network will be slower in some cases but it should not fail. Patch 13 is a micro-optimisation to avoid a function call in the common case. Patch 14 tags NBD sockets as being SOCK_MEMALLOC so they can use PFMEMALLOC if necessary. Patch 15 notes that it is still possible for the PFMEMALLOC reserve to be depleted. To prevent this, direct reclaimers get throttled on a waitqueue if 50% of the PFMEMALLOC reserves are depleted. It is expected that kswapd and the direct reclaimers already running will clean enough pages for the low watermark to be reached and the throttled processes are woken up. Patch 16 adds a statistic to track how often processes get throttled Some basic performance testing was run using kernel builds, netperf on loopback for UDP and TCP, hackbench (pipes and sockets), iozone and sysbench. Each of them were expected to use the sl*b allocators reasonably heavily but there did not appear to be significant performance variances. For testing swap-over-NBD, a machine was booted with 2G of RAM with a swapfile backed by NBD. 8*NUM_CPU processes were started that create anonymous memory mappings and read them linearly in a loop. The total size of the mappings were 4*PHYSICAL_MEMORY to use swap heavily under memory pressure. Without the patches and using SLUB, the machine locks up within minutes and runs to completion with them applied. With SLAB, the story is different as an unpatched kernel run to completion. However, the patched kernel completed the test 45% faster. MICRO 3.5.0-rc2 3.5.0-rc2 vanilla swapnbd Unrecognised test vmscan-anon-mmap-write MMTests Statistics: duration Sys Time Running Test (seconds) 197.80 173.07 User+Sys Time Running Test (seconds) 206.96 182.03 Total Elapsed Time (seconds) 3240.70 1762.09 This patch: mm: sl[au]b: add knowledge of PFMEMALLOC reserve pages Allocations of pages below the min watermark run a risk of the machine hanging due to a lack of memory. To prevent this, only callers who have PF_MEMALLOC or TIF_MEMDIE set and are not processing an interrupt are allowed to allocate with ALLOC_NO_WATERMARKS. Once they are allocated to a slab though, nothing prevents other callers consuming free objects within those slabs. This patch limits access to slab pages that were alloced from the PFMEMALLOC reserves. When this patch is applied, pages allocated from below the low watermark are returned with page->pfmemalloc set and it is up to the caller to determine how the page should be protected. SLAB restricts access to any page with page->pfmemalloc set to callers which are known to able to access the PFMEMALLOC reserve. If one is not available, an attempt is made to allocate a new page rather than use a reserve. SLUB is a bit more relaxed in that it only records if the current per-CPU page was allocated from PFMEMALLOC reserve and uses another partial slab if the caller does not have the necessary GFP or process flags. This was found to be sufficient in tests to avoid hangs due to SLUB generally maintaining smaller lists than SLAB. In low-memory conditions it does mean that !PFMEMALLOC allocators can fail a slab allocation even though free objects are available because they are being preserved for callers that are freeing pages. [a.p.zijlstra@chello.nl: Original implementation] [sebastian@breakpoint.cc: Correct order of page flag clearing] Signed-off-by: Mel Gorman <mgorman@suse.de> Cc: David Miller <davem@davemloft.net> Cc: Neil Brown <neilb@suse.de> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Christie <michaelc@cs.wisc.edu> Cc: Eric B Munson <emunson@mgebm.net> Cc: Eric Dumazet <eric.dumazet@gmail.com> Cc: Sebastian Andrzej Siewior <sebastian@breakpoint.cc> Cc: Mel Gorman <mgorman@suse.de> Cc: Christoph Lameter <cl@linux.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-08-01 07:43:58 +08:00
SetPageSlabPfmemalloc(page);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
kasan_poison_slab(page);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
start = page_address(page);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
setup_page_debug(s, page, start);
mm: slub: add kernel address sanitizer support for slub allocator With this patch kasan will be able to catch bugs in memory allocated by slub. Initially all objects in newly allocated slab page, marked as redzone. Later, when allocation of slub object happens, requested by caller number of bytes marked as accessible, and the rest of the object (including slub's metadata) marked as redzone (inaccessible). We also mark object as accessible if ksize was called for this object. There is some places in kernel where ksize function is called to inquire size of really allocated area. Such callers could validly access whole allocated memory, so it should be marked as accessible. Code in slub.c and slab_common.c files could validly access to object's metadata, so instrumentation for this files are disabled. Signed-off-by: Andrey Ryabinin <a.ryabinin@samsung.com> Signed-off-by: Dmitry Chernenkov <dmitryc@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Konstantin Serebryany <kcc@google.com> Signed-off-by: Andrey Konovalov <adech.fo@gmail.com> Cc: Yuri Gribov <tetra2005@gmail.com> Cc: Konstantin Khlebnikov <koct9i@gmail.com> Cc: Sasha Levin <sasha.levin@oracle.com> Cc: Christoph Lameter <cl@linux.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Ingo Molnar <mingo@elte.hu> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-14 06:39:42 +08:00
mm: SLUB freelist randomization Implements freelist randomization for the SLUB allocator. It was previous implemented for the SLAB allocator. Both use the same configuration option (CONFIG_SLAB_FREELIST_RANDOM). The list is randomized during initialization of a new set of pages. The order on different freelist sizes is pre-computed at boot for performance. Each kmem_cache has its own randomized freelist. This security feature reduces the predictability of the kernel SLUB allocator against heap overflows rendering attacks much less stable. For example these attacks exploit the predictability of the heap: - Linux Kernel CAN SLUB overflow (https://goo.gl/oMNWkU) - Exploiting Linux Kernel Heap corruptions (http://goo.gl/EXLn95) Performance results: slab_test impact is between 3% to 4% on average for 100000 attempts without smp. It is a very focused testing, kernbench show the overall impact on the system is way lower. Before: Single thread testing ===================== 1. Kmalloc: Repeatedly allocate then free test 100000 times kmalloc(8) -> 49 cycles kfree -> 77 cycles 100000 times kmalloc(16) -> 51 cycles kfree -> 79 cycles 100000 times kmalloc(32) -> 53 cycles kfree -> 83 cycles 100000 times kmalloc(64) -> 62 cycles kfree -> 90 cycles 100000 times kmalloc(128) -> 81 cycles kfree -> 97 cycles 100000 times kmalloc(256) -> 98 cycles kfree -> 121 cycles 100000 times kmalloc(512) -> 95 cycles kfree -> 122 cycles 100000 times kmalloc(1024) -> 96 cycles kfree -> 126 cycles 100000 times kmalloc(2048) -> 115 cycles kfree -> 140 cycles 100000 times kmalloc(4096) -> 149 cycles kfree -> 171 cycles 2. Kmalloc: alloc/free test 100000 times kmalloc(8)/kfree -> 70 cycles 100000 times kmalloc(16)/kfree -> 70 cycles 100000 times kmalloc(32)/kfree -> 70 cycles 100000 times kmalloc(64)/kfree -> 70 cycles 100000 times kmalloc(128)/kfree -> 70 cycles 100000 times kmalloc(256)/kfree -> 69 cycles 100000 times kmalloc(512)/kfree -> 70 cycles 100000 times kmalloc(1024)/kfree -> 73 cycles 100000 times kmalloc(2048)/kfree -> 72 cycles 100000 times kmalloc(4096)/kfree -> 71 cycles After: Single thread testing ===================== 1. Kmalloc: Repeatedly allocate then free test 100000 times kmalloc(8) -> 57 cycles kfree -> 78 cycles 100000 times kmalloc(16) -> 61 cycles kfree -> 81 cycles 100000 times kmalloc(32) -> 76 cycles kfree -> 93 cycles 100000 times kmalloc(64) -> 83 cycles kfree -> 94 cycles 100000 times kmalloc(128) -> 106 cycles kfree -> 107 cycles 100000 times kmalloc(256) -> 118 cycles kfree -> 117 cycles 100000 times kmalloc(512) -> 114 cycles kfree -> 116 cycles 100000 times kmalloc(1024) -> 115 cycles kfree -> 118 cycles 100000 times kmalloc(2048) -> 147 cycles kfree -> 131 cycles 100000 times kmalloc(4096) -> 214 cycles kfree -> 161 cycles 2. Kmalloc: alloc/free test 100000 times kmalloc(8)/kfree -> 66 cycles 100000 times kmalloc(16)/kfree -> 66 cycles 100000 times kmalloc(32)/kfree -> 66 cycles 100000 times kmalloc(64)/kfree -> 66 cycles 100000 times kmalloc(128)/kfree -> 65 cycles 100000 times kmalloc(256)/kfree -> 67 cycles 100000 times kmalloc(512)/kfree -> 67 cycles 100000 times kmalloc(1024)/kfree -> 64 cycles 100000 times kmalloc(2048)/kfree -> 67 cycles 100000 times kmalloc(4096)/kfree -> 67 cycles Kernbench, before: Average Optimal load -j 12 Run (std deviation): Elapsed Time 101.873 (1.16069) User Time 1045.22 (1.60447) System Time 88.969 (0.559195) Percent CPU 1112.9 (13.8279) Context Switches 189140 (2282.15) Sleeps 99008.6 (768.091) After: Average Optimal load -j 12 Run (std deviation): Elapsed Time 102.47 (0.562732) User Time 1045.3 (1.34263) System Time 88.311 (0.342554) Percent CPU 1105.8 (6.49444) Context Switches 189081 (2355.78) Sleeps 99231.5 (800.358) Link: http://lkml.kernel.org/r/1464295031-26375-3-git-send-email-thgarnie@google.com Signed-off-by: Thomas Garnier <thgarnie@google.com> Reviewed-by: Kees Cook <keescook@chromium.org> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-07-27 06:21:59 +08:00
shuffle = shuffle_freelist(s, page);
if (!shuffle) {
start = fixup_red_left(s, start);
start = setup_object(s, page, start);
page->freelist = start;
for (idx = 0, p = start; idx < page->objects - 1; idx++) {
next = p + s->size;
next = setup_object(s, page, next);
set_freepointer(s, p, next);
p = next;
}
set_freepointer(s, p, NULL);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
page->inuse = page->objects;
page->frozen = 1;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
out:
mm, page_alloc: distinguish between being unable to sleep, unwilling to sleep and avoiding waking kswapd __GFP_WAIT has been used to identify atomic context in callers that hold spinlocks or are in interrupts. They are expected to be high priority and have access one of two watermarks lower than "min" which can be referred to as the "atomic reserve". __GFP_HIGH users get access to the first lower watermark and can be called the "high priority reserve". Over time, callers had a requirement to not block when fallback options were available. Some have abused __GFP_WAIT leading to a situation where an optimisitic allocation with a fallback option can access atomic reserves. This patch uses __GFP_ATOMIC to identify callers that are truely atomic, cannot sleep and have no alternative. High priority users continue to use __GFP_HIGH. __GFP_DIRECT_RECLAIM identifies callers that can sleep and are willing to enter direct reclaim. __GFP_KSWAPD_RECLAIM to identify callers that want to wake kswapd for background reclaim. __GFP_WAIT is redefined as a caller that is willing to enter direct reclaim and wake kswapd for background reclaim. This patch then converts a number of sites o __GFP_ATOMIC is used by callers that are high priority and have memory pools for those requests. GFP_ATOMIC uses this flag. o Callers that have a limited mempool to guarantee forward progress clear __GFP_DIRECT_RECLAIM but keep __GFP_KSWAPD_RECLAIM. bio allocations fall into this category where kswapd will still be woken but atomic reserves are not used as there is a one-entry mempool to guarantee progress. o Callers that are checking if they are non-blocking should use the helper gfpflags_allow_blocking() where possible. This is because checking for __GFP_WAIT as was done historically now can trigger false positives. Some exceptions like dm-crypt.c exist where the code intent is clearer if __GFP_DIRECT_RECLAIM is used instead of the helper due to flag manipulations. o Callers that built their own GFP flags instead of starting with GFP_KERNEL and friends now also need to specify __GFP_KSWAPD_RECLAIM. The first key hazard to watch out for is callers that removed __GFP_WAIT and was depending on access to atomic reserves for inconspicuous reasons. In some cases it may be appropriate for them to use __GFP_HIGH. The second key hazard is callers that assembled their own combination of GFP flags instead of starting with something like GFP_KERNEL. They may now wish to specify __GFP_KSWAPD_RECLAIM. It's almost certainly harmless if it's missed in most cases as other activity will wake kswapd. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Acked-by: Vlastimil Babka <vbabka@suse.cz> Acked-by: Michal Hocko <mhocko@suse.com> Acked-by: Johannes Weiner <hannes@cmpxchg.org> Cc: Christoph Lameter <cl@linux.com> Cc: David Rientjes <rientjes@google.com> Cc: Vitaly Wool <vitalywool@gmail.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-07 08:28:21 +08:00
if (gfpflags_allow_blocking(flags))
local_irq_disable();
if (!page)
return NULL;
inc_slabs_node(s, page_to_nid(page), page->objects);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
return page;
}
static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
{
if (unlikely(flags & GFP_SLAB_BUG_MASK))
flags = kmalloc_fix_flags(flags);
return allocate_slab(s,
flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
}
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
static void __free_slab(struct kmem_cache *s, struct page *page)
{
int order = compound_order(page);
int pages = 1 << order;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS)) {
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
void *p;
slab_pad_check(s, page);
for_each_object(p, s, page_address(page),
page->objects)
check_object(s, page, p, SLUB_RED_INACTIVE);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
mm: sl[au]b: add knowledge of PFMEMALLOC reserve pages When a user or administrator requires swap for their application, they create a swap partition and file, format it with mkswap and activate it with swapon. Swap over the network is considered as an option in diskless systems. The two likely scenarios are when blade servers are used as part of a cluster where the form factor or maintenance costs do not allow the use of disks and thin clients. The Linux Terminal Server Project recommends the use of the Network Block Device (NBD) for swap according to the manual at https://sourceforge.net/projects/ltsp/files/Docs-Admin-Guide/LTSPManual.pdf/download There is also documentation and tutorials on how to setup swap over NBD at places like https://help.ubuntu.com/community/UbuntuLTSP/EnableNBDSWAP The nbd-client also documents the use of NBD as swap. Despite this, the fact is that a machine using NBD for swap can deadlock within minutes if swap is used intensively. This patch series addresses the problem. The core issue is that network block devices do not use mempools like normal block devices do. As the host cannot control where they receive packets from, they cannot reliably work out in advance how much memory they might need. Some years ago, Peter Zijlstra developed a series of patches that supported swap over an NFS that at least one distribution is carrying within their kernels. This patch series borrows very heavily from Peter's work to support swapping over NBD as a pre-requisite to supporting swap-over-NFS. The bulk of the complexity is concerned with preserving memory that is allocated from the PFMEMALLOC reserves for use by the network layer which is needed for both NBD and NFS. Patch 1 adds knowledge of the PFMEMALLOC reserves to SLAB and SLUB to preserve access to pages allocated under low memory situations to callers that are freeing memory. Patch 2 optimises the SLUB fast path to avoid pfmemalloc checks Patch 3 introduces __GFP_MEMALLOC to allow access to the PFMEMALLOC reserves without setting PFMEMALLOC. Patch 4 opens the possibility for softirqs to use PFMEMALLOC reserves for later use by network packet processing. Patch 5 only sets page->pfmemalloc when ALLOC_NO_WATERMARKS was required Patch 6 ignores memory policies when ALLOC_NO_WATERMARKS is set. Patches 7-12 allows network processing to use PFMEMALLOC reserves when the socket has been marked as being used by the VM to clean pages. If packets are received and stored in pages that were allocated under low-memory situations and are unrelated to the VM, the packets are dropped. Patch 11 reintroduces __skb_alloc_page which the networking folk may object to but is needed in some cases to propogate pfmemalloc from a newly allocated page to an skb. If there is a strong objection, this patch can be dropped with the impact being that swap-over-network will be slower in some cases but it should not fail. Patch 13 is a micro-optimisation to avoid a function call in the common case. Patch 14 tags NBD sockets as being SOCK_MEMALLOC so they can use PFMEMALLOC if necessary. Patch 15 notes that it is still possible for the PFMEMALLOC reserve to be depleted. To prevent this, direct reclaimers get throttled on a waitqueue if 50% of the PFMEMALLOC reserves are depleted. It is expected that kswapd and the direct reclaimers already running will clean enough pages for the low watermark to be reached and the throttled processes are woken up. Patch 16 adds a statistic to track how often processes get throttled Some basic performance testing was run using kernel builds, netperf on loopback for UDP and TCP, hackbench (pipes and sockets), iozone and sysbench. Each of them were expected to use the sl*b allocators reasonably heavily but there did not appear to be significant performance variances. For testing swap-over-NBD, a machine was booted with 2G of RAM with a swapfile backed by NBD. 8*NUM_CPU processes were started that create anonymous memory mappings and read them linearly in a loop. The total size of the mappings were 4*PHYSICAL_MEMORY to use swap heavily under memory pressure. Without the patches and using SLUB, the machine locks up within minutes and runs to completion with them applied. With SLAB, the story is different as an unpatched kernel run to completion. However, the patched kernel completed the test 45% faster. MICRO 3.5.0-rc2 3.5.0-rc2 vanilla swapnbd Unrecognised test vmscan-anon-mmap-write MMTests Statistics: duration Sys Time Running Test (seconds) 197.80 173.07 User+Sys Time Running Test (seconds) 206.96 182.03 Total Elapsed Time (seconds) 3240.70 1762.09 This patch: mm: sl[au]b: add knowledge of PFMEMALLOC reserve pages Allocations of pages below the min watermark run a risk of the machine hanging due to a lack of memory. To prevent this, only callers who have PF_MEMALLOC or TIF_MEMDIE set and are not processing an interrupt are allowed to allocate with ALLOC_NO_WATERMARKS. Once they are allocated to a slab though, nothing prevents other callers consuming free objects within those slabs. This patch limits access to slab pages that were alloced from the PFMEMALLOC reserves. When this patch is applied, pages allocated from below the low watermark are returned with page->pfmemalloc set and it is up to the caller to determine how the page should be protected. SLAB restricts access to any page with page->pfmemalloc set to callers which are known to able to access the PFMEMALLOC reserve. If one is not available, an attempt is made to allocate a new page rather than use a reserve. SLUB is a bit more relaxed in that it only records if the current per-CPU page was allocated from PFMEMALLOC reserve and uses another partial slab if the caller does not have the necessary GFP or process flags. This was found to be sufficient in tests to avoid hangs due to SLUB generally maintaining smaller lists than SLAB. In low-memory conditions it does mean that !PFMEMALLOC allocators can fail a slab allocation even though free objects are available because they are being preserved for callers that are freeing pages. [a.p.zijlstra@chello.nl: Original implementation] [sebastian@breakpoint.cc: Correct order of page flag clearing] Signed-off-by: Mel Gorman <mgorman@suse.de> Cc: David Miller <davem@davemloft.net> Cc: Neil Brown <neilb@suse.de> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Christie <michaelc@cs.wisc.edu> Cc: Eric B Munson <emunson@mgebm.net> Cc: Eric Dumazet <eric.dumazet@gmail.com> Cc: Sebastian Andrzej Siewior <sebastian@breakpoint.cc> Cc: Mel Gorman <mgorman@suse.de> Cc: Christoph Lameter <cl@linux.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-08-01 07:43:58 +08:00
__ClearPageSlabPfmemalloc(page);
__ClearPageSlab(page);
/* In union with page->mapping where page allocator expects NULL */
page->slab_cache = NULL;
if (current->reclaim_state)
current->reclaim_state->reclaimed_slab += pages;
unaccount_slab_page(page, order, s);
__free_pages(page, order);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
static void rcu_free_slab(struct rcu_head *h)
{
struct page *page = container_of(h, struct page, rcu_head);
__free_slab(page->slab_cache, page);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
static void free_slab(struct kmem_cache *s, struct page *page)
{
if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) {
call_rcu(&page->rcu_head, rcu_free_slab);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
} else
__free_slab(s, page);
}
static void discard_slab(struct kmem_cache *s, struct page *page)
{
dec_slabs_node(s, page_to_nid(page), page->objects);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
free_slab(s, page);
}
/*
* Management of partially allocated slabs.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*/
slub: do not assert not having lock in removing freed partial Vladimir reported the following issue: Commit c65c1877bd68 ("slub: use lockdep_assert_held") requires remove_partial() to be called with n->list_lock held, but free_partial() called from kmem_cache_close() on cache destruction does not follow this rule, leading to a warning: WARNING: CPU: 0 PID: 2787 at mm/slub.c:1536 __kmem_cache_shutdown+0x1b2/0x1f0() Modules linked in: CPU: 0 PID: 2787 Comm: modprobe Tainted: G W 3.14.0-rc1-mm1+ #1 Hardware name: 0000000000000600 ffff88003ae1dde8 ffffffff816d9583 0000000000000600 0000000000000000 ffff88003ae1de28 ffffffff8107c107 0000000000000000 ffff880037ab2b00 ffff88007c240d30 ffffea0001ee5280 ffffea0001ee52a0 Call Trace: __kmem_cache_shutdown+0x1b2/0x1f0 kmem_cache_destroy+0x43/0xf0 xfs_destroy_zones+0x103/0x110 [xfs] exit_xfs_fs+0x38/0x4e4 [xfs] SyS_delete_module+0x19a/0x1f0 system_call_fastpath+0x16/0x1b His solution was to add a spinlock in order to quiet lockdep. Although there would be no contention to adding the lock, that lock also requires disabling of interrupts which will have a larger impact on the system. Instead of adding a spinlock to a location where it is not needed for lockdep, make a __remove_partial() function that does not test if the list_lock is held, as no one should have it due to it being freed. Also added a __add_partial() function that does not do the lock validation either, as it is not needed for the creation of the cache. Signed-off-by: Steven Rostedt <rostedt@goodmis.org> Reported-by: Vladimir Davydov <vdavydov@parallels.com> Suggested-by: David Rientjes <rientjes@google.com> Acked-by: David Rientjes <rientjes@google.com> Acked-by: Vladimir Davydov <vdavydov@parallels.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-02-11 06:25:46 +08:00
static inline void
__add_partial(struct kmem_cache_node *n, struct page *page, int tail)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
{
n->nr_partial++;
if (tail == DEACTIVATE_TO_TAIL)
list_add_tail(&page->slab_list, &n->partial);
else
list_add(&page->slab_list, &n->partial);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
slub: do not assert not having lock in removing freed partial Vladimir reported the following issue: Commit c65c1877bd68 ("slub: use lockdep_assert_held") requires remove_partial() to be called with n->list_lock held, but free_partial() called from kmem_cache_close() on cache destruction does not follow this rule, leading to a warning: WARNING: CPU: 0 PID: 2787 at mm/slub.c:1536 __kmem_cache_shutdown+0x1b2/0x1f0() Modules linked in: CPU: 0 PID: 2787 Comm: modprobe Tainted: G W 3.14.0-rc1-mm1+ #1 Hardware name: 0000000000000600 ffff88003ae1dde8 ffffffff816d9583 0000000000000600 0000000000000000 ffff88003ae1de28 ffffffff8107c107 0000000000000000 ffff880037ab2b00 ffff88007c240d30 ffffea0001ee5280 ffffea0001ee52a0 Call Trace: __kmem_cache_shutdown+0x1b2/0x1f0 kmem_cache_destroy+0x43/0xf0 xfs_destroy_zones+0x103/0x110 [xfs] exit_xfs_fs+0x38/0x4e4 [xfs] SyS_delete_module+0x19a/0x1f0 system_call_fastpath+0x16/0x1b His solution was to add a spinlock in order to quiet lockdep. Although there would be no contention to adding the lock, that lock also requires disabling of interrupts which will have a larger impact on the system. Instead of adding a spinlock to a location where it is not needed for lockdep, make a __remove_partial() function that does not test if the list_lock is held, as no one should have it due to it being freed. Also added a __add_partial() function that does not do the lock validation either, as it is not needed for the creation of the cache. Signed-off-by: Steven Rostedt <rostedt@goodmis.org> Reported-by: Vladimir Davydov <vdavydov@parallels.com> Suggested-by: David Rientjes <rientjes@google.com> Acked-by: David Rientjes <rientjes@google.com> Acked-by: Vladimir Davydov <vdavydov@parallels.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-02-11 06:25:46 +08:00
static inline void add_partial(struct kmem_cache_node *n,
struct page *page, int tail)
{
lockdep_assert_held(&n->list_lock);
slub: do not assert not having lock in removing freed partial Vladimir reported the following issue: Commit c65c1877bd68 ("slub: use lockdep_assert_held") requires remove_partial() to be called with n->list_lock held, but free_partial() called from kmem_cache_close() on cache destruction does not follow this rule, leading to a warning: WARNING: CPU: 0 PID: 2787 at mm/slub.c:1536 __kmem_cache_shutdown+0x1b2/0x1f0() Modules linked in: CPU: 0 PID: 2787 Comm: modprobe Tainted: G W 3.14.0-rc1-mm1+ #1 Hardware name: 0000000000000600 ffff88003ae1dde8 ffffffff816d9583 0000000000000600 0000000000000000 ffff88003ae1de28 ffffffff8107c107 0000000000000000 ffff880037ab2b00 ffff88007c240d30 ffffea0001ee5280 ffffea0001ee52a0 Call Trace: __kmem_cache_shutdown+0x1b2/0x1f0 kmem_cache_destroy+0x43/0xf0 xfs_destroy_zones+0x103/0x110 [xfs] exit_xfs_fs+0x38/0x4e4 [xfs] SyS_delete_module+0x19a/0x1f0 system_call_fastpath+0x16/0x1b His solution was to add a spinlock in order to quiet lockdep. Although there would be no contention to adding the lock, that lock also requires disabling of interrupts which will have a larger impact on the system. Instead of adding a spinlock to a location where it is not needed for lockdep, make a __remove_partial() function that does not test if the list_lock is held, as no one should have it due to it being freed. Also added a __add_partial() function that does not do the lock validation either, as it is not needed for the creation of the cache. Signed-off-by: Steven Rostedt <rostedt@goodmis.org> Reported-by: Vladimir Davydov <vdavydov@parallels.com> Suggested-by: David Rientjes <rientjes@google.com> Acked-by: David Rientjes <rientjes@google.com> Acked-by: Vladimir Davydov <vdavydov@parallels.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-02-11 06:25:46 +08:00
__add_partial(n, page, tail);
}
slub: do not assert not having lock in removing freed partial Vladimir reported the following issue: Commit c65c1877bd68 ("slub: use lockdep_assert_held") requires remove_partial() to be called with n->list_lock held, but free_partial() called from kmem_cache_close() on cache destruction does not follow this rule, leading to a warning: WARNING: CPU: 0 PID: 2787 at mm/slub.c:1536 __kmem_cache_shutdown+0x1b2/0x1f0() Modules linked in: CPU: 0 PID: 2787 Comm: modprobe Tainted: G W 3.14.0-rc1-mm1+ #1 Hardware name: 0000000000000600 ffff88003ae1dde8 ffffffff816d9583 0000000000000600 0000000000000000 ffff88003ae1de28 ffffffff8107c107 0000000000000000 ffff880037ab2b00 ffff88007c240d30 ffffea0001ee5280 ffffea0001ee52a0 Call Trace: __kmem_cache_shutdown+0x1b2/0x1f0 kmem_cache_destroy+0x43/0xf0 xfs_destroy_zones+0x103/0x110 [xfs] exit_xfs_fs+0x38/0x4e4 [xfs] SyS_delete_module+0x19a/0x1f0 system_call_fastpath+0x16/0x1b His solution was to add a spinlock in order to quiet lockdep. Although there would be no contention to adding the lock, that lock also requires disabling of interrupts which will have a larger impact on the system. Instead of adding a spinlock to a location where it is not needed for lockdep, make a __remove_partial() function that does not test if the list_lock is held, as no one should have it due to it being freed. Also added a __add_partial() function that does not do the lock validation either, as it is not needed for the creation of the cache. Signed-off-by: Steven Rostedt <rostedt@goodmis.org> Reported-by: Vladimir Davydov <vdavydov@parallels.com> Suggested-by: David Rientjes <rientjes@google.com> Acked-by: David Rientjes <rientjes@google.com> Acked-by: Vladimir Davydov <vdavydov@parallels.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-02-11 06:25:46 +08:00
static inline void remove_partial(struct kmem_cache_node *n,
struct page *page)
{
lockdep_assert_held(&n->list_lock);
list_del(&page->slab_list);
mm: slab: free kmem_cache_node after destroy sysfs file When slub_debug alloc_calls_show is enabled we will try to track location and user of slab object on each online node, kmem_cache_node structure and cpu_cache/cpu_slub shouldn't be freed till there is the last reference to sysfs file. This fixes the following panic: BUG: unable to handle kernel NULL pointer dereference at 0000000000000020 IP: list_locations+0x169/0x4e0 PGD 257304067 PUD 438456067 PMD 0 Oops: 0000 [#1] SMP CPU: 3 PID: 973074 Comm: cat ve: 0 Not tainted 3.10.0-229.7.2.ovz.9.30-00007-japdoll-dirty #2 9.30 Hardware name: DEPO Computers To Be Filled By O.E.M./H67DE3, BIOS L1.60c 07/14/2011 task: ffff88042a5dc5b0 ti: ffff88037f8d8000 task.ti: ffff88037f8d8000 RIP: list_locations+0x169/0x4e0 Call Trace: alloc_calls_show+0x1d/0x30 slab_attr_show+0x1b/0x30 sysfs_read_file+0x9a/0x1a0 vfs_read+0x9c/0x170 SyS_read+0x58/0xb0 system_call_fastpath+0x16/0x1b Code: 5e 07 12 00 b9 00 04 00 00 3d 00 04 00 00 0f 4f c1 3d 00 04 00 00 89 45 b0 0f 84 c3 00 00 00 48 63 45 b0 49 8b 9c c4 f8 00 00 00 <48> 8b 43 20 48 85 c0 74 b6 48 89 df e8 46 37 44 00 48 8b 53 10 CR2: 0000000000000020 Separated __kmem_cache_release from __kmem_cache_shutdown which now called on slab_kmem_cache_release (after the last reference to sysfs file object has dropped). Reintroduced locking in free_partial as sysfs file might access cache's partial list after shutdowning - partial revert of the commit 69cb8e6b7c29 ("slub: free slabs without holding locks"). Zap __remove_partial and use remove_partial (w/o underscores) as free_partial now takes list_lock which s partial revert for commit 1e4dd9461fab ("slub: do not assert not having lock in removing freed partial") Signed-off-by: Dmitry Safonov <dsafonov@virtuozzo.com> Suggested-by: Vladimir Davydov <vdavydov@virtuozzo.com> Acked-by: Vladimir Davydov <vdavydov@virtuozzo.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-02-18 05:11:37 +08:00
n->nr_partial--;
slub: do not assert not having lock in removing freed partial Vladimir reported the following issue: Commit c65c1877bd68 ("slub: use lockdep_assert_held") requires remove_partial() to be called with n->list_lock held, but free_partial() called from kmem_cache_close() on cache destruction does not follow this rule, leading to a warning: WARNING: CPU: 0 PID: 2787 at mm/slub.c:1536 __kmem_cache_shutdown+0x1b2/0x1f0() Modules linked in: CPU: 0 PID: 2787 Comm: modprobe Tainted: G W 3.14.0-rc1-mm1+ #1 Hardware name: 0000000000000600 ffff88003ae1dde8 ffffffff816d9583 0000000000000600 0000000000000000 ffff88003ae1de28 ffffffff8107c107 0000000000000000 ffff880037ab2b00 ffff88007c240d30 ffffea0001ee5280 ffffea0001ee52a0 Call Trace: __kmem_cache_shutdown+0x1b2/0x1f0 kmem_cache_destroy+0x43/0xf0 xfs_destroy_zones+0x103/0x110 [xfs] exit_xfs_fs+0x38/0x4e4 [xfs] SyS_delete_module+0x19a/0x1f0 system_call_fastpath+0x16/0x1b His solution was to add a spinlock in order to quiet lockdep. Although there would be no contention to adding the lock, that lock also requires disabling of interrupts which will have a larger impact on the system. Instead of adding a spinlock to a location where it is not needed for lockdep, make a __remove_partial() function that does not test if the list_lock is held, as no one should have it due to it being freed. Also added a __add_partial() function that does not do the lock validation either, as it is not needed for the creation of the cache. Signed-off-by: Steven Rostedt <rostedt@goodmis.org> Reported-by: Vladimir Davydov <vdavydov@parallels.com> Suggested-by: David Rientjes <rientjes@google.com> Acked-by: David Rientjes <rientjes@google.com> Acked-by: Vladimir Davydov <vdavydov@parallels.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-02-11 06:25:46 +08:00
}
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/*
* Remove slab from the partial list, freeze it and
* return the pointer to the freelist.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*
* Returns a list of objects or NULL if it fails.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*/
static inline void *acquire_slab(struct kmem_cache *s,
struct kmem_cache_node *n, struct page *page,
int mode, int *objects)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
{
void *freelist;
unsigned long counters;
struct page new;
lockdep_assert_held(&n->list_lock);
/*
* Zap the freelist and set the frozen bit.
* The old freelist is the list of objects for the
* per cpu allocation list.
*/
freelist = page->freelist;
counters = page->counters;
new.counters = counters;
*objects = new.objects - new.inuse;
if (mode) {
new.inuse = page->objects;
new.freelist = NULL;
} else {
new.freelist = freelist;
}
VM_BUG_ON(new.frozen);
new.frozen = 1;
if (!__cmpxchg_double_slab(s, page,
freelist, counters,
new.freelist, new.counters,
"acquire_slab"))
return NULL;
remove_partial(n, page);
WARN_ON(!freelist);
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
return freelist;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
slub: consider pfmemalloc_match() in get_partial_node() get_partial() is currently not checking pfmemalloc_match() meaning that it is possible for pfmemalloc pages to leak to non-pfmemalloc users. This is a problem in the following situation. Assume that there is a request from normal allocation and there are no objects in the per-cpu cache and no node-partial slab. In this case, slab_alloc enters the slow path and new_slab_objects() is called which may return a PFMEMALLOC page. As the current user is not allowed to access PFMEMALLOC page, deactivate_slab() is called ([5091b74a: mm: slub: optimise the SLUB fast path to avoid pfmemalloc checks]) and returns an object from PFMEMALLOC page. Next time, when we get another request from normal allocation, slab_alloc() enters the slow-path and calls new_slab_objects(). In new_slab_objects(), we call get_partial() and get a partial slab which was just deactivated but is a pfmemalloc page. We extract one object from it and re-deactivate. "deactivate -> re-get in get_partial -> re-deactivate" occures repeatedly. As a result, access to PFMEMALLOC page is not properly restricted and it can cause a performance degradation due to frequent deactivation. deactivation frequently. This patch changes get_partial_node() to take pfmemalloc_match() into account and prevents the "deactivate -> re-get in get_partial() scenario. Instead, new_slab() is called. Signed-off-by: Joonsoo Kim <js1304@gmail.com> Acked-by: David Rientjes <rientjes@google.com> Signed-off-by: Mel Gorman <mgorman@suse.de> Cc: David Miller <davem@davemloft.net> Cc: Chuck Lever <chuck.lever@oracle.com> Cc: Pekka Enberg <penberg@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-09-18 05:09:09 +08:00
static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/*
* Try to allocate a partial slab from a specific node.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*/
slub: consider pfmemalloc_match() in get_partial_node() get_partial() is currently not checking pfmemalloc_match() meaning that it is possible for pfmemalloc pages to leak to non-pfmemalloc users. This is a problem in the following situation. Assume that there is a request from normal allocation and there are no objects in the per-cpu cache and no node-partial slab. In this case, slab_alloc enters the slow path and new_slab_objects() is called which may return a PFMEMALLOC page. As the current user is not allowed to access PFMEMALLOC page, deactivate_slab() is called ([5091b74a: mm: slub: optimise the SLUB fast path to avoid pfmemalloc checks]) and returns an object from PFMEMALLOC page. Next time, when we get another request from normal allocation, slab_alloc() enters the slow-path and calls new_slab_objects(). In new_slab_objects(), we call get_partial() and get a partial slab which was just deactivated but is a pfmemalloc page. We extract one object from it and re-deactivate. "deactivate -> re-get in get_partial -> re-deactivate" occures repeatedly. As a result, access to PFMEMALLOC page is not properly restricted and it can cause a performance degradation due to frequent deactivation. deactivation frequently. This patch changes get_partial_node() to take pfmemalloc_match() into account and prevents the "deactivate -> re-get in get_partial() scenario. Instead, new_slab() is called. Signed-off-by: Joonsoo Kim <js1304@gmail.com> Acked-by: David Rientjes <rientjes@google.com> Signed-off-by: Mel Gorman <mgorman@suse.de> Cc: David Miller <davem@davemloft.net> Cc: Chuck Lever <chuck.lever@oracle.com> Cc: Pekka Enberg <penberg@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-09-18 05:09:09 +08:00
static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
struct kmem_cache_cpu *c, gfp_t flags)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
{
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
struct page *page, *page2;
void *object = NULL;
unsigned int available = 0;
int objects;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/*
* Racy check. If we mistakenly see no partial slabs then we
* just allocate an empty slab. If we mistakenly try to get a
* partial slab and there is none available then get_partial()
* will return NULL.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*/
if (!n || !n->nr_partial)
return NULL;
spin_lock(&n->list_lock);
list_for_each_entry_safe(page, page2, &n->partial, slab_list) {
slub: consider pfmemalloc_match() in get_partial_node() get_partial() is currently not checking pfmemalloc_match() meaning that it is possible for pfmemalloc pages to leak to non-pfmemalloc users. This is a problem in the following situation. Assume that there is a request from normal allocation and there are no objects in the per-cpu cache and no node-partial slab. In this case, slab_alloc enters the slow path and new_slab_objects() is called which may return a PFMEMALLOC page. As the current user is not allowed to access PFMEMALLOC page, deactivate_slab() is called ([5091b74a: mm: slub: optimise the SLUB fast path to avoid pfmemalloc checks]) and returns an object from PFMEMALLOC page. Next time, when we get another request from normal allocation, slab_alloc() enters the slow-path and calls new_slab_objects(). In new_slab_objects(), we call get_partial() and get a partial slab which was just deactivated but is a pfmemalloc page. We extract one object from it and re-deactivate. "deactivate -> re-get in get_partial -> re-deactivate" occures repeatedly. As a result, access to PFMEMALLOC page is not properly restricted and it can cause a performance degradation due to frequent deactivation. deactivation frequently. This patch changes get_partial_node() to take pfmemalloc_match() into account and prevents the "deactivate -> re-get in get_partial() scenario. Instead, new_slab() is called. Signed-off-by: Joonsoo Kim <js1304@gmail.com> Acked-by: David Rientjes <rientjes@google.com> Signed-off-by: Mel Gorman <mgorman@suse.de> Cc: David Miller <davem@davemloft.net> Cc: Chuck Lever <chuck.lever@oracle.com> Cc: Pekka Enberg <penberg@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-09-18 05:09:09 +08:00
void *t;
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
slub: consider pfmemalloc_match() in get_partial_node() get_partial() is currently not checking pfmemalloc_match() meaning that it is possible for pfmemalloc pages to leak to non-pfmemalloc users. This is a problem in the following situation. Assume that there is a request from normal allocation and there are no objects in the per-cpu cache and no node-partial slab. In this case, slab_alloc enters the slow path and new_slab_objects() is called which may return a PFMEMALLOC page. As the current user is not allowed to access PFMEMALLOC page, deactivate_slab() is called ([5091b74a: mm: slub: optimise the SLUB fast path to avoid pfmemalloc checks]) and returns an object from PFMEMALLOC page. Next time, when we get another request from normal allocation, slab_alloc() enters the slow-path and calls new_slab_objects(). In new_slab_objects(), we call get_partial() and get a partial slab which was just deactivated but is a pfmemalloc page. We extract one object from it and re-deactivate. "deactivate -> re-get in get_partial -> re-deactivate" occures repeatedly. As a result, access to PFMEMALLOC page is not properly restricted and it can cause a performance degradation due to frequent deactivation. deactivation frequently. This patch changes get_partial_node() to take pfmemalloc_match() into account and prevents the "deactivate -> re-get in get_partial() scenario. Instead, new_slab() is called. Signed-off-by: Joonsoo Kim <js1304@gmail.com> Acked-by: David Rientjes <rientjes@google.com> Signed-off-by: Mel Gorman <mgorman@suse.de> Cc: David Miller <davem@davemloft.net> Cc: Chuck Lever <chuck.lever@oracle.com> Cc: Pekka Enberg <penberg@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-09-18 05:09:09 +08:00
if (!pfmemalloc_match(page, flags))
continue;
t = acquire_slab(s, n, page, object == NULL, &objects);
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
if (!t)
Revert "mm, slub: consider rest of partial list if acquire_slab() fails" This reverts commit 8ff60eb052eeba95cfb3efe16b08c9199f8121cf. The kernel test robot reports a huge performance regression due to the commit, and the reason seems fairly straightforward: when there is contention on the page list (which is what causes acquire_slab() to fail), we do _not_ want to just loop and try again, because that will transfer the contention to the 'n->list_lock' spinlock we hold, and just make things even worse. This is admittedly likely a problem only on big machines - the kernel test robot report comes from a 96-thread dual socket Intel Xeon Gold 6252 setup, but the regression there really is quite noticeable: -47.9% regression of stress-ng.rawpkt.ops_per_sec and the commit that was marked as being fixed (7ced37197196: "slub: Acquire_slab() avoid loop") actually did the loop exit early very intentionally (the hint being that "avoid loop" part of that commit message), exactly to avoid this issue. The correct thing to do may be to pick some kind of reasonable middle ground: instead of breaking out of the loop on the very first sign of contention, or trying over and over and over again, the right thing may be to re-try _once_, and then give up on the second failure (or pick your favorite value for "once"..). Reported-by: kernel test robot <oliver.sang@intel.com> Link: https://lore.kernel.org/lkml/20210301080404.GF12822@xsang-OptiPlex-9020/ Cc: Jann Horn <jannh@google.com> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Acked-by: Christoph Lameter <cl@linux.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-03-11 02:18:04 +08:00
break;
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
available += objects;
if (!object) {
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
c->page = page;
stat(s, ALLOC_FROM_PARTIAL);
object = t;
} else {
put_cpu_partial(s, page, 0);
stat(s, CPU_PARTIAL_NODE);
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
}
if (!kmem_cache_has_cpu_partial(s)
|| available > slub_cpu_partial(s) / 2)
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
break;
}
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
spin_unlock(&n->list_lock);
return object;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
/*
* Get a page from somewhere. Search in increasing NUMA distances.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*/
static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
struct kmem_cache_cpu *c)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
{
#ifdef CONFIG_NUMA
struct zonelist *zonelist;
mm: have zonelist contains structs with both a zone pointer and zone_idx Filtering zonelists requires very frequent use of zone_idx(). This is costly as it involves a lookup of another structure and a substraction operation. As the zone_idx is often required, it should be quickly accessible. The node idx could also be stored here if it was found that accessing zone->node is significant which may be the case on workloads where nodemasks are heavily used. This patch introduces a struct zoneref to store a zone pointer and a zone index. The zonelist then consists of an array of these struct zonerefs which are looked up as necessary. Helpers are given for accessing the zone index as well as the node index. [kamezawa.hiroyu@jp.fujitsu.com: Suggested struct zoneref instead of embedding information in pointers] [hugh@veritas.com: mm-have-zonelist: fix memcg ooms] [hugh@veritas.com: just return do_try_to_free_pages] [hugh@veritas.com: do_try_to_free_pages gfp_mask redundant] Signed-off-by: Mel Gorman <mel@csn.ul.ie> Acked-by: Christoph Lameter <clameter@sgi.com> Acked-by: David Rientjes <rientjes@google.com> Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Christoph Lameter <clameter@sgi.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-28 17:12:17 +08:00
struct zoneref *z;
struct zone *zone;
enum zone_type highest_zoneidx = gfp_zone(flags);
void *object;
cpuset: mm: reduce large amounts of memory barrier related damage v3 Commit c0ff7453bb5c ("cpuset,mm: fix no node to alloc memory when changing cpuset's mems") wins a super prize for the largest number of memory barriers entered into fast paths for one commit. [get|put]_mems_allowed is incredibly heavy with pairs of full memory barriers inserted into a number of hot paths. This was detected while investigating at large page allocator slowdown introduced some time after 2.6.32. The largest portion of this overhead was shown by oprofile to be at an mfence introduced by this commit into the page allocator hot path. For extra style points, the commit introduced the use of yield() in an implementation of what looks like a spinning mutex. This patch replaces the full memory barriers on both read and write sides with a sequence counter with just read barriers on the fast path side. This is much cheaper on some architectures, including x86. The main bulk of the patch is the retry logic if the nodemask changes in a manner that can cause a false failure. While updating the nodemask, a check is made to see if a false failure is a risk. If it is, the sequence number gets bumped and parallel allocators will briefly stall while the nodemask update takes place. In a page fault test microbenchmark, oprofile samples from __alloc_pages_nodemask went from 4.53% of all samples to 1.15%. The actual results were 3.3.0-rc3 3.3.0-rc3 rc3-vanilla nobarrier-v2r1 Clients 1 UserTime 0.07 ( 0.00%) 0.08 (-14.19%) Clients 2 UserTime 0.07 ( 0.00%) 0.07 ( 2.72%) Clients 4 UserTime 0.08 ( 0.00%) 0.07 ( 3.29%) Clients 1 SysTime 0.70 ( 0.00%) 0.65 ( 6.65%) Clients 2 SysTime 0.85 ( 0.00%) 0.82 ( 3.65%) Clients 4 SysTime 1.41 ( 0.00%) 1.41 ( 0.32%) Clients 1 WallTime 0.77 ( 0.00%) 0.74 ( 4.19%) Clients 2 WallTime 0.47 ( 0.00%) 0.45 ( 3.73%) Clients 4 WallTime 0.38 ( 0.00%) 0.37 ( 1.58%) Clients 1 Flt/sec/cpu 497620.28 ( 0.00%) 520294.53 ( 4.56%) Clients 2 Flt/sec/cpu 414639.05 ( 0.00%) 429882.01 ( 3.68%) Clients 4 Flt/sec/cpu 257959.16 ( 0.00%) 258761.48 ( 0.31%) Clients 1 Flt/sec 495161.39 ( 0.00%) 517292.87 ( 4.47%) Clients 2 Flt/sec 820325.95 ( 0.00%) 850289.77 ( 3.65%) Clients 4 Flt/sec 1020068.93 ( 0.00%) 1022674.06 ( 0.26%) MMTests Statistics: duration Sys Time Running Test (seconds) 135.68 132.17 User+Sys Time Running Test (seconds) 164.2 160.13 Total Elapsed Time (seconds) 123.46 120.87 The overall improvement is small but the System CPU time is much improved and roughly in correlation to what oprofile reported (these performance figures are without profiling so skew is expected). The actual number of page faults is noticeably improved. For benchmarks like kernel builds, the overall benefit is marginal but the system CPU time is slightly reduced. To test the actual bug the commit fixed I opened two terminals. The first ran within a cpuset and continually ran a small program that faulted 100M of anonymous data. In a second window, the nodemask of the cpuset was continually randomised in a loop. Without the commit, the program would fail every so often (usually within 10 seconds) and obviously with the commit everything worked fine. With this patch applied, it also worked fine so the fix should be functionally equivalent. Signed-off-by: Mel Gorman <mgorman@suse.de> Cc: Miao Xie <miaox@cn.fujitsu.com> Cc: David Rientjes <rientjes@google.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Christoph Lameter <cl@linux.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-03-22 07:34:11 +08:00
unsigned int cpuset_mems_cookie;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/*
* The defrag ratio allows a configuration of the tradeoffs between
* inter node defragmentation and node local allocations. A lower
* defrag_ratio increases the tendency to do local allocations
* instead of attempting to obtain partial slabs from other nodes.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*
* If the defrag_ratio is set to 0 then kmalloc() always
* returns node local objects. If the ratio is higher then kmalloc()
* may return off node objects because partial slabs are obtained
* from other nodes and filled up.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*
* If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
* (which makes defrag_ratio = 1000) then every (well almost)
* allocation will first attempt to defrag slab caches on other nodes.
* This means scanning over all nodes to look for partial slabs which
* may be expensive if we do it every time we are trying to find a slab
* with available objects.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*/
if (!s->remote_node_defrag_ratio ||
get_cycles() % 1024 > s->remote_node_defrag_ratio)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
return NULL;
cpuset: mm: reduce large amounts of memory barrier related damage v3 Commit c0ff7453bb5c ("cpuset,mm: fix no node to alloc memory when changing cpuset's mems") wins a super prize for the largest number of memory barriers entered into fast paths for one commit. [get|put]_mems_allowed is incredibly heavy with pairs of full memory barriers inserted into a number of hot paths. This was detected while investigating at large page allocator slowdown introduced some time after 2.6.32. The largest portion of this overhead was shown by oprofile to be at an mfence introduced by this commit into the page allocator hot path. For extra style points, the commit introduced the use of yield() in an implementation of what looks like a spinning mutex. This patch replaces the full memory barriers on both read and write sides with a sequence counter with just read barriers on the fast path side. This is much cheaper on some architectures, including x86. The main bulk of the patch is the retry logic if the nodemask changes in a manner that can cause a false failure. While updating the nodemask, a check is made to see if a false failure is a risk. If it is, the sequence number gets bumped and parallel allocators will briefly stall while the nodemask update takes place. In a page fault test microbenchmark, oprofile samples from __alloc_pages_nodemask went from 4.53% of all samples to 1.15%. The actual results were 3.3.0-rc3 3.3.0-rc3 rc3-vanilla nobarrier-v2r1 Clients 1 UserTime 0.07 ( 0.00%) 0.08 (-14.19%) Clients 2 UserTime 0.07 ( 0.00%) 0.07 ( 2.72%) Clients 4 UserTime 0.08 ( 0.00%) 0.07 ( 3.29%) Clients 1 SysTime 0.70 ( 0.00%) 0.65 ( 6.65%) Clients 2 SysTime 0.85 ( 0.00%) 0.82 ( 3.65%) Clients 4 SysTime 1.41 ( 0.00%) 1.41 ( 0.32%) Clients 1 WallTime 0.77 ( 0.00%) 0.74 ( 4.19%) Clients 2 WallTime 0.47 ( 0.00%) 0.45 ( 3.73%) Clients 4 WallTime 0.38 ( 0.00%) 0.37 ( 1.58%) Clients 1 Flt/sec/cpu 497620.28 ( 0.00%) 520294.53 ( 4.56%) Clients 2 Flt/sec/cpu 414639.05 ( 0.00%) 429882.01 ( 3.68%) Clients 4 Flt/sec/cpu 257959.16 ( 0.00%) 258761.48 ( 0.31%) Clients 1 Flt/sec 495161.39 ( 0.00%) 517292.87 ( 4.47%) Clients 2 Flt/sec 820325.95 ( 0.00%) 850289.77 ( 3.65%) Clients 4 Flt/sec 1020068.93 ( 0.00%) 1022674.06 ( 0.26%) MMTests Statistics: duration Sys Time Running Test (seconds) 135.68 132.17 User+Sys Time Running Test (seconds) 164.2 160.13 Total Elapsed Time (seconds) 123.46 120.87 The overall improvement is small but the System CPU time is much improved and roughly in correlation to what oprofile reported (these performance figures are without profiling so skew is expected). The actual number of page faults is noticeably improved. For benchmarks like kernel builds, the overall benefit is marginal but the system CPU time is slightly reduced. To test the actual bug the commit fixed I opened two terminals. The first ran within a cpuset and continually ran a small program that faulted 100M of anonymous data. In a second window, the nodemask of the cpuset was continually randomised in a loop. Without the commit, the program would fail every so often (usually within 10 seconds) and obviously with the commit everything worked fine. With this patch applied, it also worked fine so the fix should be functionally equivalent. Signed-off-by: Mel Gorman <mgorman@suse.de> Cc: Miao Xie <miaox@cn.fujitsu.com> Cc: David Rientjes <rientjes@google.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Christoph Lameter <cl@linux.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-03-22 07:34:11 +08:00
do {
cpuset_mems_cookie = read_mems_allowed_begin();
zonelist = node_zonelist(mempolicy_slab_node(), flags);
for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) {
cpuset: mm: reduce large amounts of memory barrier related damage v3 Commit c0ff7453bb5c ("cpuset,mm: fix no node to alloc memory when changing cpuset's mems") wins a super prize for the largest number of memory barriers entered into fast paths for one commit. [get|put]_mems_allowed is incredibly heavy with pairs of full memory barriers inserted into a number of hot paths. This was detected while investigating at large page allocator slowdown introduced some time after 2.6.32. The largest portion of this overhead was shown by oprofile to be at an mfence introduced by this commit into the page allocator hot path. For extra style points, the commit introduced the use of yield() in an implementation of what looks like a spinning mutex. This patch replaces the full memory barriers on both read and write sides with a sequence counter with just read barriers on the fast path side. This is much cheaper on some architectures, including x86. The main bulk of the patch is the retry logic if the nodemask changes in a manner that can cause a false failure. While updating the nodemask, a check is made to see if a false failure is a risk. If it is, the sequence number gets bumped and parallel allocators will briefly stall while the nodemask update takes place. In a page fault test microbenchmark, oprofile samples from __alloc_pages_nodemask went from 4.53% of all samples to 1.15%. The actual results were 3.3.0-rc3 3.3.0-rc3 rc3-vanilla nobarrier-v2r1 Clients 1 UserTime 0.07 ( 0.00%) 0.08 (-14.19%) Clients 2 UserTime 0.07 ( 0.00%) 0.07 ( 2.72%) Clients 4 UserTime 0.08 ( 0.00%) 0.07 ( 3.29%) Clients 1 SysTime 0.70 ( 0.00%) 0.65 ( 6.65%) Clients 2 SysTime 0.85 ( 0.00%) 0.82 ( 3.65%) Clients 4 SysTime 1.41 ( 0.00%) 1.41 ( 0.32%) Clients 1 WallTime 0.77 ( 0.00%) 0.74 ( 4.19%) Clients 2 WallTime 0.47 ( 0.00%) 0.45 ( 3.73%) Clients 4 WallTime 0.38 ( 0.00%) 0.37 ( 1.58%) Clients 1 Flt/sec/cpu 497620.28 ( 0.00%) 520294.53 ( 4.56%) Clients 2 Flt/sec/cpu 414639.05 ( 0.00%) 429882.01 ( 3.68%) Clients 4 Flt/sec/cpu 257959.16 ( 0.00%) 258761.48 ( 0.31%) Clients 1 Flt/sec 495161.39 ( 0.00%) 517292.87 ( 4.47%) Clients 2 Flt/sec 820325.95 ( 0.00%) 850289.77 ( 3.65%) Clients 4 Flt/sec 1020068.93 ( 0.00%) 1022674.06 ( 0.26%) MMTests Statistics: duration Sys Time Running Test (seconds) 135.68 132.17 User+Sys Time Running Test (seconds) 164.2 160.13 Total Elapsed Time (seconds) 123.46 120.87 The overall improvement is small but the System CPU time is much improved and roughly in correlation to what oprofile reported (these performance figures are without profiling so skew is expected). The actual number of page faults is noticeably improved. For benchmarks like kernel builds, the overall benefit is marginal but the system CPU time is slightly reduced. To test the actual bug the commit fixed I opened two terminals. The first ran within a cpuset and continually ran a small program that faulted 100M of anonymous data. In a second window, the nodemask of the cpuset was continually randomised in a loop. Without the commit, the program would fail every so often (usually within 10 seconds) and obviously with the commit everything worked fine. With this patch applied, it also worked fine so the fix should be functionally equivalent. Signed-off-by: Mel Gorman <mgorman@suse.de> Cc: Miao Xie <miaox@cn.fujitsu.com> Cc: David Rientjes <rientjes@google.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Christoph Lameter <cl@linux.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-03-22 07:34:11 +08:00
struct kmem_cache_node *n;
n = get_node(s, zone_to_nid(zone));
if (n && cpuset_zone_allowed(zone, flags) &&
cpuset: mm: reduce large amounts of memory barrier related damage v3 Commit c0ff7453bb5c ("cpuset,mm: fix no node to alloc memory when changing cpuset's mems") wins a super prize for the largest number of memory barriers entered into fast paths for one commit. [get|put]_mems_allowed is incredibly heavy with pairs of full memory barriers inserted into a number of hot paths. This was detected while investigating at large page allocator slowdown introduced some time after 2.6.32. The largest portion of this overhead was shown by oprofile to be at an mfence introduced by this commit into the page allocator hot path. For extra style points, the commit introduced the use of yield() in an implementation of what looks like a spinning mutex. This patch replaces the full memory barriers on both read and write sides with a sequence counter with just read barriers on the fast path side. This is much cheaper on some architectures, including x86. The main bulk of the patch is the retry logic if the nodemask changes in a manner that can cause a false failure. While updating the nodemask, a check is made to see if a false failure is a risk. If it is, the sequence number gets bumped and parallel allocators will briefly stall while the nodemask update takes place. In a page fault test microbenchmark, oprofile samples from __alloc_pages_nodemask went from 4.53% of all samples to 1.15%. The actual results were 3.3.0-rc3 3.3.0-rc3 rc3-vanilla nobarrier-v2r1 Clients 1 UserTime 0.07 ( 0.00%) 0.08 (-14.19%) Clients 2 UserTime 0.07 ( 0.00%) 0.07 ( 2.72%) Clients 4 UserTime 0.08 ( 0.00%) 0.07 ( 3.29%) Clients 1 SysTime 0.70 ( 0.00%) 0.65 ( 6.65%) Clients 2 SysTime 0.85 ( 0.00%) 0.82 ( 3.65%) Clients 4 SysTime 1.41 ( 0.00%) 1.41 ( 0.32%) Clients 1 WallTime 0.77 ( 0.00%) 0.74 ( 4.19%) Clients 2 WallTime 0.47 ( 0.00%) 0.45 ( 3.73%) Clients 4 WallTime 0.38 ( 0.00%) 0.37 ( 1.58%) Clients 1 Flt/sec/cpu 497620.28 ( 0.00%) 520294.53 ( 4.56%) Clients 2 Flt/sec/cpu 414639.05 ( 0.00%) 429882.01 ( 3.68%) Clients 4 Flt/sec/cpu 257959.16 ( 0.00%) 258761.48 ( 0.31%) Clients 1 Flt/sec 495161.39 ( 0.00%) 517292.87 ( 4.47%) Clients 2 Flt/sec 820325.95 ( 0.00%) 850289.77 ( 3.65%) Clients 4 Flt/sec 1020068.93 ( 0.00%) 1022674.06 ( 0.26%) MMTests Statistics: duration Sys Time Running Test (seconds) 135.68 132.17 User+Sys Time Running Test (seconds) 164.2 160.13 Total Elapsed Time (seconds) 123.46 120.87 The overall improvement is small but the System CPU time is much improved and roughly in correlation to what oprofile reported (these performance figures are without profiling so skew is expected). The actual number of page faults is noticeably improved. For benchmarks like kernel builds, the overall benefit is marginal but the system CPU time is slightly reduced. To test the actual bug the commit fixed I opened two terminals. The first ran within a cpuset and continually ran a small program that faulted 100M of anonymous data. In a second window, the nodemask of the cpuset was continually randomised in a loop. Without the commit, the program would fail every so often (usually within 10 seconds) and obviously with the commit everything worked fine. With this patch applied, it also worked fine so the fix should be functionally equivalent. Signed-off-by: Mel Gorman <mgorman@suse.de> Cc: Miao Xie <miaox@cn.fujitsu.com> Cc: David Rientjes <rientjes@google.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Christoph Lameter <cl@linux.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-03-22 07:34:11 +08:00
n->nr_partial > s->min_partial) {
slub: consider pfmemalloc_match() in get_partial_node() get_partial() is currently not checking pfmemalloc_match() meaning that it is possible for pfmemalloc pages to leak to non-pfmemalloc users. This is a problem in the following situation. Assume that there is a request from normal allocation and there are no objects in the per-cpu cache and no node-partial slab. In this case, slab_alloc enters the slow path and new_slab_objects() is called which may return a PFMEMALLOC page. As the current user is not allowed to access PFMEMALLOC page, deactivate_slab() is called ([5091b74a: mm: slub: optimise the SLUB fast path to avoid pfmemalloc checks]) and returns an object from PFMEMALLOC page. Next time, when we get another request from normal allocation, slab_alloc() enters the slow-path and calls new_slab_objects(). In new_slab_objects(), we call get_partial() and get a partial slab which was just deactivated but is a pfmemalloc page. We extract one object from it and re-deactivate. "deactivate -> re-get in get_partial -> re-deactivate" occures repeatedly. As a result, access to PFMEMALLOC page is not properly restricted and it can cause a performance degradation due to frequent deactivation. deactivation frequently. This patch changes get_partial_node() to take pfmemalloc_match() into account and prevents the "deactivate -> re-get in get_partial() scenario. Instead, new_slab() is called. Signed-off-by: Joonsoo Kim <js1304@gmail.com> Acked-by: David Rientjes <rientjes@google.com> Signed-off-by: Mel Gorman <mgorman@suse.de> Cc: David Miller <davem@davemloft.net> Cc: Chuck Lever <chuck.lever@oracle.com> Cc: Pekka Enberg <penberg@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-09-18 05:09:09 +08:00
object = get_partial_node(s, n, c, flags);
cpuset: mm: reduce large amounts of memory barrier related damage v3 Commit c0ff7453bb5c ("cpuset,mm: fix no node to alloc memory when changing cpuset's mems") wins a super prize for the largest number of memory barriers entered into fast paths for one commit. [get|put]_mems_allowed is incredibly heavy with pairs of full memory barriers inserted into a number of hot paths. This was detected while investigating at large page allocator slowdown introduced some time after 2.6.32. The largest portion of this overhead was shown by oprofile to be at an mfence introduced by this commit into the page allocator hot path. For extra style points, the commit introduced the use of yield() in an implementation of what looks like a spinning mutex. This patch replaces the full memory barriers on both read and write sides with a sequence counter with just read barriers on the fast path side. This is much cheaper on some architectures, including x86. The main bulk of the patch is the retry logic if the nodemask changes in a manner that can cause a false failure. While updating the nodemask, a check is made to see if a false failure is a risk. If it is, the sequence number gets bumped and parallel allocators will briefly stall while the nodemask update takes place. In a page fault test microbenchmark, oprofile samples from __alloc_pages_nodemask went from 4.53% of all samples to 1.15%. The actual results were 3.3.0-rc3 3.3.0-rc3 rc3-vanilla nobarrier-v2r1 Clients 1 UserTime 0.07 ( 0.00%) 0.08 (-14.19%) Clients 2 UserTime 0.07 ( 0.00%) 0.07 ( 2.72%) Clients 4 UserTime 0.08 ( 0.00%) 0.07 ( 3.29%) Clients 1 SysTime 0.70 ( 0.00%) 0.65 ( 6.65%) Clients 2 SysTime 0.85 ( 0.00%) 0.82 ( 3.65%) Clients 4 SysTime 1.41 ( 0.00%) 1.41 ( 0.32%) Clients 1 WallTime 0.77 ( 0.00%) 0.74 ( 4.19%) Clients 2 WallTime 0.47 ( 0.00%) 0.45 ( 3.73%) Clients 4 WallTime 0.38 ( 0.00%) 0.37 ( 1.58%) Clients 1 Flt/sec/cpu 497620.28 ( 0.00%) 520294.53 ( 4.56%) Clients 2 Flt/sec/cpu 414639.05 ( 0.00%) 429882.01 ( 3.68%) Clients 4 Flt/sec/cpu 257959.16 ( 0.00%) 258761.48 ( 0.31%) Clients 1 Flt/sec 495161.39 ( 0.00%) 517292.87 ( 4.47%) Clients 2 Flt/sec 820325.95 ( 0.00%) 850289.77 ( 3.65%) Clients 4 Flt/sec 1020068.93 ( 0.00%) 1022674.06 ( 0.26%) MMTests Statistics: duration Sys Time Running Test (seconds) 135.68 132.17 User+Sys Time Running Test (seconds) 164.2 160.13 Total Elapsed Time (seconds) 123.46 120.87 The overall improvement is small but the System CPU time is much improved and roughly in correlation to what oprofile reported (these performance figures are without profiling so skew is expected). The actual number of page faults is noticeably improved. For benchmarks like kernel builds, the overall benefit is marginal but the system CPU time is slightly reduced. To test the actual bug the commit fixed I opened two terminals. The first ran within a cpuset and continually ran a small program that faulted 100M of anonymous data. In a second window, the nodemask of the cpuset was continually randomised in a loop. Without the commit, the program would fail every so often (usually within 10 seconds) and obviously with the commit everything worked fine. With this patch applied, it also worked fine so the fix should be functionally equivalent. Signed-off-by: Mel Gorman <mgorman@suse.de> Cc: Miao Xie <miaox@cn.fujitsu.com> Cc: David Rientjes <rientjes@google.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Christoph Lameter <cl@linux.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-03-22 07:34:11 +08:00
if (object) {
/*
* Don't check read_mems_allowed_retry()
* here - if mems_allowed was updated in
* parallel, that was a harmless race
* between allocation and the cpuset
* update
cpuset: mm: reduce large amounts of memory barrier related damage v3 Commit c0ff7453bb5c ("cpuset,mm: fix no node to alloc memory when changing cpuset's mems") wins a super prize for the largest number of memory barriers entered into fast paths for one commit. [get|put]_mems_allowed is incredibly heavy with pairs of full memory barriers inserted into a number of hot paths. This was detected while investigating at large page allocator slowdown introduced some time after 2.6.32. The largest portion of this overhead was shown by oprofile to be at an mfence introduced by this commit into the page allocator hot path. For extra style points, the commit introduced the use of yield() in an implementation of what looks like a spinning mutex. This patch replaces the full memory barriers on both read and write sides with a sequence counter with just read barriers on the fast path side. This is much cheaper on some architectures, including x86. The main bulk of the patch is the retry logic if the nodemask changes in a manner that can cause a false failure. While updating the nodemask, a check is made to see if a false failure is a risk. If it is, the sequence number gets bumped and parallel allocators will briefly stall while the nodemask update takes place. In a page fault test microbenchmark, oprofile samples from __alloc_pages_nodemask went from 4.53% of all samples to 1.15%. The actual results were 3.3.0-rc3 3.3.0-rc3 rc3-vanilla nobarrier-v2r1 Clients 1 UserTime 0.07 ( 0.00%) 0.08 (-14.19%) Clients 2 UserTime 0.07 ( 0.00%) 0.07 ( 2.72%) Clients 4 UserTime 0.08 ( 0.00%) 0.07 ( 3.29%) Clients 1 SysTime 0.70 ( 0.00%) 0.65 ( 6.65%) Clients 2 SysTime 0.85 ( 0.00%) 0.82 ( 3.65%) Clients 4 SysTime 1.41 ( 0.00%) 1.41 ( 0.32%) Clients 1 WallTime 0.77 ( 0.00%) 0.74 ( 4.19%) Clients 2 WallTime 0.47 ( 0.00%) 0.45 ( 3.73%) Clients 4 WallTime 0.38 ( 0.00%) 0.37 ( 1.58%) Clients 1 Flt/sec/cpu 497620.28 ( 0.00%) 520294.53 ( 4.56%) Clients 2 Flt/sec/cpu 414639.05 ( 0.00%) 429882.01 ( 3.68%) Clients 4 Flt/sec/cpu 257959.16 ( 0.00%) 258761.48 ( 0.31%) Clients 1 Flt/sec 495161.39 ( 0.00%) 517292.87 ( 4.47%) Clients 2 Flt/sec 820325.95 ( 0.00%) 850289.77 ( 3.65%) Clients 4 Flt/sec 1020068.93 ( 0.00%) 1022674.06 ( 0.26%) MMTests Statistics: duration Sys Time Running Test (seconds) 135.68 132.17 User+Sys Time Running Test (seconds) 164.2 160.13 Total Elapsed Time (seconds) 123.46 120.87 The overall improvement is small but the System CPU time is much improved and roughly in correlation to what oprofile reported (these performance figures are without profiling so skew is expected). The actual number of page faults is noticeably improved. For benchmarks like kernel builds, the overall benefit is marginal but the system CPU time is slightly reduced. To test the actual bug the commit fixed I opened two terminals. The first ran within a cpuset and continually ran a small program that faulted 100M of anonymous data. In a second window, the nodemask of the cpuset was continually randomised in a loop. Without the commit, the program would fail every so often (usually within 10 seconds) and obviously with the commit everything worked fine. With this patch applied, it also worked fine so the fix should be functionally equivalent. Signed-off-by: Mel Gorman <mgorman@suse.de> Cc: Miao Xie <miaox@cn.fujitsu.com> Cc: David Rientjes <rientjes@google.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Christoph Lameter <cl@linux.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-03-22 07:34:11 +08:00
*/
return object;
}
cpuset,mm: fix no node to alloc memory when changing cpuset's mems Before applying this patch, cpuset updates task->mems_allowed and mempolicy by setting all new bits in the nodemask first, and clearing all old unallowed bits later. But in the way, the allocator may find that there is no node to alloc memory. The reason is that cpuset rebinds the task's mempolicy, it cleans the nodes which the allocater can alloc pages on, for example: (mpol: mempolicy) task1 task1's mpol task2 alloc page 1 alloc on node0? NO 1 1 change mems from 1 to 0 1 rebind task1's mpol 0-1 set new bits 0 clear disallowed bits alloc on node1? NO 0 ... can't alloc page goto oom This patch fixes this problem by expanding the nodes range first(set newly allowed bits) and shrink it lazily(clear newly disallowed bits). So we use a variable to tell the write-side task that read-side task is reading nodemask, and the write-side task clears newly disallowed nodes after read-side task ends the current memory allocation. [akpm@linux-foundation.org: fix spello] Signed-off-by: Miao Xie <miaox@cn.fujitsu.com> Cc: David Rientjes <rientjes@google.com> Cc: Nick Piggin <npiggin@suse.de> Cc: Paul Menage <menage@google.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: Ravikiran Thirumalai <kiran@scalex86.org> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Christoph Lameter <cl@linux-foundation.org> Cc: Andi Kleen <andi@firstfloor.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-25 05:32:08 +08:00
}
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
} while (read_mems_allowed_retry(cpuset_mems_cookie));
#endif /* CONFIG_NUMA */
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
return NULL;
}
/*
* Get a partial page, lock it and return it.
*/
static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
struct kmem_cache_cpu *c)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
{
void *object;
int searchnode = node;
if (node == NUMA_NO_NODE)
searchnode = numa_mem_id();
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
slub: consider pfmemalloc_match() in get_partial_node() get_partial() is currently not checking pfmemalloc_match() meaning that it is possible for pfmemalloc pages to leak to non-pfmemalloc users. This is a problem in the following situation. Assume that there is a request from normal allocation and there are no objects in the per-cpu cache and no node-partial slab. In this case, slab_alloc enters the slow path and new_slab_objects() is called which may return a PFMEMALLOC page. As the current user is not allowed to access PFMEMALLOC page, deactivate_slab() is called ([5091b74a: mm: slub: optimise the SLUB fast path to avoid pfmemalloc checks]) and returns an object from PFMEMALLOC page. Next time, when we get another request from normal allocation, slab_alloc() enters the slow-path and calls new_slab_objects(). In new_slab_objects(), we call get_partial() and get a partial slab which was just deactivated but is a pfmemalloc page. We extract one object from it and re-deactivate. "deactivate -> re-get in get_partial -> re-deactivate" occures repeatedly. As a result, access to PFMEMALLOC page is not properly restricted and it can cause a performance degradation due to frequent deactivation. deactivation frequently. This patch changes get_partial_node() to take pfmemalloc_match() into account and prevents the "deactivate -> re-get in get_partial() scenario. Instead, new_slab() is called. Signed-off-by: Joonsoo Kim <js1304@gmail.com> Acked-by: David Rientjes <rientjes@google.com> Signed-off-by: Mel Gorman <mgorman@suse.de> Cc: David Miller <davem@davemloft.net> Cc: Chuck Lever <chuck.lever@oracle.com> Cc: Pekka Enberg <penberg@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-09-18 05:09:09 +08:00
object = get_partial_node(s, get_node(s, searchnode), c, flags);
if (object || node != NUMA_NO_NODE)
return object;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
return get_any_partial(s, flags, c);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
#ifdef CONFIG_PREEMPTION
Lockless (and preemptless) fastpaths for slub Use the this_cpu_cmpxchg_double functionality to implement a lockless allocation algorithm on arches that support fast this_cpu_ops. Each of the per cpu pointers is paired with a transaction id that ensures that updates of the per cpu information can only occur in sequence on a certain cpu. A transaction id is a "long" integer that is comprised of an event number and the cpu number. The event number is incremented for every change to the per cpu state. This means that the cmpxchg instruction can verify for an update that nothing interfered and that we are updating the percpu structure for the processor where we picked up the information and that we are also currently on that processor when we update the information. This results in a significant decrease of the overhead in the fastpaths. It also makes it easy to adopt the fast path for realtime kernels since this is lockless and does not require the use of the current per cpu area over the critical section. It is only important that the per cpu area is current at the beginning of the critical section and at the end. So there is no need even to disable preemption. Test results show that the fastpath cycle count is reduced by up to ~ 40% (alloc/free test goes from ~140 cycles down to ~80). The slowpath for kfree adds a few cycles. Sadly this does nothing for the slowpath which is where the main issues with performance in slub are but the best case performance rises significantly. (For that see the more complex slub patches that require cmpxchg_double) Kmalloc: alloc/free test Before: 10000 times kmalloc(8)/kfree -> 134 cycles 10000 times kmalloc(16)/kfree -> 152 cycles 10000 times kmalloc(32)/kfree -> 144 cycles 10000 times kmalloc(64)/kfree -> 142 cycles 10000 times kmalloc(128)/kfree -> 142 cycles 10000 times kmalloc(256)/kfree -> 132 cycles 10000 times kmalloc(512)/kfree -> 132 cycles 10000 times kmalloc(1024)/kfree -> 135 cycles 10000 times kmalloc(2048)/kfree -> 135 cycles 10000 times kmalloc(4096)/kfree -> 135 cycles 10000 times kmalloc(8192)/kfree -> 144 cycles 10000 times kmalloc(16384)/kfree -> 754 cycles After: 10000 times kmalloc(8)/kfree -> 78 cycles 10000 times kmalloc(16)/kfree -> 78 cycles 10000 times kmalloc(32)/kfree -> 82 cycles 10000 times kmalloc(64)/kfree -> 88 cycles 10000 times kmalloc(128)/kfree -> 79 cycles 10000 times kmalloc(256)/kfree -> 79 cycles 10000 times kmalloc(512)/kfree -> 85 cycles 10000 times kmalloc(1024)/kfree -> 82 cycles 10000 times kmalloc(2048)/kfree -> 82 cycles 10000 times kmalloc(4096)/kfree -> 85 cycles 10000 times kmalloc(8192)/kfree -> 82 cycles 10000 times kmalloc(16384)/kfree -> 706 cycles Kmalloc: Repeatedly allocate then free test Before: 10000 times kmalloc(8) -> 211 cycles kfree -> 113 cycles 10000 times kmalloc(16) -> 174 cycles kfree -> 115 cycles 10000 times kmalloc(32) -> 235 cycles kfree -> 129 cycles 10000 times kmalloc(64) -> 222 cycles kfree -> 120 cycles 10000 times kmalloc(128) -> 343 cycles kfree -> 139 cycles 10000 times kmalloc(256) -> 827 cycles kfree -> 147 cycles 10000 times kmalloc(512) -> 1048 cycles kfree -> 272 cycles 10000 times kmalloc(1024) -> 2043 cycles kfree -> 528 cycles 10000 times kmalloc(2048) -> 4002 cycles kfree -> 571 cycles 10000 times kmalloc(4096) -> 7740 cycles kfree -> 628 cycles 10000 times kmalloc(8192) -> 8062 cycles kfree -> 850 cycles 10000 times kmalloc(16384) -> 8895 cycles kfree -> 1249 cycles After: 10000 times kmalloc(8) -> 190 cycles kfree -> 129 cycles 10000 times kmalloc(16) -> 76 cycles kfree -> 123 cycles 10000 times kmalloc(32) -> 126 cycles kfree -> 124 cycles 10000 times kmalloc(64) -> 181 cycles kfree -> 128 cycles 10000 times kmalloc(128) -> 310 cycles kfree -> 140 cycles 10000 times kmalloc(256) -> 809 cycles kfree -> 165 cycles 10000 times kmalloc(512) -> 1005 cycles kfree -> 269 cycles 10000 times kmalloc(1024) -> 1999 cycles kfree -> 527 cycles 10000 times kmalloc(2048) -> 3967 cycles kfree -> 570 cycles 10000 times kmalloc(4096) -> 7658 cycles kfree -> 637 cycles 10000 times kmalloc(8192) -> 8111 cycles kfree -> 859 cycles 10000 times kmalloc(16384) -> 8791 cycles kfree -> 1173 cycles Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-02-26 01:38:54 +08:00
/*
* Calculate the next globally unique transaction for disambiguation
Lockless (and preemptless) fastpaths for slub Use the this_cpu_cmpxchg_double functionality to implement a lockless allocation algorithm on arches that support fast this_cpu_ops. Each of the per cpu pointers is paired with a transaction id that ensures that updates of the per cpu information can only occur in sequence on a certain cpu. A transaction id is a "long" integer that is comprised of an event number and the cpu number. The event number is incremented for every change to the per cpu state. This means that the cmpxchg instruction can verify for an update that nothing interfered and that we are updating the percpu structure for the processor where we picked up the information and that we are also currently on that processor when we update the information. This results in a significant decrease of the overhead in the fastpaths. It also makes it easy to adopt the fast path for realtime kernels since this is lockless and does not require the use of the current per cpu area over the critical section. It is only important that the per cpu area is current at the beginning of the critical section and at the end. So there is no need even to disable preemption. Test results show that the fastpath cycle count is reduced by up to ~ 40% (alloc/free test goes from ~140 cycles down to ~80). The slowpath for kfree adds a few cycles. Sadly this does nothing for the slowpath which is where the main issues with performance in slub are but the best case performance rises significantly. (For that see the more complex slub patches that require cmpxchg_double) Kmalloc: alloc/free test Before: 10000 times kmalloc(8)/kfree -> 134 cycles 10000 times kmalloc(16)/kfree -> 152 cycles 10000 times kmalloc(32)/kfree -> 144 cycles 10000 times kmalloc(64)/kfree -> 142 cycles 10000 times kmalloc(128)/kfree -> 142 cycles 10000 times kmalloc(256)/kfree -> 132 cycles 10000 times kmalloc(512)/kfree -> 132 cycles 10000 times kmalloc(1024)/kfree -> 135 cycles 10000 times kmalloc(2048)/kfree -> 135 cycles 10000 times kmalloc(4096)/kfree -> 135 cycles 10000 times kmalloc(8192)/kfree -> 144 cycles 10000 times kmalloc(16384)/kfree -> 754 cycles After: 10000 times kmalloc(8)/kfree -> 78 cycles 10000 times kmalloc(16)/kfree -> 78 cycles 10000 times kmalloc(32)/kfree -> 82 cycles 10000 times kmalloc(64)/kfree -> 88 cycles 10000 times kmalloc(128)/kfree -> 79 cycles 10000 times kmalloc(256)/kfree -> 79 cycles 10000 times kmalloc(512)/kfree -> 85 cycles 10000 times kmalloc(1024)/kfree -> 82 cycles 10000 times kmalloc(2048)/kfree -> 82 cycles 10000 times kmalloc(4096)/kfree -> 85 cycles 10000 times kmalloc(8192)/kfree -> 82 cycles 10000 times kmalloc(16384)/kfree -> 706 cycles Kmalloc: Repeatedly allocate then free test Before: 10000 times kmalloc(8) -> 211 cycles kfree -> 113 cycles 10000 times kmalloc(16) -> 174 cycles kfree -> 115 cycles 10000 times kmalloc(32) -> 235 cycles kfree -> 129 cycles 10000 times kmalloc(64) -> 222 cycles kfree -> 120 cycles 10000 times kmalloc(128) -> 343 cycles kfree -> 139 cycles 10000 times kmalloc(256) -> 827 cycles kfree -> 147 cycles 10000 times kmalloc(512) -> 1048 cycles kfree -> 272 cycles 10000 times kmalloc(1024) -> 2043 cycles kfree -> 528 cycles 10000 times kmalloc(2048) -> 4002 cycles kfree -> 571 cycles 10000 times kmalloc(4096) -> 7740 cycles kfree -> 628 cycles 10000 times kmalloc(8192) -> 8062 cycles kfree -> 850 cycles 10000 times kmalloc(16384) -> 8895 cycles kfree -> 1249 cycles After: 10000 times kmalloc(8) -> 190 cycles kfree -> 129 cycles 10000 times kmalloc(16) -> 76 cycles kfree -> 123 cycles 10000 times kmalloc(32) -> 126 cycles kfree -> 124 cycles 10000 times kmalloc(64) -> 181 cycles kfree -> 128 cycles 10000 times kmalloc(128) -> 310 cycles kfree -> 140 cycles 10000 times kmalloc(256) -> 809 cycles kfree -> 165 cycles 10000 times kmalloc(512) -> 1005 cycles kfree -> 269 cycles 10000 times kmalloc(1024) -> 1999 cycles kfree -> 527 cycles 10000 times kmalloc(2048) -> 3967 cycles kfree -> 570 cycles 10000 times kmalloc(4096) -> 7658 cycles kfree -> 637 cycles 10000 times kmalloc(8192) -> 8111 cycles kfree -> 859 cycles 10000 times kmalloc(16384) -> 8791 cycles kfree -> 1173 cycles Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-02-26 01:38:54 +08:00
* during cmpxchg. The transactions start with the cpu number and are then
* incremented by CONFIG_NR_CPUS.
*/
#define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
#else
/*
* No preemption supported therefore also no need to check for
* different cpus.
*/
#define TID_STEP 1
#endif
static inline unsigned long next_tid(unsigned long tid)
{
return tid + TID_STEP;
}
#ifdef SLUB_DEBUG_CMPXCHG
Lockless (and preemptless) fastpaths for slub Use the this_cpu_cmpxchg_double functionality to implement a lockless allocation algorithm on arches that support fast this_cpu_ops. Each of the per cpu pointers is paired with a transaction id that ensures that updates of the per cpu information can only occur in sequence on a certain cpu. A transaction id is a "long" integer that is comprised of an event number and the cpu number. The event number is incremented for every change to the per cpu state. This means that the cmpxchg instruction can verify for an update that nothing interfered and that we are updating the percpu structure for the processor where we picked up the information and that we are also currently on that processor when we update the information. This results in a significant decrease of the overhead in the fastpaths. It also makes it easy to adopt the fast path for realtime kernels since this is lockless and does not require the use of the current per cpu area over the critical section. It is only important that the per cpu area is current at the beginning of the critical section and at the end. So there is no need even to disable preemption. Test results show that the fastpath cycle count is reduced by up to ~ 40% (alloc/free test goes from ~140 cycles down to ~80). The slowpath for kfree adds a few cycles. Sadly this does nothing for the slowpath which is where the main issues with performance in slub are but the best case performance rises significantly. (For that see the more complex slub patches that require cmpxchg_double) Kmalloc: alloc/free test Before: 10000 times kmalloc(8)/kfree -> 134 cycles 10000 times kmalloc(16)/kfree -> 152 cycles 10000 times kmalloc(32)/kfree -> 144 cycles 10000 times kmalloc(64)/kfree -> 142 cycles 10000 times kmalloc(128)/kfree -> 142 cycles 10000 times kmalloc(256)/kfree -> 132 cycles 10000 times kmalloc(512)/kfree -> 132 cycles 10000 times kmalloc(1024)/kfree -> 135 cycles 10000 times kmalloc(2048)/kfree -> 135 cycles 10000 times kmalloc(4096)/kfree -> 135 cycles 10000 times kmalloc(8192)/kfree -> 144 cycles 10000 times kmalloc(16384)/kfree -> 754 cycles After: 10000 times kmalloc(8)/kfree -> 78 cycles 10000 times kmalloc(16)/kfree -> 78 cycles 10000 times kmalloc(32)/kfree -> 82 cycles 10000 times kmalloc(64)/kfree -> 88 cycles 10000 times kmalloc(128)/kfree -> 79 cycles 10000 times kmalloc(256)/kfree -> 79 cycles 10000 times kmalloc(512)/kfree -> 85 cycles 10000 times kmalloc(1024)/kfree -> 82 cycles 10000 times kmalloc(2048)/kfree -> 82 cycles 10000 times kmalloc(4096)/kfree -> 85 cycles 10000 times kmalloc(8192)/kfree -> 82 cycles 10000 times kmalloc(16384)/kfree -> 706 cycles Kmalloc: Repeatedly allocate then free test Before: 10000 times kmalloc(8) -> 211 cycles kfree -> 113 cycles 10000 times kmalloc(16) -> 174 cycles kfree -> 115 cycles 10000 times kmalloc(32) -> 235 cycles kfree -> 129 cycles 10000 times kmalloc(64) -> 222 cycles kfree -> 120 cycles 10000 times kmalloc(128) -> 343 cycles kfree -> 139 cycles 10000 times kmalloc(256) -> 827 cycles kfree -> 147 cycles 10000 times kmalloc(512) -> 1048 cycles kfree -> 272 cycles 10000 times kmalloc(1024) -> 2043 cycles kfree -> 528 cycles 10000 times kmalloc(2048) -> 4002 cycles kfree -> 571 cycles 10000 times kmalloc(4096) -> 7740 cycles kfree -> 628 cycles 10000 times kmalloc(8192) -> 8062 cycles kfree -> 850 cycles 10000 times kmalloc(16384) -> 8895 cycles kfree -> 1249 cycles After: 10000 times kmalloc(8) -> 190 cycles kfree -> 129 cycles 10000 times kmalloc(16) -> 76 cycles kfree -> 123 cycles 10000 times kmalloc(32) -> 126 cycles kfree -> 124 cycles 10000 times kmalloc(64) -> 181 cycles kfree -> 128 cycles 10000 times kmalloc(128) -> 310 cycles kfree -> 140 cycles 10000 times kmalloc(256) -> 809 cycles kfree -> 165 cycles 10000 times kmalloc(512) -> 1005 cycles kfree -> 269 cycles 10000 times kmalloc(1024) -> 1999 cycles kfree -> 527 cycles 10000 times kmalloc(2048) -> 3967 cycles kfree -> 570 cycles 10000 times kmalloc(4096) -> 7658 cycles kfree -> 637 cycles 10000 times kmalloc(8192) -> 8111 cycles kfree -> 859 cycles 10000 times kmalloc(16384) -> 8791 cycles kfree -> 1173 cycles Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-02-26 01:38:54 +08:00
static inline unsigned int tid_to_cpu(unsigned long tid)
{
return tid % TID_STEP;
}
static inline unsigned long tid_to_event(unsigned long tid)
{
return tid / TID_STEP;
}
#endif
Lockless (and preemptless) fastpaths for slub Use the this_cpu_cmpxchg_double functionality to implement a lockless allocation algorithm on arches that support fast this_cpu_ops. Each of the per cpu pointers is paired with a transaction id that ensures that updates of the per cpu information can only occur in sequence on a certain cpu. A transaction id is a "long" integer that is comprised of an event number and the cpu number. The event number is incremented for every change to the per cpu state. This means that the cmpxchg instruction can verify for an update that nothing interfered and that we are updating the percpu structure for the processor where we picked up the information and that we are also currently on that processor when we update the information. This results in a significant decrease of the overhead in the fastpaths. It also makes it easy to adopt the fast path for realtime kernels since this is lockless and does not require the use of the current per cpu area over the critical section. It is only important that the per cpu area is current at the beginning of the critical section and at the end. So there is no need even to disable preemption. Test results show that the fastpath cycle count is reduced by up to ~ 40% (alloc/free test goes from ~140 cycles down to ~80). The slowpath for kfree adds a few cycles. Sadly this does nothing for the slowpath which is where the main issues with performance in slub are but the best case performance rises significantly. (For that see the more complex slub patches that require cmpxchg_double) Kmalloc: alloc/free test Before: 10000 times kmalloc(8)/kfree -> 134 cycles 10000 times kmalloc(16)/kfree -> 152 cycles 10000 times kmalloc(32)/kfree -> 144 cycles 10000 times kmalloc(64)/kfree -> 142 cycles 10000 times kmalloc(128)/kfree -> 142 cycles 10000 times kmalloc(256)/kfree -> 132 cycles 10000 times kmalloc(512)/kfree -> 132 cycles 10000 times kmalloc(1024)/kfree -> 135 cycles 10000 times kmalloc(2048)/kfree -> 135 cycles 10000 times kmalloc(4096)/kfree -> 135 cycles 10000 times kmalloc(8192)/kfree -> 144 cycles 10000 times kmalloc(16384)/kfree -> 754 cycles After: 10000 times kmalloc(8)/kfree -> 78 cycles 10000 times kmalloc(16)/kfree -> 78 cycles 10000 times kmalloc(32)/kfree -> 82 cycles 10000 times kmalloc(64)/kfree -> 88 cycles 10000 times kmalloc(128)/kfree -> 79 cycles 10000 times kmalloc(256)/kfree -> 79 cycles 10000 times kmalloc(512)/kfree -> 85 cycles 10000 times kmalloc(1024)/kfree -> 82 cycles 10000 times kmalloc(2048)/kfree -> 82 cycles 10000 times kmalloc(4096)/kfree -> 85 cycles 10000 times kmalloc(8192)/kfree -> 82 cycles 10000 times kmalloc(16384)/kfree -> 706 cycles Kmalloc: Repeatedly allocate then free test Before: 10000 times kmalloc(8) -> 211 cycles kfree -> 113 cycles 10000 times kmalloc(16) -> 174 cycles kfree -> 115 cycles 10000 times kmalloc(32) -> 235 cycles kfree -> 129 cycles 10000 times kmalloc(64) -> 222 cycles kfree -> 120 cycles 10000 times kmalloc(128) -> 343 cycles kfree -> 139 cycles 10000 times kmalloc(256) -> 827 cycles kfree -> 147 cycles 10000 times kmalloc(512) -> 1048 cycles kfree -> 272 cycles 10000 times kmalloc(1024) -> 2043 cycles kfree -> 528 cycles 10000 times kmalloc(2048) -> 4002 cycles kfree -> 571 cycles 10000 times kmalloc(4096) -> 7740 cycles kfree -> 628 cycles 10000 times kmalloc(8192) -> 8062 cycles kfree -> 850 cycles 10000 times kmalloc(16384) -> 8895 cycles kfree -> 1249 cycles After: 10000 times kmalloc(8) -> 190 cycles kfree -> 129 cycles 10000 times kmalloc(16) -> 76 cycles kfree -> 123 cycles 10000 times kmalloc(32) -> 126 cycles kfree -> 124 cycles 10000 times kmalloc(64) -> 181 cycles kfree -> 128 cycles 10000 times kmalloc(128) -> 310 cycles kfree -> 140 cycles 10000 times kmalloc(256) -> 809 cycles kfree -> 165 cycles 10000 times kmalloc(512) -> 1005 cycles kfree -> 269 cycles 10000 times kmalloc(1024) -> 1999 cycles kfree -> 527 cycles 10000 times kmalloc(2048) -> 3967 cycles kfree -> 570 cycles 10000 times kmalloc(4096) -> 7658 cycles kfree -> 637 cycles 10000 times kmalloc(8192) -> 8111 cycles kfree -> 859 cycles 10000 times kmalloc(16384) -> 8791 cycles kfree -> 1173 cycles Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-02-26 01:38:54 +08:00
static inline unsigned int init_tid(int cpu)
{
return cpu;
}
static inline void note_cmpxchg_failure(const char *n,
const struct kmem_cache *s, unsigned long tid)
{
#ifdef SLUB_DEBUG_CMPXCHG
unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
pr_info("%s %s: cmpxchg redo ", n, s->name);
Lockless (and preemptless) fastpaths for slub Use the this_cpu_cmpxchg_double functionality to implement a lockless allocation algorithm on arches that support fast this_cpu_ops. Each of the per cpu pointers is paired with a transaction id that ensures that updates of the per cpu information can only occur in sequence on a certain cpu. A transaction id is a "long" integer that is comprised of an event number and the cpu number. The event number is incremented for every change to the per cpu state. This means that the cmpxchg instruction can verify for an update that nothing interfered and that we are updating the percpu structure for the processor where we picked up the information and that we are also currently on that processor when we update the information. This results in a significant decrease of the overhead in the fastpaths. It also makes it easy to adopt the fast path for realtime kernels since this is lockless and does not require the use of the current per cpu area over the critical section. It is only important that the per cpu area is current at the beginning of the critical section and at the end. So there is no need even to disable preemption. Test results show that the fastpath cycle count is reduced by up to ~ 40% (alloc/free test goes from ~140 cycles down to ~80). The slowpath for kfree adds a few cycles. Sadly this does nothing for the slowpath which is where the main issues with performance in slub are but the best case performance rises significantly. (For that see the more complex slub patches that require cmpxchg_double) Kmalloc: alloc/free test Before: 10000 times kmalloc(8)/kfree -> 134 cycles 10000 times kmalloc(16)/kfree -> 152 cycles 10000 times kmalloc(32)/kfree -> 144 cycles 10000 times kmalloc(64)/kfree -> 142 cycles 10000 times kmalloc(128)/kfree -> 142 cycles 10000 times kmalloc(256)/kfree -> 132 cycles 10000 times kmalloc(512)/kfree -> 132 cycles 10000 times kmalloc(1024)/kfree -> 135 cycles 10000 times kmalloc(2048)/kfree -> 135 cycles 10000 times kmalloc(4096)/kfree -> 135 cycles 10000 times kmalloc(8192)/kfree -> 144 cycles 10000 times kmalloc(16384)/kfree -> 754 cycles After: 10000 times kmalloc(8)/kfree -> 78 cycles 10000 times kmalloc(16)/kfree -> 78 cycles 10000 times kmalloc(32)/kfree -> 82 cycles 10000 times kmalloc(64)/kfree -> 88 cycles 10000 times kmalloc(128)/kfree -> 79 cycles 10000 times kmalloc(256)/kfree -> 79 cycles 10000 times kmalloc(512)/kfree -> 85 cycles 10000 times kmalloc(1024)/kfree -> 82 cycles 10000 times kmalloc(2048)/kfree -> 82 cycles 10000 times kmalloc(4096)/kfree -> 85 cycles 10000 times kmalloc(8192)/kfree -> 82 cycles 10000 times kmalloc(16384)/kfree -> 706 cycles Kmalloc: Repeatedly allocate then free test Before: 10000 times kmalloc(8) -> 211 cycles kfree -> 113 cycles 10000 times kmalloc(16) -> 174 cycles kfree -> 115 cycles 10000 times kmalloc(32) -> 235 cycles kfree -> 129 cycles 10000 times kmalloc(64) -> 222 cycles kfree -> 120 cycles 10000 times kmalloc(128) -> 343 cycles kfree -> 139 cycles 10000 times kmalloc(256) -> 827 cycles kfree -> 147 cycles 10000 times kmalloc(512) -> 1048 cycles kfree -> 272 cycles 10000 times kmalloc(1024) -> 2043 cycles kfree -> 528 cycles 10000 times kmalloc(2048) -> 4002 cycles kfree -> 571 cycles 10000 times kmalloc(4096) -> 7740 cycles kfree -> 628 cycles 10000 times kmalloc(8192) -> 8062 cycles kfree -> 850 cycles 10000 times kmalloc(16384) -> 8895 cycles kfree -> 1249 cycles After: 10000 times kmalloc(8) -> 190 cycles kfree -> 129 cycles 10000 times kmalloc(16) -> 76 cycles kfree -> 123 cycles 10000 times kmalloc(32) -> 126 cycles kfree -> 124 cycles 10000 times kmalloc(64) -> 181 cycles kfree -> 128 cycles 10000 times kmalloc(128) -> 310 cycles kfree -> 140 cycles 10000 times kmalloc(256) -> 809 cycles kfree -> 165 cycles 10000 times kmalloc(512) -> 1005 cycles kfree -> 269 cycles 10000 times kmalloc(1024) -> 1999 cycles kfree -> 527 cycles 10000 times kmalloc(2048) -> 3967 cycles kfree -> 570 cycles 10000 times kmalloc(4096) -> 7658 cycles kfree -> 637 cycles 10000 times kmalloc(8192) -> 8111 cycles kfree -> 859 cycles 10000 times kmalloc(16384) -> 8791 cycles kfree -> 1173 cycles Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-02-26 01:38:54 +08:00
#ifdef CONFIG_PREEMPTION
Lockless (and preemptless) fastpaths for slub Use the this_cpu_cmpxchg_double functionality to implement a lockless allocation algorithm on arches that support fast this_cpu_ops. Each of the per cpu pointers is paired with a transaction id that ensures that updates of the per cpu information can only occur in sequence on a certain cpu. A transaction id is a "long" integer that is comprised of an event number and the cpu number. The event number is incremented for every change to the per cpu state. This means that the cmpxchg instruction can verify for an update that nothing interfered and that we are updating the percpu structure for the processor where we picked up the information and that we are also currently on that processor when we update the information. This results in a significant decrease of the overhead in the fastpaths. It also makes it easy to adopt the fast path for realtime kernels since this is lockless and does not require the use of the current per cpu area over the critical section. It is only important that the per cpu area is current at the beginning of the critical section and at the end. So there is no need even to disable preemption. Test results show that the fastpath cycle count is reduced by up to ~ 40% (alloc/free test goes from ~140 cycles down to ~80). The slowpath for kfree adds a few cycles. Sadly this does nothing for the slowpath which is where the main issues with performance in slub are but the best case performance rises significantly. (For that see the more complex slub patches that require cmpxchg_double) Kmalloc: alloc/free test Before: 10000 times kmalloc(8)/kfree -> 134 cycles 10000 times kmalloc(16)/kfree -> 152 cycles 10000 times kmalloc(32)/kfree -> 144 cycles 10000 times kmalloc(64)/kfree -> 142 cycles 10000 times kmalloc(128)/kfree -> 142 cycles 10000 times kmalloc(256)/kfree -> 132 cycles 10000 times kmalloc(512)/kfree -> 132 cycles 10000 times kmalloc(1024)/kfree -> 135 cycles 10000 times kmalloc(2048)/kfree -> 135 cycles 10000 times kmalloc(4096)/kfree -> 135 cycles 10000 times kmalloc(8192)/kfree -> 144 cycles 10000 times kmalloc(16384)/kfree -> 754 cycles After: 10000 times kmalloc(8)/kfree -> 78 cycles 10000 times kmalloc(16)/kfree -> 78 cycles 10000 times kmalloc(32)/kfree -> 82 cycles 10000 times kmalloc(64)/kfree -> 88 cycles 10000 times kmalloc(128)/kfree -> 79 cycles 10000 times kmalloc(256)/kfree -> 79 cycles 10000 times kmalloc(512)/kfree -> 85 cycles 10000 times kmalloc(1024)/kfree -> 82 cycles 10000 times kmalloc(2048)/kfree -> 82 cycles 10000 times kmalloc(4096)/kfree -> 85 cycles 10000 times kmalloc(8192)/kfree -> 82 cycles 10000 times kmalloc(16384)/kfree -> 706 cycles Kmalloc: Repeatedly allocate then free test Before: 10000 times kmalloc(8) -> 211 cycles kfree -> 113 cycles 10000 times kmalloc(16) -> 174 cycles kfree -> 115 cycles 10000 times kmalloc(32) -> 235 cycles kfree -> 129 cycles 10000 times kmalloc(64) -> 222 cycles kfree -> 120 cycles 10000 times kmalloc(128) -> 343 cycles kfree -> 139 cycles 10000 times kmalloc(256) -> 827 cycles kfree -> 147 cycles 10000 times kmalloc(512) -> 1048 cycles kfree -> 272 cycles 10000 times kmalloc(1024) -> 2043 cycles kfree -> 528 cycles 10000 times kmalloc(2048) -> 4002 cycles kfree -> 571 cycles 10000 times kmalloc(4096) -> 7740 cycles kfree -> 628 cycles 10000 times kmalloc(8192) -> 8062 cycles kfree -> 850 cycles 10000 times kmalloc(16384) -> 8895 cycles kfree -> 1249 cycles After: 10000 times kmalloc(8) -> 190 cycles kfree -> 129 cycles 10000 times kmalloc(16) -> 76 cycles kfree -> 123 cycles 10000 times kmalloc(32) -> 126 cycles kfree -> 124 cycles 10000 times kmalloc(64) -> 181 cycles kfree -> 128 cycles 10000 times kmalloc(128) -> 310 cycles kfree -> 140 cycles 10000 times kmalloc(256) -> 809 cycles kfree -> 165 cycles 10000 times kmalloc(512) -> 1005 cycles kfree -> 269 cycles 10000 times kmalloc(1024) -> 1999 cycles kfree -> 527 cycles 10000 times kmalloc(2048) -> 3967 cycles kfree -> 570 cycles 10000 times kmalloc(4096) -> 7658 cycles kfree -> 637 cycles 10000 times kmalloc(8192) -> 8111 cycles kfree -> 859 cycles 10000 times kmalloc(16384) -> 8791 cycles kfree -> 1173 cycles Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-02-26 01:38:54 +08:00
if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
pr_warn("due to cpu change %d -> %d\n",
Lockless (and preemptless) fastpaths for slub Use the this_cpu_cmpxchg_double functionality to implement a lockless allocation algorithm on arches that support fast this_cpu_ops. Each of the per cpu pointers is paired with a transaction id that ensures that updates of the per cpu information can only occur in sequence on a certain cpu. A transaction id is a "long" integer that is comprised of an event number and the cpu number. The event number is incremented for every change to the per cpu state. This means that the cmpxchg instruction can verify for an update that nothing interfered and that we are updating the percpu structure for the processor where we picked up the information and that we are also currently on that processor when we update the information. This results in a significant decrease of the overhead in the fastpaths. It also makes it easy to adopt the fast path for realtime kernels since this is lockless and does not require the use of the current per cpu area over the critical section. It is only important that the per cpu area is current at the beginning of the critical section and at the end. So there is no need even to disable preemption. Test results show that the fastpath cycle count is reduced by up to ~ 40% (alloc/free test goes from ~140 cycles down to ~80). The slowpath for kfree adds a few cycles. Sadly this does nothing for the slowpath which is where the main issues with performance in slub are but the best case performance rises significantly. (For that see the more complex slub patches that require cmpxchg_double) Kmalloc: alloc/free test Before: 10000 times kmalloc(8)/kfree -> 134 cycles 10000 times kmalloc(16)/kfree -> 152 cycles 10000 times kmalloc(32)/kfree -> 144 cycles 10000 times kmalloc(64)/kfree -> 142 cycles 10000 times kmalloc(128)/kfree -> 142 cycles 10000 times kmalloc(256)/kfree -> 132 cycles 10000 times kmalloc(512)/kfree -> 132 cycles 10000 times kmalloc(1024)/kfree -> 135 cycles 10000 times kmalloc(2048)/kfree -> 135 cycles 10000 times kmalloc(4096)/kfree -> 135 cycles 10000 times kmalloc(8192)/kfree -> 144 cycles 10000 times kmalloc(16384)/kfree -> 754 cycles After: 10000 times kmalloc(8)/kfree -> 78 cycles 10000 times kmalloc(16)/kfree -> 78 cycles 10000 times kmalloc(32)/kfree -> 82 cycles 10000 times kmalloc(64)/kfree -> 88 cycles 10000 times kmalloc(128)/kfree -> 79 cycles 10000 times kmalloc(256)/kfree -> 79 cycles 10000 times kmalloc(512)/kfree -> 85 cycles 10000 times kmalloc(1024)/kfree -> 82 cycles 10000 times kmalloc(2048)/kfree -> 82 cycles 10000 times kmalloc(4096)/kfree -> 85 cycles 10000 times kmalloc(8192)/kfree -> 82 cycles 10000 times kmalloc(16384)/kfree -> 706 cycles Kmalloc: Repeatedly allocate then free test Before: 10000 times kmalloc(8) -> 211 cycles kfree -> 113 cycles 10000 times kmalloc(16) -> 174 cycles kfree -> 115 cycles 10000 times kmalloc(32) -> 235 cycles kfree -> 129 cycles 10000 times kmalloc(64) -> 222 cycles kfree -> 120 cycles 10000 times kmalloc(128) -> 343 cycles kfree -> 139 cycles 10000 times kmalloc(256) -> 827 cycles kfree -> 147 cycles 10000 times kmalloc(512) -> 1048 cycles kfree -> 272 cycles 10000 times kmalloc(1024) -> 2043 cycles kfree -> 528 cycles 10000 times kmalloc(2048) -> 4002 cycles kfree -> 571 cycles 10000 times kmalloc(4096) -> 7740 cycles kfree -> 628 cycles 10000 times kmalloc(8192) -> 8062 cycles kfree -> 850 cycles 10000 times kmalloc(16384) -> 8895 cycles kfree -> 1249 cycles After: 10000 times kmalloc(8) -> 190 cycles kfree -> 129 cycles 10000 times kmalloc(16) -> 76 cycles kfree -> 123 cycles 10000 times kmalloc(32) -> 126 cycles kfree -> 124 cycles 10000 times kmalloc(64) -> 181 cycles kfree -> 128 cycles 10000 times kmalloc(128) -> 310 cycles kfree -> 140 cycles 10000 times kmalloc(256) -> 809 cycles kfree -> 165 cycles 10000 times kmalloc(512) -> 1005 cycles kfree -> 269 cycles 10000 times kmalloc(1024) -> 1999 cycles kfree -> 527 cycles 10000 times kmalloc(2048) -> 3967 cycles kfree -> 570 cycles 10000 times kmalloc(4096) -> 7658 cycles kfree -> 637 cycles 10000 times kmalloc(8192) -> 8111 cycles kfree -> 859 cycles 10000 times kmalloc(16384) -> 8791 cycles kfree -> 1173 cycles Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-02-26 01:38:54 +08:00
tid_to_cpu(tid), tid_to_cpu(actual_tid));
else
#endif
if (tid_to_event(tid) != tid_to_event(actual_tid))
pr_warn("due to cpu running other code. Event %ld->%ld\n",
Lockless (and preemptless) fastpaths for slub Use the this_cpu_cmpxchg_double functionality to implement a lockless allocation algorithm on arches that support fast this_cpu_ops. Each of the per cpu pointers is paired with a transaction id that ensures that updates of the per cpu information can only occur in sequence on a certain cpu. A transaction id is a "long" integer that is comprised of an event number and the cpu number. The event number is incremented for every change to the per cpu state. This means that the cmpxchg instruction can verify for an update that nothing interfered and that we are updating the percpu structure for the processor where we picked up the information and that we are also currently on that processor when we update the information. This results in a significant decrease of the overhead in the fastpaths. It also makes it easy to adopt the fast path for realtime kernels since this is lockless and does not require the use of the current per cpu area over the critical section. It is only important that the per cpu area is current at the beginning of the critical section and at the end. So there is no need even to disable preemption. Test results show that the fastpath cycle count is reduced by up to ~ 40% (alloc/free test goes from ~140 cycles down to ~80). The slowpath for kfree adds a few cycles. Sadly this does nothing for the slowpath which is where the main issues with performance in slub are but the best case performance rises significantly. (For that see the more complex slub patches that require cmpxchg_double) Kmalloc: alloc/free test Before: 10000 times kmalloc(8)/kfree -> 134 cycles 10000 times kmalloc(16)/kfree -> 152 cycles 10000 times kmalloc(32)/kfree -> 144 cycles 10000 times kmalloc(64)/kfree -> 142 cycles 10000 times kmalloc(128)/kfree -> 142 cycles 10000 times kmalloc(256)/kfree -> 132 cycles 10000 times kmalloc(512)/kfree -> 132 cycles 10000 times kmalloc(1024)/kfree -> 135 cycles 10000 times kmalloc(2048)/kfree -> 135 cycles 10000 times kmalloc(4096)/kfree -> 135 cycles 10000 times kmalloc(8192)/kfree -> 144 cycles 10000 times kmalloc(16384)/kfree -> 754 cycles After: 10000 times kmalloc(8)/kfree -> 78 cycles 10000 times kmalloc(16)/kfree -> 78 cycles 10000 times kmalloc(32)/kfree -> 82 cycles 10000 times kmalloc(64)/kfree -> 88 cycles 10000 times kmalloc(128)/kfree -> 79 cycles 10000 times kmalloc(256)/kfree -> 79 cycles 10000 times kmalloc(512)/kfree -> 85 cycles 10000 times kmalloc(1024)/kfree -> 82 cycles 10000 times kmalloc(2048)/kfree -> 82 cycles 10000 times kmalloc(4096)/kfree -> 85 cycles 10000 times kmalloc(8192)/kfree -> 82 cycles 10000 times kmalloc(16384)/kfree -> 706 cycles Kmalloc: Repeatedly allocate then free test Before: 10000 times kmalloc(8) -> 211 cycles kfree -> 113 cycles 10000 times kmalloc(16) -> 174 cycles kfree -> 115 cycles 10000 times kmalloc(32) -> 235 cycles kfree -> 129 cycles 10000 times kmalloc(64) -> 222 cycles kfree -> 120 cycles 10000 times kmalloc(128) -> 343 cycles kfree -> 139 cycles 10000 times kmalloc(256) -> 827 cycles kfree -> 147 cycles 10000 times kmalloc(512) -> 1048 cycles kfree -> 272 cycles 10000 times kmalloc(1024) -> 2043 cycles kfree -> 528 cycles 10000 times kmalloc(2048) -> 4002 cycles kfree -> 571 cycles 10000 times kmalloc(4096) -> 7740 cycles kfree -> 628 cycles 10000 times kmalloc(8192) -> 8062 cycles kfree -> 850 cycles 10000 times kmalloc(16384) -> 8895 cycles kfree -> 1249 cycles After: 10000 times kmalloc(8) -> 190 cycles kfree -> 129 cycles 10000 times kmalloc(16) -> 76 cycles kfree -> 123 cycles 10000 times kmalloc(32) -> 126 cycles kfree -> 124 cycles 10000 times kmalloc(64) -> 181 cycles kfree -> 128 cycles 10000 times kmalloc(128) -> 310 cycles kfree -> 140 cycles 10000 times kmalloc(256) -> 809 cycles kfree -> 165 cycles 10000 times kmalloc(512) -> 1005 cycles kfree -> 269 cycles 10000 times kmalloc(1024) -> 1999 cycles kfree -> 527 cycles 10000 times kmalloc(2048) -> 3967 cycles kfree -> 570 cycles 10000 times kmalloc(4096) -> 7658 cycles kfree -> 637 cycles 10000 times kmalloc(8192) -> 8111 cycles kfree -> 859 cycles 10000 times kmalloc(16384) -> 8791 cycles kfree -> 1173 cycles Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-02-26 01:38:54 +08:00
tid_to_event(tid), tid_to_event(actual_tid));
else
pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
Lockless (and preemptless) fastpaths for slub Use the this_cpu_cmpxchg_double functionality to implement a lockless allocation algorithm on arches that support fast this_cpu_ops. Each of the per cpu pointers is paired with a transaction id that ensures that updates of the per cpu information can only occur in sequence on a certain cpu. A transaction id is a "long" integer that is comprised of an event number and the cpu number. The event number is incremented for every change to the per cpu state. This means that the cmpxchg instruction can verify for an update that nothing interfered and that we are updating the percpu structure for the processor where we picked up the information and that we are also currently on that processor when we update the information. This results in a significant decrease of the overhead in the fastpaths. It also makes it easy to adopt the fast path for realtime kernels since this is lockless and does not require the use of the current per cpu area over the critical section. It is only important that the per cpu area is current at the beginning of the critical section and at the end. So there is no need even to disable preemption. Test results show that the fastpath cycle count is reduced by up to ~ 40% (alloc/free test goes from ~140 cycles down to ~80). The slowpath for kfree adds a few cycles. Sadly this does nothing for the slowpath which is where the main issues with performance in slub are but the best case performance rises significantly. (For that see the more complex slub patches that require cmpxchg_double) Kmalloc: alloc/free test Before: 10000 times kmalloc(8)/kfree -> 134 cycles 10000 times kmalloc(16)/kfree -> 152 cycles 10000 times kmalloc(32)/kfree -> 144 cycles 10000 times kmalloc(64)/kfree -> 142 cycles 10000 times kmalloc(128)/kfree -> 142 cycles 10000 times kmalloc(256)/kfree -> 132 cycles 10000 times kmalloc(512)/kfree -> 132 cycles 10000 times kmalloc(1024)/kfree -> 135 cycles 10000 times kmalloc(2048)/kfree -> 135 cycles 10000 times kmalloc(4096)/kfree -> 135 cycles 10000 times kmalloc(8192)/kfree -> 144 cycles 10000 times kmalloc(16384)/kfree -> 754 cycles After: 10000 times kmalloc(8)/kfree -> 78 cycles 10000 times kmalloc(16)/kfree -> 78 cycles 10000 times kmalloc(32)/kfree -> 82 cycles 10000 times kmalloc(64)/kfree -> 88 cycles 10000 times kmalloc(128)/kfree -> 79 cycles 10000 times kmalloc(256)/kfree -> 79 cycles 10000 times kmalloc(512)/kfree -> 85 cycles 10000 times kmalloc(1024)/kfree -> 82 cycles 10000 times kmalloc(2048)/kfree -> 82 cycles 10000 times kmalloc(4096)/kfree -> 85 cycles 10000 times kmalloc(8192)/kfree -> 82 cycles 10000 times kmalloc(16384)/kfree -> 706 cycles Kmalloc: Repeatedly allocate then free test Before: 10000 times kmalloc(8) -> 211 cycles kfree -> 113 cycles 10000 times kmalloc(16) -> 174 cycles kfree -> 115 cycles 10000 times kmalloc(32) -> 235 cycles kfree -> 129 cycles 10000 times kmalloc(64) -> 222 cycles kfree -> 120 cycles 10000 times kmalloc(128) -> 343 cycles kfree -> 139 cycles 10000 times kmalloc(256) -> 827 cycles kfree -> 147 cycles 10000 times kmalloc(512) -> 1048 cycles kfree -> 272 cycles 10000 times kmalloc(1024) -> 2043 cycles kfree -> 528 cycles 10000 times kmalloc(2048) -> 4002 cycles kfree -> 571 cycles 10000 times kmalloc(4096) -> 7740 cycles kfree -> 628 cycles 10000 times kmalloc(8192) -> 8062 cycles kfree -> 850 cycles 10000 times kmalloc(16384) -> 8895 cycles kfree -> 1249 cycles After: 10000 times kmalloc(8) -> 190 cycles kfree -> 129 cycles 10000 times kmalloc(16) -> 76 cycles kfree -> 123 cycles 10000 times kmalloc(32) -> 126 cycles kfree -> 124 cycles 10000 times kmalloc(64) -> 181 cycles kfree -> 128 cycles 10000 times kmalloc(128) -> 310 cycles kfree -> 140 cycles 10000 times kmalloc(256) -> 809 cycles kfree -> 165 cycles 10000 times kmalloc(512) -> 1005 cycles kfree -> 269 cycles 10000 times kmalloc(1024) -> 1999 cycles kfree -> 527 cycles 10000 times kmalloc(2048) -> 3967 cycles kfree -> 570 cycles 10000 times kmalloc(4096) -> 7658 cycles kfree -> 637 cycles 10000 times kmalloc(8192) -> 8111 cycles kfree -> 859 cycles 10000 times kmalloc(16384) -> 8791 cycles kfree -> 1173 cycles Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-02-26 01:38:54 +08:00
actual_tid, tid, next_tid(tid));
#endif
stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
Lockless (and preemptless) fastpaths for slub Use the this_cpu_cmpxchg_double functionality to implement a lockless allocation algorithm on arches that support fast this_cpu_ops. Each of the per cpu pointers is paired with a transaction id that ensures that updates of the per cpu information can only occur in sequence on a certain cpu. A transaction id is a "long" integer that is comprised of an event number and the cpu number. The event number is incremented for every change to the per cpu state. This means that the cmpxchg instruction can verify for an update that nothing interfered and that we are updating the percpu structure for the processor where we picked up the information and that we are also currently on that processor when we update the information. This results in a significant decrease of the overhead in the fastpaths. It also makes it easy to adopt the fast path for realtime kernels since this is lockless and does not require the use of the current per cpu area over the critical section. It is only important that the per cpu area is current at the beginning of the critical section and at the end. So there is no need even to disable preemption. Test results show that the fastpath cycle count is reduced by up to ~ 40% (alloc/free test goes from ~140 cycles down to ~80). The slowpath for kfree adds a few cycles. Sadly this does nothing for the slowpath which is where the main issues with performance in slub are but the best case performance rises significantly. (For that see the more complex slub patches that require cmpxchg_double) Kmalloc: alloc/free test Before: 10000 times kmalloc(8)/kfree -> 134 cycles 10000 times kmalloc(16)/kfree -> 152 cycles 10000 times kmalloc(32)/kfree -> 144 cycles 10000 times kmalloc(64)/kfree -> 142 cycles 10000 times kmalloc(128)/kfree -> 142 cycles 10000 times kmalloc(256)/kfree -> 132 cycles 10000 times kmalloc(512)/kfree -> 132 cycles 10000 times kmalloc(1024)/kfree -> 135 cycles 10000 times kmalloc(2048)/kfree -> 135 cycles 10000 times kmalloc(4096)/kfree -> 135 cycles 10000 times kmalloc(8192)/kfree -> 144 cycles 10000 times kmalloc(16384)/kfree -> 754 cycles After: 10000 times kmalloc(8)/kfree -> 78 cycles 10000 times kmalloc(16)/kfree -> 78 cycles 10000 times kmalloc(32)/kfree -> 82 cycles 10000 times kmalloc(64)/kfree -> 88 cycles 10000 times kmalloc(128)/kfree -> 79 cycles 10000 times kmalloc(256)/kfree -> 79 cycles 10000 times kmalloc(512)/kfree -> 85 cycles 10000 times kmalloc(1024)/kfree -> 82 cycles 10000 times kmalloc(2048)/kfree -> 82 cycles 10000 times kmalloc(4096)/kfree -> 85 cycles 10000 times kmalloc(8192)/kfree -> 82 cycles 10000 times kmalloc(16384)/kfree -> 706 cycles Kmalloc: Repeatedly allocate then free test Before: 10000 times kmalloc(8) -> 211 cycles kfree -> 113 cycles 10000 times kmalloc(16) -> 174 cycles kfree -> 115 cycles 10000 times kmalloc(32) -> 235 cycles kfree -> 129 cycles 10000 times kmalloc(64) -> 222 cycles kfree -> 120 cycles 10000 times kmalloc(128) -> 343 cycles kfree -> 139 cycles 10000 times kmalloc(256) -> 827 cycles kfree -> 147 cycles 10000 times kmalloc(512) -> 1048 cycles kfree -> 272 cycles 10000 times kmalloc(1024) -> 2043 cycles kfree -> 528 cycles 10000 times kmalloc(2048) -> 4002 cycles kfree -> 571 cycles 10000 times kmalloc(4096) -> 7740 cycles kfree -> 628 cycles 10000 times kmalloc(8192) -> 8062 cycles kfree -> 850 cycles 10000 times kmalloc(16384) -> 8895 cycles kfree -> 1249 cycles After: 10000 times kmalloc(8) -> 190 cycles kfree -> 129 cycles 10000 times kmalloc(16) -> 76 cycles kfree -> 123 cycles 10000 times kmalloc(32) -> 126 cycles kfree -> 124 cycles 10000 times kmalloc(64) -> 181 cycles kfree -> 128 cycles 10000 times kmalloc(128) -> 310 cycles kfree -> 140 cycles 10000 times kmalloc(256) -> 809 cycles kfree -> 165 cycles 10000 times kmalloc(512) -> 1005 cycles kfree -> 269 cycles 10000 times kmalloc(1024) -> 1999 cycles kfree -> 527 cycles 10000 times kmalloc(2048) -> 3967 cycles kfree -> 570 cycles 10000 times kmalloc(4096) -> 7658 cycles kfree -> 637 cycles 10000 times kmalloc(8192) -> 8111 cycles kfree -> 859 cycles 10000 times kmalloc(16384) -> 8791 cycles kfree -> 1173 cycles Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-02-26 01:38:54 +08:00
}
static void init_kmem_cache_cpus(struct kmem_cache *s)
Lockless (and preemptless) fastpaths for slub Use the this_cpu_cmpxchg_double functionality to implement a lockless allocation algorithm on arches that support fast this_cpu_ops. Each of the per cpu pointers is paired with a transaction id that ensures that updates of the per cpu information can only occur in sequence on a certain cpu. A transaction id is a "long" integer that is comprised of an event number and the cpu number. The event number is incremented for every change to the per cpu state. This means that the cmpxchg instruction can verify for an update that nothing interfered and that we are updating the percpu structure for the processor where we picked up the information and that we are also currently on that processor when we update the information. This results in a significant decrease of the overhead in the fastpaths. It also makes it easy to adopt the fast path for realtime kernels since this is lockless and does not require the use of the current per cpu area over the critical section. It is only important that the per cpu area is current at the beginning of the critical section and at the end. So there is no need even to disable preemption. Test results show that the fastpath cycle count is reduced by up to ~ 40% (alloc/free test goes from ~140 cycles down to ~80). The slowpath for kfree adds a few cycles. Sadly this does nothing for the slowpath which is where the main issues with performance in slub are but the best case performance rises significantly. (For that see the more complex slub patches that require cmpxchg_double) Kmalloc: alloc/free test Before: 10000 times kmalloc(8)/kfree -> 134 cycles 10000 times kmalloc(16)/kfree -> 152 cycles 10000 times kmalloc(32)/kfree -> 144 cycles 10000 times kmalloc(64)/kfree -> 142 cycles 10000 times kmalloc(128)/kfree -> 142 cycles 10000 times kmalloc(256)/kfree -> 132 cycles 10000 times kmalloc(512)/kfree -> 132 cycles 10000 times kmalloc(1024)/kfree -> 135 cycles 10000 times kmalloc(2048)/kfree -> 135 cycles 10000 times kmalloc(4096)/kfree -> 135 cycles 10000 times kmalloc(8192)/kfree -> 144 cycles 10000 times kmalloc(16384)/kfree -> 754 cycles After: 10000 times kmalloc(8)/kfree -> 78 cycles 10000 times kmalloc(16)/kfree -> 78 cycles 10000 times kmalloc(32)/kfree -> 82 cycles 10000 times kmalloc(64)/kfree -> 88 cycles 10000 times kmalloc(128)/kfree -> 79 cycles 10000 times kmalloc(256)/kfree -> 79 cycles 10000 times kmalloc(512)/kfree -> 85 cycles 10000 times kmalloc(1024)/kfree -> 82 cycles 10000 times kmalloc(2048)/kfree -> 82 cycles 10000 times kmalloc(4096)/kfree -> 85 cycles 10000 times kmalloc(8192)/kfree -> 82 cycles 10000 times kmalloc(16384)/kfree -> 706 cycles Kmalloc: Repeatedly allocate then free test Before: 10000 times kmalloc(8) -> 211 cycles kfree -> 113 cycles 10000 times kmalloc(16) -> 174 cycles kfree -> 115 cycles 10000 times kmalloc(32) -> 235 cycles kfree -> 129 cycles 10000 times kmalloc(64) -> 222 cycles kfree -> 120 cycles 10000 times kmalloc(128) -> 343 cycles kfree -> 139 cycles 10000 times kmalloc(256) -> 827 cycles kfree -> 147 cycles 10000 times kmalloc(512) -> 1048 cycles kfree -> 272 cycles 10000 times kmalloc(1024) -> 2043 cycles kfree -> 528 cycles 10000 times kmalloc(2048) -> 4002 cycles kfree -> 571 cycles 10000 times kmalloc(4096) -> 7740 cycles kfree -> 628 cycles 10000 times kmalloc(8192) -> 8062 cycles kfree -> 850 cycles 10000 times kmalloc(16384) -> 8895 cycles kfree -> 1249 cycles After: 10000 times kmalloc(8) -> 190 cycles kfree -> 129 cycles 10000 times kmalloc(16) -> 76 cycles kfree -> 123 cycles 10000 times kmalloc(32) -> 126 cycles kfree -> 124 cycles 10000 times kmalloc(64) -> 181 cycles kfree -> 128 cycles 10000 times kmalloc(128) -> 310 cycles kfree -> 140 cycles 10000 times kmalloc(256) -> 809 cycles kfree -> 165 cycles 10000 times kmalloc(512) -> 1005 cycles kfree -> 269 cycles 10000 times kmalloc(1024) -> 1999 cycles kfree -> 527 cycles 10000 times kmalloc(2048) -> 3967 cycles kfree -> 570 cycles 10000 times kmalloc(4096) -> 7658 cycles kfree -> 637 cycles 10000 times kmalloc(8192) -> 8111 cycles kfree -> 859 cycles 10000 times kmalloc(16384) -> 8791 cycles kfree -> 1173 cycles Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-02-26 01:38:54 +08:00
{
int cpu;
for_each_possible_cpu(cpu)
per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
}
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/*
* Remove the cpu slab
*/
static void deactivate_slab(struct kmem_cache *s, struct page *page,
void *freelist, struct kmem_cache_cpu *c)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
{
enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
struct kmem_cache_node *n = get_node(s, page_to_nid(page));
int lock = 0, free_delta = 0;
enum slab_modes l = M_NONE, m = M_NONE;
void *nextfree, *freelist_iter, *freelist_tail;
int tail = DEACTIVATE_TO_HEAD;
struct page new;
struct page old;
if (page->freelist) {
stat(s, DEACTIVATE_REMOTE_FREES);
tail = DEACTIVATE_TO_TAIL;
}
slub: support concurrent local and remote frees and allocs on a slab Avoid atomic overhead in slab_alloc and slab_free SLUB needs to use the slab_lock for the per cpu slabs to synchronize with potential kfree operations. This patch avoids that need by moving all free objects onto a lockless_freelist. The regular freelist continues to exist and will be used to free objects. So while we consume the lockless_freelist the regular freelist may build up objects. If we are out of objects on the lockless_freelist then we may check the regular freelist. If it has objects then we move those over to the lockless_freelist and do this again. There is a significant savings in terms of atomic operations that have to be performed. We can even free directly to the lockless_freelist if we know that we are running on the same processor. So this speeds up short lived objects. They may be allocated and freed without taking the slab_lock. This is particular good for netperf. In order to maximize the effect of the new faster hotpath we extract the hottest performance pieces into inlined functions. These are then inlined into kmem_cache_alloc and kmem_cache_free. So hotpath allocation and freeing no longer requires a subroutine call within SLUB. [I am not sure that it is worth doing this because it changes the easy to read structure of slub just to reduce atomic ops. However, there is someone out there with a benchmark on 4 way and 8 way processor systems that seems to show a 5% regression vs. Slab. Seems that the regression is due to increased atomic operations use vs. SLAB in SLUB). I wonder if this is applicable or discernable at all in a real workload?] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-10 18:15:16 +08:00
/*
* Stage one: Count the objects on cpu's freelist as free_delta and
* remember the last object in freelist_tail for later splicing.
*/
freelist_tail = NULL;
freelist_iter = freelist;
while (freelist_iter) {
nextfree = get_freepointer(s, freelist_iter);
mm/slub.c: fix corrupted freechain in deactivate_slab() The slub_debug is able to fix the corrupted slab freelist/page. However, alloc_debug_processing() only checks the validity of current and next freepointer during allocation path. As a result, once some objects have their freepointers corrupted, deactivate_slab() may lead to page fault. Below is from a test kernel module when 'slub_debug=PUF,kmalloc-128 slub_nomerge'. The test kernel corrupts the freepointer of one free object on purpose. Unfortunately, deactivate_slab() does not detect it when iterating the freechain. BUG: unable to handle page fault for address: 00000000123456f8 #PF: supervisor read access in kernel mode #PF: error_code(0x0000) - not-present page PGD 0 P4D 0 Oops: 0000 [#1] SMP PTI ... ... RIP: 0010:deactivate_slab.isra.92+0xed/0x490 ... ... Call Trace: ___slab_alloc+0x536/0x570 __slab_alloc+0x17/0x30 __kmalloc+0x1d9/0x200 ext4_htree_store_dirent+0x30/0xf0 htree_dirblock_to_tree+0xcb/0x1c0 ext4_htree_fill_tree+0x1bc/0x2d0 ext4_readdir+0x54f/0x920 iterate_dir+0x88/0x190 __x64_sys_getdents+0xa6/0x140 do_syscall_64+0x49/0x170 entry_SYSCALL_64_after_hwframe+0x44/0xa9 Therefore, this patch adds extra consistency check in deactivate_slab(). Once an object's freepointer is corrupted, all following objects starting at this object are isolated. [akpm@linux-foundation.org: fix build with CONFIG_SLAB_DEBUG=n] Signed-off-by: Dongli Zhang <dongli.zhang@oracle.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Cc: Joe Jin <joe.jin@oracle.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Link: http://lkml.kernel.org/r/20200331031450.12182-1-dongli.zhang@oracle.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-02 12:45:47 +08:00
/*
* If 'nextfree' is invalid, it is possible that the object at
* 'freelist_iter' is already corrupted. So isolate all objects
* starting at 'freelist_iter' by skipping them.
mm/slub.c: fix corrupted freechain in deactivate_slab() The slub_debug is able to fix the corrupted slab freelist/page. However, alloc_debug_processing() only checks the validity of current and next freepointer during allocation path. As a result, once some objects have their freepointers corrupted, deactivate_slab() may lead to page fault. Below is from a test kernel module when 'slub_debug=PUF,kmalloc-128 slub_nomerge'. The test kernel corrupts the freepointer of one free object on purpose. Unfortunately, deactivate_slab() does not detect it when iterating the freechain. BUG: unable to handle page fault for address: 00000000123456f8 #PF: supervisor read access in kernel mode #PF: error_code(0x0000) - not-present page PGD 0 P4D 0 Oops: 0000 [#1] SMP PTI ... ... RIP: 0010:deactivate_slab.isra.92+0xed/0x490 ... ... Call Trace: ___slab_alloc+0x536/0x570 __slab_alloc+0x17/0x30 __kmalloc+0x1d9/0x200 ext4_htree_store_dirent+0x30/0xf0 htree_dirblock_to_tree+0xcb/0x1c0 ext4_htree_fill_tree+0x1bc/0x2d0 ext4_readdir+0x54f/0x920 iterate_dir+0x88/0x190 __x64_sys_getdents+0xa6/0x140 do_syscall_64+0x49/0x170 entry_SYSCALL_64_after_hwframe+0x44/0xa9 Therefore, this patch adds extra consistency check in deactivate_slab(). Once an object's freepointer is corrupted, all following objects starting at this object are isolated. [akpm@linux-foundation.org: fix build with CONFIG_SLAB_DEBUG=n] Signed-off-by: Dongli Zhang <dongli.zhang@oracle.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Cc: Joe Jin <joe.jin@oracle.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Link: http://lkml.kernel.org/r/20200331031450.12182-1-dongli.zhang@oracle.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-02 12:45:47 +08:00
*/
if (freelist_corrupted(s, page, &freelist_iter, nextfree))
mm/slub.c: fix corrupted freechain in deactivate_slab() The slub_debug is able to fix the corrupted slab freelist/page. However, alloc_debug_processing() only checks the validity of current and next freepointer during allocation path. As a result, once some objects have their freepointers corrupted, deactivate_slab() may lead to page fault. Below is from a test kernel module when 'slub_debug=PUF,kmalloc-128 slub_nomerge'. The test kernel corrupts the freepointer of one free object on purpose. Unfortunately, deactivate_slab() does not detect it when iterating the freechain. BUG: unable to handle page fault for address: 00000000123456f8 #PF: supervisor read access in kernel mode #PF: error_code(0x0000) - not-present page PGD 0 P4D 0 Oops: 0000 [#1] SMP PTI ... ... RIP: 0010:deactivate_slab.isra.92+0xed/0x490 ... ... Call Trace: ___slab_alloc+0x536/0x570 __slab_alloc+0x17/0x30 __kmalloc+0x1d9/0x200 ext4_htree_store_dirent+0x30/0xf0 htree_dirblock_to_tree+0xcb/0x1c0 ext4_htree_fill_tree+0x1bc/0x2d0 ext4_readdir+0x54f/0x920 iterate_dir+0x88/0x190 __x64_sys_getdents+0xa6/0x140 do_syscall_64+0x49/0x170 entry_SYSCALL_64_after_hwframe+0x44/0xa9 Therefore, this patch adds extra consistency check in deactivate_slab(). Once an object's freepointer is corrupted, all following objects starting at this object are isolated. [akpm@linux-foundation.org: fix build with CONFIG_SLAB_DEBUG=n] Signed-off-by: Dongli Zhang <dongli.zhang@oracle.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Cc: Joe Jin <joe.jin@oracle.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Link: http://lkml.kernel.org/r/20200331031450.12182-1-dongli.zhang@oracle.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-02 12:45:47 +08:00
break;
freelist_tail = freelist_iter;
free_delta++;
freelist_iter = nextfree;
}
slub: support concurrent local and remote frees and allocs on a slab Avoid atomic overhead in slab_alloc and slab_free SLUB needs to use the slab_lock for the per cpu slabs to synchronize with potential kfree operations. This patch avoids that need by moving all free objects onto a lockless_freelist. The regular freelist continues to exist and will be used to free objects. So while we consume the lockless_freelist the regular freelist may build up objects. If we are out of objects on the lockless_freelist then we may check the regular freelist. If it has objects then we move those over to the lockless_freelist and do this again. There is a significant savings in terms of atomic operations that have to be performed. We can even free directly to the lockless_freelist if we know that we are running on the same processor. So this speeds up short lived objects. They may be allocated and freed without taking the slab_lock. This is particular good for netperf. In order to maximize the effect of the new faster hotpath we extract the hottest performance pieces into inlined functions. These are then inlined into kmem_cache_alloc and kmem_cache_free. So hotpath allocation and freeing no longer requires a subroutine call within SLUB. [I am not sure that it is worth doing this because it changes the easy to read structure of slub just to reduce atomic ops. However, there is someone out there with a benchmark on 4 way and 8 way processor systems that seems to show a 5% regression vs. Slab. Seems that the regression is due to increased atomic operations use vs. SLAB in SLUB). I wonder if this is applicable or discernable at all in a real workload?] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-10 18:15:16 +08:00
/*
* Stage two: Unfreeze the page while splicing the per-cpu
* freelist to the head of page's freelist.
*
* Ensure that the page is unfrozen while the list presence
* reflects the actual number of objects during unfreeze.
*
* We setup the list membership and then perform a cmpxchg
* with the count. If there is a mismatch then the page
* is not unfrozen but the page is on the wrong list.
*
* Then we restart the process which may have to remove
* the page from the list that we just put it on again
* because the number of objects in the slab may have
* changed.
slub: support concurrent local and remote frees and allocs on a slab Avoid atomic overhead in slab_alloc and slab_free SLUB needs to use the slab_lock for the per cpu slabs to synchronize with potential kfree operations. This patch avoids that need by moving all free objects onto a lockless_freelist. The regular freelist continues to exist and will be used to free objects. So while we consume the lockless_freelist the regular freelist may build up objects. If we are out of objects on the lockless_freelist then we may check the regular freelist. If it has objects then we move those over to the lockless_freelist and do this again. There is a significant savings in terms of atomic operations that have to be performed. We can even free directly to the lockless_freelist if we know that we are running on the same processor. So this speeds up short lived objects. They may be allocated and freed without taking the slab_lock. This is particular good for netperf. In order to maximize the effect of the new faster hotpath we extract the hottest performance pieces into inlined functions. These are then inlined into kmem_cache_alloc and kmem_cache_free. So hotpath allocation and freeing no longer requires a subroutine call within SLUB. [I am not sure that it is worth doing this because it changes the easy to read structure of slub just to reduce atomic ops. However, there is someone out there with a benchmark on 4 way and 8 way processor systems that seems to show a 5% regression vs. Slab. Seems that the regression is due to increased atomic operations use vs. SLAB in SLUB). I wonder if this is applicable or discernable at all in a real workload?] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-10 18:15:16 +08:00
*/
redo:
slub: support concurrent local and remote frees and allocs on a slab Avoid atomic overhead in slab_alloc and slab_free SLUB needs to use the slab_lock for the per cpu slabs to synchronize with potential kfree operations. This patch avoids that need by moving all free objects onto a lockless_freelist. The regular freelist continues to exist and will be used to free objects. So while we consume the lockless_freelist the regular freelist may build up objects. If we are out of objects on the lockless_freelist then we may check the regular freelist. If it has objects then we move those over to the lockless_freelist and do this again. There is a significant savings in terms of atomic operations that have to be performed. We can even free directly to the lockless_freelist if we know that we are running on the same processor. So this speeds up short lived objects. They may be allocated and freed without taking the slab_lock. This is particular good for netperf. In order to maximize the effect of the new faster hotpath we extract the hottest performance pieces into inlined functions. These are then inlined into kmem_cache_alloc and kmem_cache_free. So hotpath allocation and freeing no longer requires a subroutine call within SLUB. [I am not sure that it is worth doing this because it changes the easy to read structure of slub just to reduce atomic ops. However, there is someone out there with a benchmark on 4 way and 8 way processor systems that seems to show a 5% regression vs. Slab. Seems that the regression is due to increased atomic operations use vs. SLAB in SLUB). I wonder if this is applicable or discernable at all in a real workload?] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-10 18:15:16 +08:00
old.freelist = READ_ONCE(page->freelist);
old.counters = READ_ONCE(page->counters);
VM_BUG_ON(!old.frozen);
/* Determine target state of the slab */
new.counters = old.counters;
if (freelist_tail) {
new.inuse -= free_delta;
set_freepointer(s, freelist_tail, old.freelist);
new.freelist = freelist;
} else
new.freelist = old.freelist;
new.frozen = 0;
if (!new.inuse && n->nr_partial >= s->min_partial)
m = M_FREE;
else if (new.freelist) {
m = M_PARTIAL;
if (!lock) {
lock = 1;
/*
* Taking the spinlock removes the possibility
* that acquire_slab() will see a slab page that
* is frozen
*/
spin_lock(&n->list_lock);
}
} else {
m = M_FULL;
if (kmem_cache_debug_flags(s, SLAB_STORE_USER) && !lock) {
lock = 1;
/*
* This also ensures that the scanning of full
* slabs from diagnostic functions will not see
* any frozen slabs.
*/
spin_lock(&n->list_lock);
}
}
if (l != m) {
if (l == M_PARTIAL)
remove_partial(n, page);
else if (l == M_FULL)
remove_full(s, n, page);
if (m == M_PARTIAL)
add_partial(n, page, tail);
else if (m == M_FULL)
add_full(s, n, page);
}
l = m;
if (!__cmpxchg_double_slab(s, page,
old.freelist, old.counters,
new.freelist, new.counters,
"unfreezing slab"))
goto redo;
if (lock)
spin_unlock(&n->list_lock);
if (m == M_PARTIAL)
stat(s, tail);
else if (m == M_FULL)
stat(s, DEACTIVATE_FULL);
else if (m == M_FREE) {
stat(s, DEACTIVATE_EMPTY);
discard_slab(s, page);
stat(s, FREE_SLAB);
slub: support concurrent local and remote frees and allocs on a slab Avoid atomic overhead in slab_alloc and slab_free SLUB needs to use the slab_lock for the per cpu slabs to synchronize with potential kfree operations. This patch avoids that need by moving all free objects onto a lockless_freelist. The regular freelist continues to exist and will be used to free objects. So while we consume the lockless_freelist the regular freelist may build up objects. If we are out of objects on the lockless_freelist then we may check the regular freelist. If it has objects then we move those over to the lockless_freelist and do this again. There is a significant savings in terms of atomic operations that have to be performed. We can even free directly to the lockless_freelist if we know that we are running on the same processor. So this speeds up short lived objects. They may be allocated and freed without taking the slab_lock. This is particular good for netperf. In order to maximize the effect of the new faster hotpath we extract the hottest performance pieces into inlined functions. These are then inlined into kmem_cache_alloc and kmem_cache_free. So hotpath allocation and freeing no longer requires a subroutine call within SLUB. [I am not sure that it is worth doing this because it changes the easy to read structure of slub just to reduce atomic ops. However, there is someone out there with a benchmark on 4 way and 8 way processor systems that seems to show a 5% regression vs. Slab. Seems that the regression is due to increased atomic operations use vs. SLAB in SLUB). I wonder if this is applicable or discernable at all in a real workload?] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-10 18:15:16 +08:00
}
c->page = NULL;
c->freelist = NULL;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
/*
* Unfreeze all the cpu partial slabs.
*
* This function must be called with interrupts disabled
* for the cpu using c (or some other guarantee must be there
* to guarantee no concurrent accesses).
*/
static void unfreeze_partials(struct kmem_cache *s,
struct kmem_cache_cpu *c)
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
{
#ifdef CONFIG_SLUB_CPU_PARTIAL
slub: refactoring unfreeze_partials() Current implementation of unfreeze_partials() is so complicated, but benefit from it is insignificant. In addition many code in do {} while loop have a bad influence to a fail rate of cmpxchg_double_slab. Under current implementation which test status of cpu partial slab and acquire list_lock in do {} while loop, we don't need to acquire a list_lock and gain a little benefit when front of the cpu partial slab is to be discarded, but this is a rare case. In case that add_partial is performed and cmpxchg_double_slab is failed, remove_partial should be called case by case. I think that these are disadvantages of current implementation, so I do refactoring unfreeze_partials(). Minimizing code in do {} while loop introduce a reduced fail rate of cmpxchg_double_slab. Below is output of 'slabinfo -r kmalloc-256' when './perf stat -r 33 hackbench 50 process 4000 > /dev/null' is done. ** before ** Cmpxchg_double Looping ------------------------ Locked Cmpxchg Double redos 182685 Unlocked Cmpxchg Double redos 0 ** after ** Cmpxchg_double Looping ------------------------ Locked Cmpxchg Double redos 177995 Unlocked Cmpxchg Double redos 1 We can see cmpxchg_double_slab fail rate is improved slightly. Bolow is output of './perf stat -r 30 hackbench 50 process 4000 > /dev/null'. ** before ** Performance counter stats for './hackbench 50 process 4000' (30 runs): 108517.190463 task-clock # 7.926 CPUs utilized ( +- 0.24% ) 2,919,550 context-switches # 0.027 M/sec ( +- 3.07% ) 100,774 CPU-migrations # 0.929 K/sec ( +- 4.72% ) 124,201 page-faults # 0.001 M/sec ( +- 0.15% ) 401,500,234,387 cycles # 3.700 GHz ( +- 0.24% ) <not supported> stalled-cycles-frontend <not supported> stalled-cycles-backend 250,576,913,354 instructions # 0.62 insns per cycle ( +- 0.13% ) 45,934,956,860 branches # 423.297 M/sec ( +- 0.14% ) 188,219,787 branch-misses # 0.41% of all branches ( +- 0.56% ) 13.691837307 seconds time elapsed ( +- 0.24% ) ** after ** Performance counter stats for './hackbench 50 process 4000' (30 runs): 107784.479767 task-clock # 7.928 CPUs utilized ( +- 0.22% ) 2,834,781 context-switches # 0.026 M/sec ( +- 2.33% ) 93,083 CPU-migrations # 0.864 K/sec ( +- 3.45% ) 123,967 page-faults # 0.001 M/sec ( +- 0.15% ) 398,781,421,836 cycles # 3.700 GHz ( +- 0.22% ) <not supported> stalled-cycles-frontend <not supported> stalled-cycles-backend 250,189,160,419 instructions # 0.63 insns per cycle ( +- 0.09% ) 45,855,370,128 branches # 425.436 M/sec ( +- 0.10% ) 169,881,248 branch-misses # 0.37% of all branches ( +- 0.43% ) 13.596272341 seconds time elapsed ( +- 0.22% ) No regression is found, but rather we can see slightly better result. Acked-by: Christoph Lameter <cl@linux.com> Signed-off-by: Joonsoo Kim <js1304@gmail.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2012-06-09 01:23:16 +08:00
struct kmem_cache_node *n = NULL, *n2 = NULL;
struct page *page, *discard_page = NULL;
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
while ((page = slub_percpu_partial(c))) {
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
struct page new;
struct page old;
slub_set_percpu_partial(c, page);
slub: refactoring unfreeze_partials() Current implementation of unfreeze_partials() is so complicated, but benefit from it is insignificant. In addition many code in do {} while loop have a bad influence to a fail rate of cmpxchg_double_slab. Under current implementation which test status of cpu partial slab and acquire list_lock in do {} while loop, we don't need to acquire a list_lock and gain a little benefit when front of the cpu partial slab is to be discarded, but this is a rare case. In case that add_partial is performed and cmpxchg_double_slab is failed, remove_partial should be called case by case. I think that these are disadvantages of current implementation, so I do refactoring unfreeze_partials(). Minimizing code in do {} while loop introduce a reduced fail rate of cmpxchg_double_slab. Below is output of 'slabinfo -r kmalloc-256' when './perf stat -r 33 hackbench 50 process 4000 > /dev/null' is done. ** before ** Cmpxchg_double Looping ------------------------ Locked Cmpxchg Double redos 182685 Unlocked Cmpxchg Double redos 0 ** after ** Cmpxchg_double Looping ------------------------ Locked Cmpxchg Double redos 177995 Unlocked Cmpxchg Double redos 1 We can see cmpxchg_double_slab fail rate is improved slightly. Bolow is output of './perf stat -r 30 hackbench 50 process 4000 > /dev/null'. ** before ** Performance counter stats for './hackbench 50 process 4000' (30 runs): 108517.190463 task-clock # 7.926 CPUs utilized ( +- 0.24% ) 2,919,550 context-switches # 0.027 M/sec ( +- 3.07% ) 100,774 CPU-migrations # 0.929 K/sec ( +- 4.72% ) 124,201 page-faults # 0.001 M/sec ( +- 0.15% ) 401,500,234,387 cycles # 3.700 GHz ( +- 0.24% ) <not supported> stalled-cycles-frontend <not supported> stalled-cycles-backend 250,576,913,354 instructions # 0.62 insns per cycle ( +- 0.13% ) 45,934,956,860 branches # 423.297 M/sec ( +- 0.14% ) 188,219,787 branch-misses # 0.41% of all branches ( +- 0.56% ) 13.691837307 seconds time elapsed ( +- 0.24% ) ** after ** Performance counter stats for './hackbench 50 process 4000' (30 runs): 107784.479767 task-clock # 7.928 CPUs utilized ( +- 0.22% ) 2,834,781 context-switches # 0.026 M/sec ( +- 2.33% ) 93,083 CPU-migrations # 0.864 K/sec ( +- 3.45% ) 123,967 page-faults # 0.001 M/sec ( +- 0.15% ) 398,781,421,836 cycles # 3.700 GHz ( +- 0.22% ) <not supported> stalled-cycles-frontend <not supported> stalled-cycles-backend 250,189,160,419 instructions # 0.63 insns per cycle ( +- 0.09% ) 45,855,370,128 branches # 425.436 M/sec ( +- 0.10% ) 169,881,248 branch-misses # 0.37% of all branches ( +- 0.43% ) 13.596272341 seconds time elapsed ( +- 0.22% ) No regression is found, but rather we can see slightly better result. Acked-by: Christoph Lameter <cl@linux.com> Signed-off-by: Joonsoo Kim <js1304@gmail.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2012-06-09 01:23:16 +08:00
n2 = get_node(s, page_to_nid(page));
if (n != n2) {
if (n)
spin_unlock(&n->list_lock);
n = n2;
spin_lock(&n->list_lock);
}
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
do {
old.freelist = page->freelist;
old.counters = page->counters;
VM_BUG_ON(!old.frozen);
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
new.counters = old.counters;
new.freelist = old.freelist;
new.frozen = 0;
} while (!__cmpxchg_double_slab(s, page,
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
old.freelist, old.counters,
new.freelist, new.counters,
"unfreezing slab"));
if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
page->next = discard_page;
discard_page = page;
slub: refactoring unfreeze_partials() Current implementation of unfreeze_partials() is so complicated, but benefit from it is insignificant. In addition many code in do {} while loop have a bad influence to a fail rate of cmpxchg_double_slab. Under current implementation which test status of cpu partial slab and acquire list_lock in do {} while loop, we don't need to acquire a list_lock and gain a little benefit when front of the cpu partial slab is to be discarded, but this is a rare case. In case that add_partial is performed and cmpxchg_double_slab is failed, remove_partial should be called case by case. I think that these are disadvantages of current implementation, so I do refactoring unfreeze_partials(). Minimizing code in do {} while loop introduce a reduced fail rate of cmpxchg_double_slab. Below is output of 'slabinfo -r kmalloc-256' when './perf stat -r 33 hackbench 50 process 4000 > /dev/null' is done. ** before ** Cmpxchg_double Looping ------------------------ Locked Cmpxchg Double redos 182685 Unlocked Cmpxchg Double redos 0 ** after ** Cmpxchg_double Looping ------------------------ Locked Cmpxchg Double redos 177995 Unlocked Cmpxchg Double redos 1 We can see cmpxchg_double_slab fail rate is improved slightly. Bolow is output of './perf stat -r 30 hackbench 50 process 4000 > /dev/null'. ** before ** Performance counter stats for './hackbench 50 process 4000' (30 runs): 108517.190463 task-clock # 7.926 CPUs utilized ( +- 0.24% ) 2,919,550 context-switches # 0.027 M/sec ( +- 3.07% ) 100,774 CPU-migrations # 0.929 K/sec ( +- 4.72% ) 124,201 page-faults # 0.001 M/sec ( +- 0.15% ) 401,500,234,387 cycles # 3.700 GHz ( +- 0.24% ) <not supported> stalled-cycles-frontend <not supported> stalled-cycles-backend 250,576,913,354 instructions # 0.62 insns per cycle ( +- 0.13% ) 45,934,956,860 branches # 423.297 M/sec ( +- 0.14% ) 188,219,787 branch-misses # 0.41% of all branches ( +- 0.56% ) 13.691837307 seconds time elapsed ( +- 0.24% ) ** after ** Performance counter stats for './hackbench 50 process 4000' (30 runs): 107784.479767 task-clock # 7.928 CPUs utilized ( +- 0.22% ) 2,834,781 context-switches # 0.026 M/sec ( +- 2.33% ) 93,083 CPU-migrations # 0.864 K/sec ( +- 3.45% ) 123,967 page-faults # 0.001 M/sec ( +- 0.15% ) 398,781,421,836 cycles # 3.700 GHz ( +- 0.22% ) <not supported> stalled-cycles-frontend <not supported> stalled-cycles-backend 250,189,160,419 instructions # 0.63 insns per cycle ( +- 0.09% ) 45,855,370,128 branches # 425.436 M/sec ( +- 0.10% ) 169,881,248 branch-misses # 0.37% of all branches ( +- 0.43% ) 13.596272341 seconds time elapsed ( +- 0.22% ) No regression is found, but rather we can see slightly better result. Acked-by: Christoph Lameter <cl@linux.com> Signed-off-by: Joonsoo Kim <js1304@gmail.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2012-06-09 01:23:16 +08:00
} else {
add_partial(n, page, DEACTIVATE_TO_TAIL);
stat(s, FREE_ADD_PARTIAL);
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
}
}
if (n)
spin_unlock(&n->list_lock);
while (discard_page) {
page = discard_page;
discard_page = discard_page->next;
stat(s, DEACTIVATE_EMPTY);
discard_slab(s, page);
stat(s, FREE_SLAB);
}
#endif /* CONFIG_SLUB_CPU_PARTIAL */
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
}
/*
* Put a page that was just frozen (in __slab_free|get_partial_node) into a
* partial page slot if available.
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
*
* If we did not find a slot then simply move all the partials to the
* per node partial list.
*/
static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
{
#ifdef CONFIG_SLUB_CPU_PARTIAL
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
struct page *oldpage;
int pages;
int pobjects;
slub: make dead caches discard free slabs immediately To speed up further allocations SLUB may store empty slabs in per cpu/node partial lists instead of freeing them immediately. This prevents per memcg caches destruction, because kmem caches created for a memory cgroup are only destroyed after the last page charged to the cgroup is freed. To fix this issue, this patch resurrects approach first proposed in [1]. It forbids SLUB to cache empty slabs after the memory cgroup that the cache belongs to was destroyed. It is achieved by setting kmem_cache's cpu_partial and min_partial constants to 0 and tuning put_cpu_partial() so that it would drop frozen empty slabs immediately if cpu_partial = 0. The runtime overhead is minimal. From all the hot functions, we only touch relatively cold put_cpu_partial(): we make it call unfreeze_partials() after freezing a slab that belongs to an offline memory cgroup. Since slab freezing exists to avoid moving slabs from/to a partial list on free/alloc, and there can't be allocations from dead caches, it shouldn't cause any overhead. We do have to disable preemption for put_cpu_partial() to achieve that though. The original patch was accepted well and even merged to the mm tree. However, I decided to withdraw it due to changes happening to the memcg core at that time. I had an idea of introducing per-memcg shrinkers for kmem caches, but now, as memcg has finally settled down, I do not see it as an option, because SLUB shrinker would be too costly to call since SLUB does not keep free slabs on a separate list. Besides, we currently do not even call per-memcg shrinkers for offline memcgs. Overall, it would introduce much more complexity to both SLUB and memcg than this small patch. Regarding to SLAB, there's no problem with it, because it shrinks per-cpu/node caches periodically. Thanks to list_lru reparenting, we no longer keep entries for offline cgroups in per-memcg arrays (such as memcg_cache_params->memcg_caches), so we do not have to bother if a per-memcg cache will be shrunk a bit later than it could be. [1] http://thread.gmane.org/gmane.linux.kernel.mm/118649/focus=118650 Signed-off-by: Vladimir Davydov <vdavydov@parallels.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-13 06:59:47 +08:00
preempt_disable();
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
do {
pages = 0;
pobjects = 0;
oldpage = this_cpu_read(s->cpu_slab->partial);
if (oldpage) {
pobjects = oldpage->pobjects;
pages = oldpage->pages;
if (drain && pobjects > slub_cpu_partial(s)) {
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
unsigned long flags;
/*
* partial array is full. Move the existing
* set to the per node partial list.
*/
local_irq_save(flags);
unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
local_irq_restore(flags);
oldpage = NULL;
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
pobjects = 0;
pages = 0;
stat(s, CPU_PARTIAL_DRAIN);
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
}
}
pages++;
pobjects += page->objects - page->inuse;
page->pages = pages;
page->pobjects = pobjects;
page->next = oldpage;
} while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
!= oldpage);
if (unlikely(!slub_cpu_partial(s))) {
slub: make dead caches discard free slabs immediately To speed up further allocations SLUB may store empty slabs in per cpu/node partial lists instead of freeing them immediately. This prevents per memcg caches destruction, because kmem caches created for a memory cgroup are only destroyed after the last page charged to the cgroup is freed. To fix this issue, this patch resurrects approach first proposed in [1]. It forbids SLUB to cache empty slabs after the memory cgroup that the cache belongs to was destroyed. It is achieved by setting kmem_cache's cpu_partial and min_partial constants to 0 and tuning put_cpu_partial() so that it would drop frozen empty slabs immediately if cpu_partial = 0. The runtime overhead is minimal. From all the hot functions, we only touch relatively cold put_cpu_partial(): we make it call unfreeze_partials() after freezing a slab that belongs to an offline memory cgroup. Since slab freezing exists to avoid moving slabs from/to a partial list on free/alloc, and there can't be allocations from dead caches, it shouldn't cause any overhead. We do have to disable preemption for put_cpu_partial() to achieve that though. The original patch was accepted well and even merged to the mm tree. However, I decided to withdraw it due to changes happening to the memcg core at that time. I had an idea of introducing per-memcg shrinkers for kmem caches, but now, as memcg has finally settled down, I do not see it as an option, because SLUB shrinker would be too costly to call since SLUB does not keep free slabs on a separate list. Besides, we currently do not even call per-memcg shrinkers for offline memcgs. Overall, it would introduce much more complexity to both SLUB and memcg than this small patch. Regarding to SLAB, there's no problem with it, because it shrinks per-cpu/node caches periodically. Thanks to list_lru reparenting, we no longer keep entries for offline cgroups in per-memcg arrays (such as memcg_cache_params->memcg_caches), so we do not have to bother if a per-memcg cache will be shrunk a bit later than it could be. [1] http://thread.gmane.org/gmane.linux.kernel.mm/118649/focus=118650 Signed-off-by: Vladimir Davydov <vdavydov@parallels.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-13 06:59:47 +08:00
unsigned long flags;
local_irq_save(flags);
unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
local_irq_restore(flags);
}
preempt_enable();
#endif /* CONFIG_SLUB_CPU_PARTIAL */
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
}
SLUB: Avoid page struct cacheline bouncing due to remote frees to cpu slab A remote free may access the same page struct that also contains the lockless freelist for the cpu slab. If objects have a short lifetime and are freed by a different processor then remote frees back to the slab from which we are currently allocating are frequent. The cacheline with the page struct needs to be repeately acquired in exclusive mode by both the allocating thread and the freeing thread. If this is frequent enough then performance will suffer because of cacheline bouncing. This patchset puts the lockless_freelist pointer in its own cacheline. In order to make that happen we introduce a per cpu structure called kmem_cache_cpu. Instead of keeping an array of pointers to page structs we now keep an array to a per cpu structure that--among other things--contains the pointer to the lockless freelist. The freeing thread can then keep possession of exclusive access to the page struct cacheline while the allocating thread keeps its exclusive access to the cacheline containing the per cpu structure. This works as long as the allocating cpu is able to service its request from the lockless freelist. If the lockless freelist runs empty then the allocating thread needs to acquire exclusive access to the cacheline with the page struct lock the slab. The allocating thread will then check if new objects were freed to the per cpu slab. If so it will keep the slab as the cpu slab and continue with the recently remote freed objects. So the allocating thread can take a series of just freed remote pages and dish them out again. Ideally allocations could be just recycling objects in the same slab this way which will lead to an ideal allocation / remote free pattern. The number of objects that can be handled in this way is limited by the capacity of one slab. Increasing slab size via slub_min_objects/ slub_max_order may increase the number of objects and therefore performance. If the allocating thread runs out of objects and finds that no objects were put back by the remote processor then it will retrieve a new slab (from the partial lists or from the page allocator) and start with a whole new set of objects while the remote thread may still be freeing objects to the old cpu slab. This may then repeat until the new slab is also exhausted. If remote freeing has freed objects in the earlier slab then that earlier slab will now be on the partial freelist and the allocating thread will pick that slab next for allocation. So the loop is extended. However, both threads need to take the list_lock to make the swizzling via the partial list happen. It is likely that this kind of scheme will keep the objects being passed around to a small set that can be kept in the cpu caches leading to increased performance. More code cleanups become possible: - Instead of passing a cpu we can now pass a kmem_cache_cpu structure around. Allows reducing the number of parameters to various functions. - Can define a new node_match() function for NUMA to encapsulate locality checks. Effect on allocations: Cachelines touched before this patch: Write: page cache struct and first cacheline of object Cachelines touched after this patch: Write: kmem_cache_cpu cacheline and first cacheline of object Read: page cache struct (but see later patch that avoids touching that cacheline) The handling when the lockless alloc list runs empty gets to be a bit more complicated since another cacheline has now to be written to. But that is halfway out of the hot path. Effect on freeing: Cachelines touched before this patch: Write: page_struct and first cacheline of object Cachelines touched after this patch depending on how we free: Write(to cpu_slab): kmem_cache_cpu struct and first cacheline of object Write(to other): page struct and first cacheline of object Read(to cpu_slab): page struct to id slab etc. (but see later patch that avoids touching the page struct on free) Read(to other): cpu local kmem_cache_cpu struct to verify its not the cpu slab. Summary: Pro: - Distinct cachelines so that concurrent remote frees and local allocs on a cpuslab can occur without cacheline bouncing. - Avoids potential bouncing cachelines because of neighboring per cpu pointer updates in kmem_cache's cpu_slab structure since it now grows to a cacheline (Therefore remove the comment that talks about that concern). Cons: - Freeing objects now requires the reading of one additional cacheline. That can be mitigated for some cases by the following patches but its not possible to completely eliminate these references. - Memory usage grows slightly. The size of each per cpu object is blown up from one word (pointing to the page_struct) to one cacheline with various data. So this is NR_CPUS*NR_SLABS*L1_BYTES more memory use. Lets say NR_SLABS is 100 and a cache line size of 128 then we have just increased SLAB metadata requirements by 12.8k per cpu. (Another later patch reduces these requirements) Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 16:26:05 +08:00
static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
{
stat(s, CPUSLAB_FLUSH);
deactivate_slab(s, c->page, c->freelist, c);
c->tid = next_tid(c->tid);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
/*
* Flush cpu slab.
*
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
* Called from IPI handler with interrupts disabled.
*/
static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
{
struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (c->page)
flush_slab(s, c);
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
unfreeze_partials(s, c);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
static void flush_cpu_slab(void *d)
{
struct kmem_cache *s = d;
SLUB: Avoid page struct cacheline bouncing due to remote frees to cpu slab A remote free may access the same page struct that also contains the lockless freelist for the cpu slab. If objects have a short lifetime and are freed by a different processor then remote frees back to the slab from which we are currently allocating are frequent. The cacheline with the page struct needs to be repeately acquired in exclusive mode by both the allocating thread and the freeing thread. If this is frequent enough then performance will suffer because of cacheline bouncing. This patchset puts the lockless_freelist pointer in its own cacheline. In order to make that happen we introduce a per cpu structure called kmem_cache_cpu. Instead of keeping an array of pointers to page structs we now keep an array to a per cpu structure that--among other things--contains the pointer to the lockless freelist. The freeing thread can then keep possession of exclusive access to the page struct cacheline while the allocating thread keeps its exclusive access to the cacheline containing the per cpu structure. This works as long as the allocating cpu is able to service its request from the lockless freelist. If the lockless freelist runs empty then the allocating thread needs to acquire exclusive access to the cacheline with the page struct lock the slab. The allocating thread will then check if new objects were freed to the per cpu slab. If so it will keep the slab as the cpu slab and continue with the recently remote freed objects. So the allocating thread can take a series of just freed remote pages and dish them out again. Ideally allocations could be just recycling objects in the same slab this way which will lead to an ideal allocation / remote free pattern. The number of objects that can be handled in this way is limited by the capacity of one slab. Increasing slab size via slub_min_objects/ slub_max_order may increase the number of objects and therefore performance. If the allocating thread runs out of objects and finds that no objects were put back by the remote processor then it will retrieve a new slab (from the partial lists or from the page allocator) and start with a whole new set of objects while the remote thread may still be freeing objects to the old cpu slab. This may then repeat until the new slab is also exhausted. If remote freeing has freed objects in the earlier slab then that earlier slab will now be on the partial freelist and the allocating thread will pick that slab next for allocation. So the loop is extended. However, both threads need to take the list_lock to make the swizzling via the partial list happen. It is likely that this kind of scheme will keep the objects being passed around to a small set that can be kept in the cpu caches leading to increased performance. More code cleanups become possible: - Instead of passing a cpu we can now pass a kmem_cache_cpu structure around. Allows reducing the number of parameters to various functions. - Can define a new node_match() function for NUMA to encapsulate locality checks. Effect on allocations: Cachelines touched before this patch: Write: page cache struct and first cacheline of object Cachelines touched after this patch: Write: kmem_cache_cpu cacheline and first cacheline of object Read: page cache struct (but see later patch that avoids touching that cacheline) The handling when the lockless alloc list runs empty gets to be a bit more complicated since another cacheline has now to be written to. But that is halfway out of the hot path. Effect on freeing: Cachelines touched before this patch: Write: page_struct and first cacheline of object Cachelines touched after this patch depending on how we free: Write(to cpu_slab): kmem_cache_cpu struct and first cacheline of object Write(to other): page struct and first cacheline of object Read(to cpu_slab): page struct to id slab etc. (but see later patch that avoids touching the page struct on free) Read(to other): cpu local kmem_cache_cpu struct to verify its not the cpu slab. Summary: Pro: - Distinct cachelines so that concurrent remote frees and local allocs on a cpuslab can occur without cacheline bouncing. - Avoids potential bouncing cachelines because of neighboring per cpu pointer updates in kmem_cache's cpu_slab structure since it now grows to a cacheline (Therefore remove the comment that talks about that concern). Cons: - Freeing objects now requires the reading of one additional cacheline. That can be mitigated for some cases by the following patches but its not possible to completely eliminate these references. - Memory usage grows slightly. The size of each per cpu object is blown up from one word (pointing to the page_struct) to one cacheline with various data. So this is NR_CPUS*NR_SLABS*L1_BYTES more memory use. Lets say NR_SLABS is 100 and a cache line size of 128 then we have just increased SLAB metadata requirements by 12.8k per cpu. (Another later patch reduces these requirements) Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 16:26:05 +08:00
__flush_cpu_slab(s, smp_processor_id());
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
slub: only IPI CPUs that have per cpu obj to flush flush_all() is called for each kmem_cache_destroy(). So every cache being destroyed dynamically ends up sending an IPI to each CPU in the system, regardless if the cache has ever been used there. For example, if you close the Infinband ipath driver char device file, the close file ops calls kmem_cache_destroy(). So running some infiniband config tool on one a single CPU dedicated to system tasks might interrupt the rest of the 127 CPUs dedicated to some CPU intensive or latency sensitive task. I suspect there is a good chance that every line in the output of "git grep kmem_cache_destroy linux/ | grep '\->'" has a similar scenario. This patch attempts to rectify this issue by sending an IPI to flush the per cpu objects back to the free lists only to CPUs that seem to have such objects. The check which CPU to IPI is racy but we don't care since asking a CPU without per cpu objects to flush does no damage and as far as I can tell the flush_all by itself is racy against allocs on remote CPUs anyway, so if you required the flush_all to be determinstic, you had to arrange for locking regardless. Without this patch the following artificial test case: $ cd /sys/kernel/slab $ for DIR in *; do cat $DIR/alloc_calls > /dev/null; done produces 166 IPIs on an cpuset isolated CPU. With it it produces none. The code path of memory allocation failure for CPUMASK_OFFSTACK=y config was tested using fault injection framework. Signed-off-by: Gilad Ben-Yossef <gilad@benyossef.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Chris Metcalf <cmetcalf@tilera.com> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: Russell King <linux@arm.linux.org.uk> Cc: Pekka Enberg <penberg@kernel.org> Cc: Matt Mackall <mpm@selenic.com> Cc: Sasha Levin <levinsasha928@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Avi Kivity <avi@redhat.com> Cc: Michal Nazarewicz <mina86@mina86.org> Cc: Kosaki Motohiro <kosaki.motohiro@gmail.com> Cc: Milton Miller <miltonm@bga.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-03-29 05:42:44 +08:00
static bool has_cpu_slab(int cpu, void *info)
{
struct kmem_cache *s = info;
struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
return c->page || slub_percpu_partial(c);
slub: only IPI CPUs that have per cpu obj to flush flush_all() is called for each kmem_cache_destroy(). So every cache being destroyed dynamically ends up sending an IPI to each CPU in the system, regardless if the cache has ever been used there. For example, if you close the Infinband ipath driver char device file, the close file ops calls kmem_cache_destroy(). So running some infiniband config tool on one a single CPU dedicated to system tasks might interrupt the rest of the 127 CPUs dedicated to some CPU intensive or latency sensitive task. I suspect there is a good chance that every line in the output of "git grep kmem_cache_destroy linux/ | grep '\->'" has a similar scenario. This patch attempts to rectify this issue by sending an IPI to flush the per cpu objects back to the free lists only to CPUs that seem to have such objects. The check which CPU to IPI is racy but we don't care since asking a CPU without per cpu objects to flush does no damage and as far as I can tell the flush_all by itself is racy against allocs on remote CPUs anyway, so if you required the flush_all to be determinstic, you had to arrange for locking regardless. Without this patch the following artificial test case: $ cd /sys/kernel/slab $ for DIR in *; do cat $DIR/alloc_calls > /dev/null; done produces 166 IPIs on an cpuset isolated CPU. With it it produces none. The code path of memory allocation failure for CPUMASK_OFFSTACK=y config was tested using fault injection framework. Signed-off-by: Gilad Ben-Yossef <gilad@benyossef.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Chris Metcalf <cmetcalf@tilera.com> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: Russell King <linux@arm.linux.org.uk> Cc: Pekka Enberg <penberg@kernel.org> Cc: Matt Mackall <mpm@selenic.com> Cc: Sasha Levin <levinsasha928@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Mel Gorman <mel@csn.ul.ie> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Avi Kivity <avi@redhat.com> Cc: Michal Nazarewicz <mina86@mina86.org> Cc: Kosaki Motohiro <kosaki.motohiro@gmail.com> Cc: Milton Miller <miltonm@bga.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-03-29 05:42:44 +08:00
}
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
static void flush_all(struct kmem_cache *s)
{
on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
/*
* Use the cpu notifier to insure that the cpu slabs are flushed when
* necessary.
*/
static int slub_cpu_dead(unsigned int cpu)
{
struct kmem_cache *s;
unsigned long flags;
mutex_lock(&slab_mutex);
list_for_each_entry(s, &slab_caches, list) {
local_irq_save(flags);
__flush_cpu_slab(s, cpu);
local_irq_restore(flags);
}
mutex_unlock(&slab_mutex);
return 0;
}
SLUB: Avoid page struct cacheline bouncing due to remote frees to cpu slab A remote free may access the same page struct that also contains the lockless freelist for the cpu slab. If objects have a short lifetime and are freed by a different processor then remote frees back to the slab from which we are currently allocating are frequent. The cacheline with the page struct needs to be repeately acquired in exclusive mode by both the allocating thread and the freeing thread. If this is frequent enough then performance will suffer because of cacheline bouncing. This patchset puts the lockless_freelist pointer in its own cacheline. In order to make that happen we introduce a per cpu structure called kmem_cache_cpu. Instead of keeping an array of pointers to page structs we now keep an array to a per cpu structure that--among other things--contains the pointer to the lockless freelist. The freeing thread can then keep possession of exclusive access to the page struct cacheline while the allocating thread keeps its exclusive access to the cacheline containing the per cpu structure. This works as long as the allocating cpu is able to service its request from the lockless freelist. If the lockless freelist runs empty then the allocating thread needs to acquire exclusive access to the cacheline with the page struct lock the slab. The allocating thread will then check if new objects were freed to the per cpu slab. If so it will keep the slab as the cpu slab and continue with the recently remote freed objects. So the allocating thread can take a series of just freed remote pages and dish them out again. Ideally allocations could be just recycling objects in the same slab this way which will lead to an ideal allocation / remote free pattern. The number of objects that can be handled in this way is limited by the capacity of one slab. Increasing slab size via slub_min_objects/ slub_max_order may increase the number of objects and therefore performance. If the allocating thread runs out of objects and finds that no objects were put back by the remote processor then it will retrieve a new slab (from the partial lists or from the page allocator) and start with a whole new set of objects while the remote thread may still be freeing objects to the old cpu slab. This may then repeat until the new slab is also exhausted. If remote freeing has freed objects in the earlier slab then that earlier slab will now be on the partial freelist and the allocating thread will pick that slab next for allocation. So the loop is extended. However, both threads need to take the list_lock to make the swizzling via the partial list happen. It is likely that this kind of scheme will keep the objects being passed around to a small set that can be kept in the cpu caches leading to increased performance. More code cleanups become possible: - Instead of passing a cpu we can now pass a kmem_cache_cpu structure around. Allows reducing the number of parameters to various functions. - Can define a new node_match() function for NUMA to encapsulate locality checks. Effect on allocations: Cachelines touched before this patch: Write: page cache struct and first cacheline of object Cachelines touched after this patch: Write: kmem_cache_cpu cacheline and first cacheline of object Read: page cache struct (but see later patch that avoids touching that cacheline) The handling when the lockless alloc list runs empty gets to be a bit more complicated since another cacheline has now to be written to. But that is halfway out of the hot path. Effect on freeing: Cachelines touched before this patch: Write: page_struct and first cacheline of object Cachelines touched after this patch depending on how we free: Write(to cpu_slab): kmem_cache_cpu struct and first cacheline of object Write(to other): page struct and first cacheline of object Read(to cpu_slab): page struct to id slab etc. (but see later patch that avoids touching the page struct on free) Read(to other): cpu local kmem_cache_cpu struct to verify its not the cpu slab. Summary: Pro: - Distinct cachelines so that concurrent remote frees and local allocs on a cpuslab can occur without cacheline bouncing. - Avoids potential bouncing cachelines because of neighboring per cpu pointer updates in kmem_cache's cpu_slab structure since it now grows to a cacheline (Therefore remove the comment that talks about that concern). Cons: - Freeing objects now requires the reading of one additional cacheline. That can be mitigated for some cases by the following patches but its not possible to completely eliminate these references. - Memory usage grows slightly. The size of each per cpu object is blown up from one word (pointing to the page_struct) to one cacheline with various data. So this is NR_CPUS*NR_SLABS*L1_BYTES more memory use. Lets say NR_SLABS is 100 and a cache line size of 128 then we have just increased SLAB metadata requirements by 12.8k per cpu. (Another later patch reduces these requirements) Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 16:26:05 +08:00
/*
* Check if the objects in a per cpu structure fit numa
* locality expectations.
*/
static inline int node_match(struct page *page, int node)
SLUB: Avoid page struct cacheline bouncing due to remote frees to cpu slab A remote free may access the same page struct that also contains the lockless freelist for the cpu slab. If objects have a short lifetime and are freed by a different processor then remote frees back to the slab from which we are currently allocating are frequent. The cacheline with the page struct needs to be repeately acquired in exclusive mode by both the allocating thread and the freeing thread. If this is frequent enough then performance will suffer because of cacheline bouncing. This patchset puts the lockless_freelist pointer in its own cacheline. In order to make that happen we introduce a per cpu structure called kmem_cache_cpu. Instead of keeping an array of pointers to page structs we now keep an array to a per cpu structure that--among other things--contains the pointer to the lockless freelist. The freeing thread can then keep possession of exclusive access to the page struct cacheline while the allocating thread keeps its exclusive access to the cacheline containing the per cpu structure. This works as long as the allocating cpu is able to service its request from the lockless freelist. If the lockless freelist runs empty then the allocating thread needs to acquire exclusive access to the cacheline with the page struct lock the slab. The allocating thread will then check if new objects were freed to the per cpu slab. If so it will keep the slab as the cpu slab and continue with the recently remote freed objects. So the allocating thread can take a series of just freed remote pages and dish them out again. Ideally allocations could be just recycling objects in the same slab this way which will lead to an ideal allocation / remote free pattern. The number of objects that can be handled in this way is limited by the capacity of one slab. Increasing slab size via slub_min_objects/ slub_max_order may increase the number of objects and therefore performance. If the allocating thread runs out of objects and finds that no objects were put back by the remote processor then it will retrieve a new slab (from the partial lists or from the page allocator) and start with a whole new set of objects while the remote thread may still be freeing objects to the old cpu slab. This may then repeat until the new slab is also exhausted. If remote freeing has freed objects in the earlier slab then that earlier slab will now be on the partial freelist and the allocating thread will pick that slab next for allocation. So the loop is extended. However, both threads need to take the list_lock to make the swizzling via the partial list happen. It is likely that this kind of scheme will keep the objects being passed around to a small set that can be kept in the cpu caches leading to increased performance. More code cleanups become possible: - Instead of passing a cpu we can now pass a kmem_cache_cpu structure around. Allows reducing the number of parameters to various functions. - Can define a new node_match() function for NUMA to encapsulate locality checks. Effect on allocations: Cachelines touched before this patch: Write: page cache struct and first cacheline of object Cachelines touched after this patch: Write: kmem_cache_cpu cacheline and first cacheline of object Read: page cache struct (but see later patch that avoids touching that cacheline) The handling when the lockless alloc list runs empty gets to be a bit more complicated since another cacheline has now to be written to. But that is halfway out of the hot path. Effect on freeing: Cachelines touched before this patch: Write: page_struct and first cacheline of object Cachelines touched after this patch depending on how we free: Write(to cpu_slab): kmem_cache_cpu struct and first cacheline of object Write(to other): page struct and first cacheline of object Read(to cpu_slab): page struct to id slab etc. (but see later patch that avoids touching the page struct on free) Read(to other): cpu local kmem_cache_cpu struct to verify its not the cpu slab. Summary: Pro: - Distinct cachelines so that concurrent remote frees and local allocs on a cpuslab can occur without cacheline bouncing. - Avoids potential bouncing cachelines because of neighboring per cpu pointer updates in kmem_cache's cpu_slab structure since it now grows to a cacheline (Therefore remove the comment that talks about that concern). Cons: - Freeing objects now requires the reading of one additional cacheline. That can be mitigated for some cases by the following patches but its not possible to completely eliminate these references. - Memory usage grows slightly. The size of each per cpu object is blown up from one word (pointing to the page_struct) to one cacheline with various data. So this is NR_CPUS*NR_SLABS*L1_BYTES more memory use. Lets say NR_SLABS is 100 and a cache line size of 128 then we have just increased SLAB metadata requirements by 12.8k per cpu. (Another later patch reduces these requirements) Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 16:26:05 +08:00
{
#ifdef CONFIG_NUMA
if (node != NUMA_NO_NODE && page_to_nid(page) != node)
SLUB: Avoid page struct cacheline bouncing due to remote frees to cpu slab A remote free may access the same page struct that also contains the lockless freelist for the cpu slab. If objects have a short lifetime and are freed by a different processor then remote frees back to the slab from which we are currently allocating are frequent. The cacheline with the page struct needs to be repeately acquired in exclusive mode by both the allocating thread and the freeing thread. If this is frequent enough then performance will suffer because of cacheline bouncing. This patchset puts the lockless_freelist pointer in its own cacheline. In order to make that happen we introduce a per cpu structure called kmem_cache_cpu. Instead of keeping an array of pointers to page structs we now keep an array to a per cpu structure that--among other things--contains the pointer to the lockless freelist. The freeing thread can then keep possession of exclusive access to the page struct cacheline while the allocating thread keeps its exclusive access to the cacheline containing the per cpu structure. This works as long as the allocating cpu is able to service its request from the lockless freelist. If the lockless freelist runs empty then the allocating thread needs to acquire exclusive access to the cacheline with the page struct lock the slab. The allocating thread will then check if new objects were freed to the per cpu slab. If so it will keep the slab as the cpu slab and continue with the recently remote freed objects. So the allocating thread can take a series of just freed remote pages and dish them out again. Ideally allocations could be just recycling objects in the same slab this way which will lead to an ideal allocation / remote free pattern. The number of objects that can be handled in this way is limited by the capacity of one slab. Increasing slab size via slub_min_objects/ slub_max_order may increase the number of objects and therefore performance. If the allocating thread runs out of objects and finds that no objects were put back by the remote processor then it will retrieve a new slab (from the partial lists or from the page allocator) and start with a whole new set of objects while the remote thread may still be freeing objects to the old cpu slab. This may then repeat until the new slab is also exhausted. If remote freeing has freed objects in the earlier slab then that earlier slab will now be on the partial freelist and the allocating thread will pick that slab next for allocation. So the loop is extended. However, both threads need to take the list_lock to make the swizzling via the partial list happen. It is likely that this kind of scheme will keep the objects being passed around to a small set that can be kept in the cpu caches leading to increased performance. More code cleanups become possible: - Instead of passing a cpu we can now pass a kmem_cache_cpu structure around. Allows reducing the number of parameters to various functions. - Can define a new node_match() function for NUMA to encapsulate locality checks. Effect on allocations: Cachelines touched before this patch: Write: page cache struct and first cacheline of object Cachelines touched after this patch: Write: kmem_cache_cpu cacheline and first cacheline of object Read: page cache struct (but see later patch that avoids touching that cacheline) The handling when the lockless alloc list runs empty gets to be a bit more complicated since another cacheline has now to be written to. But that is halfway out of the hot path. Effect on freeing: Cachelines touched before this patch: Write: page_struct and first cacheline of object Cachelines touched after this patch depending on how we free: Write(to cpu_slab): kmem_cache_cpu struct and first cacheline of object Write(to other): page struct and first cacheline of object Read(to cpu_slab): page struct to id slab etc. (but see later patch that avoids touching the page struct on free) Read(to other): cpu local kmem_cache_cpu struct to verify its not the cpu slab. Summary: Pro: - Distinct cachelines so that concurrent remote frees and local allocs on a cpuslab can occur without cacheline bouncing. - Avoids potential bouncing cachelines because of neighboring per cpu pointer updates in kmem_cache's cpu_slab structure since it now grows to a cacheline (Therefore remove the comment that talks about that concern). Cons: - Freeing objects now requires the reading of one additional cacheline. That can be mitigated for some cases by the following patches but its not possible to completely eliminate these references. - Memory usage grows slightly. The size of each per cpu object is blown up from one word (pointing to the page_struct) to one cacheline with various data. So this is NR_CPUS*NR_SLABS*L1_BYTES more memory use. Lets say NR_SLABS is 100 and a cache line size of 128 then we have just increased SLAB metadata requirements by 12.8k per cpu. (Another later patch reduces these requirements) Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 16:26:05 +08:00
return 0;
#endif
return 1;
}
#ifdef CONFIG_SLUB_DEBUG
static int count_free(struct page *page)
{
return page->objects - page->inuse;
}
static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
{
return atomic_long_read(&n->total_objects);
}
#endif /* CONFIG_SLUB_DEBUG */
#if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
static unsigned long count_partial(struct kmem_cache_node *n,
int (*get_count)(struct page *))
{
unsigned long flags;
unsigned long x = 0;
struct page *page;
spin_lock_irqsave(&n->list_lock, flags);
list_for_each_entry(page, &n->partial, slab_list)
x += get_count(page);
spin_unlock_irqrestore(&n->list_lock, flags);
return x;
}
#endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
static noinline void
slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
{
#ifdef CONFIG_SLUB_DEBUG
static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
DEFAULT_RATELIMIT_BURST);
int node;
struct kmem_cache_node *n;
if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
return;
pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
nid, gfpflags, &gfpflags);
pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
s->name, s->object_size, s->size, oo_order(s->oo),
oo_order(s->min));
if (oo_order(s->min) > get_order(s->object_size))
pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
s->name);
for_each_kmem_cache_node(s, node, n) {
unsigned long nr_slabs;
unsigned long nr_objs;
unsigned long nr_free;
nr_free = count_partial(n, count_free);
nr_slabs = node_nr_slabs(n);
nr_objs = node_nr_objs(n);
pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
node, nr_slabs, nr_objs, nr_free);
}
#endif
}
static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
int node, struct kmem_cache_cpu **pc)
{
void *freelist;
struct kmem_cache_cpu *c = *pc;
struct page *page;
WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
freelist = get_partial(s, flags, node, c);
if (freelist)
return freelist;
page = new_slab(s, flags, node);
if (page) {
c = raw_cpu_ptr(s->cpu_slab);
if (c->page)
flush_slab(s, c);
/*
* No other reference to the page yet so we can
* muck around with it freely without cmpxchg
*/
freelist = page->freelist;
page->freelist = NULL;
stat(s, ALLOC_SLAB);
c->page = page;
*pc = c;
}
return freelist;
}
mm: sl[au]b: add knowledge of PFMEMALLOC reserve pages When a user or administrator requires swap for their application, they create a swap partition and file, format it with mkswap and activate it with swapon. Swap over the network is considered as an option in diskless systems. The two likely scenarios are when blade servers are used as part of a cluster where the form factor or maintenance costs do not allow the use of disks and thin clients. The Linux Terminal Server Project recommends the use of the Network Block Device (NBD) for swap according to the manual at https://sourceforge.net/projects/ltsp/files/Docs-Admin-Guide/LTSPManual.pdf/download There is also documentation and tutorials on how to setup swap over NBD at places like https://help.ubuntu.com/community/UbuntuLTSP/EnableNBDSWAP The nbd-client also documents the use of NBD as swap. Despite this, the fact is that a machine using NBD for swap can deadlock within minutes if swap is used intensively. This patch series addresses the problem. The core issue is that network block devices do not use mempools like normal block devices do. As the host cannot control where they receive packets from, they cannot reliably work out in advance how much memory they might need. Some years ago, Peter Zijlstra developed a series of patches that supported swap over an NFS that at least one distribution is carrying within their kernels. This patch series borrows very heavily from Peter's work to support swapping over NBD as a pre-requisite to supporting swap-over-NFS. The bulk of the complexity is concerned with preserving memory that is allocated from the PFMEMALLOC reserves for use by the network layer which is needed for both NBD and NFS. Patch 1 adds knowledge of the PFMEMALLOC reserves to SLAB and SLUB to preserve access to pages allocated under low memory situations to callers that are freeing memory. Patch 2 optimises the SLUB fast path to avoid pfmemalloc checks Patch 3 introduces __GFP_MEMALLOC to allow access to the PFMEMALLOC reserves without setting PFMEMALLOC. Patch 4 opens the possibility for softirqs to use PFMEMALLOC reserves for later use by network packet processing. Patch 5 only sets page->pfmemalloc when ALLOC_NO_WATERMARKS was required Patch 6 ignores memory policies when ALLOC_NO_WATERMARKS is set. Patches 7-12 allows network processing to use PFMEMALLOC reserves when the socket has been marked as being used by the VM to clean pages. If packets are received and stored in pages that were allocated under low-memory situations and are unrelated to the VM, the packets are dropped. Patch 11 reintroduces __skb_alloc_page which the networking folk may object to but is needed in some cases to propogate pfmemalloc from a newly allocated page to an skb. If there is a strong objection, this patch can be dropped with the impact being that swap-over-network will be slower in some cases but it should not fail. Patch 13 is a micro-optimisation to avoid a function call in the common case. Patch 14 tags NBD sockets as being SOCK_MEMALLOC so they can use PFMEMALLOC if necessary. Patch 15 notes that it is still possible for the PFMEMALLOC reserve to be depleted. To prevent this, direct reclaimers get throttled on a waitqueue if 50% of the PFMEMALLOC reserves are depleted. It is expected that kswapd and the direct reclaimers already running will clean enough pages for the low watermark to be reached and the throttled processes are woken up. Patch 16 adds a statistic to track how often processes get throttled Some basic performance testing was run using kernel builds, netperf on loopback for UDP and TCP, hackbench (pipes and sockets), iozone and sysbench. Each of them were expected to use the sl*b allocators reasonably heavily but there did not appear to be significant performance variances. For testing swap-over-NBD, a machine was booted with 2G of RAM with a swapfile backed by NBD. 8*NUM_CPU processes were started that create anonymous memory mappings and read them linearly in a loop. The total size of the mappings were 4*PHYSICAL_MEMORY to use swap heavily under memory pressure. Without the patches and using SLUB, the machine locks up within minutes and runs to completion with them applied. With SLAB, the story is different as an unpatched kernel run to completion. However, the patched kernel completed the test 45% faster. MICRO 3.5.0-rc2 3.5.0-rc2 vanilla swapnbd Unrecognised test vmscan-anon-mmap-write MMTests Statistics: duration Sys Time Running Test (seconds) 197.80 173.07 User+Sys Time Running Test (seconds) 206.96 182.03 Total Elapsed Time (seconds) 3240.70 1762.09 This patch: mm: sl[au]b: add knowledge of PFMEMALLOC reserve pages Allocations of pages below the min watermark run a risk of the machine hanging due to a lack of memory. To prevent this, only callers who have PF_MEMALLOC or TIF_MEMDIE set and are not processing an interrupt are allowed to allocate with ALLOC_NO_WATERMARKS. Once they are allocated to a slab though, nothing prevents other callers consuming free objects within those slabs. This patch limits access to slab pages that were alloced from the PFMEMALLOC reserves. When this patch is applied, pages allocated from below the low watermark are returned with page->pfmemalloc set and it is up to the caller to determine how the page should be protected. SLAB restricts access to any page with page->pfmemalloc set to callers which are known to able to access the PFMEMALLOC reserve. If one is not available, an attempt is made to allocate a new page rather than use a reserve. SLUB is a bit more relaxed in that it only records if the current per-CPU page was allocated from PFMEMALLOC reserve and uses another partial slab if the caller does not have the necessary GFP or process flags. This was found to be sufficient in tests to avoid hangs due to SLUB generally maintaining smaller lists than SLAB. In low-memory conditions it does mean that !PFMEMALLOC allocators can fail a slab allocation even though free objects are available because they are being preserved for callers that are freeing pages. [a.p.zijlstra@chello.nl: Original implementation] [sebastian@breakpoint.cc: Correct order of page flag clearing] Signed-off-by: Mel Gorman <mgorman@suse.de> Cc: David Miller <davem@davemloft.net> Cc: Neil Brown <neilb@suse.de> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Christie <michaelc@cs.wisc.edu> Cc: Eric B Munson <emunson@mgebm.net> Cc: Eric Dumazet <eric.dumazet@gmail.com> Cc: Sebastian Andrzej Siewior <sebastian@breakpoint.cc> Cc: Mel Gorman <mgorman@suse.de> Cc: Christoph Lameter <cl@linux.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-08-01 07:43:58 +08:00
static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
{
if (unlikely(PageSlabPfmemalloc(page)))
return gfp_pfmemalloc_allowed(gfpflags);
return true;
}
/*
* Check the page->freelist of a page and either transfer the freelist to the
* per cpu freelist or deactivate the page.
*
* The page is still frozen if the return value is not NULL.
*
* If this function returns NULL then the page has been unfrozen.
*
* This function must be called with interrupt disabled.
*/
static inline void *get_freelist(struct kmem_cache *s, struct page *page)
{
struct page new;
unsigned long counters;
void *freelist;
do {
freelist = page->freelist;
counters = page->counters;
new.counters = counters;
VM_BUG_ON(!new.frozen);
new.inuse = page->objects;
new.frozen = freelist != NULL;
} while (!__cmpxchg_double_slab(s, page,
freelist, counters,
NULL, new.counters,
"get_freelist"));
return freelist;
}
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/*
slub: support concurrent local and remote frees and allocs on a slab Avoid atomic overhead in slab_alloc and slab_free SLUB needs to use the slab_lock for the per cpu slabs to synchronize with potential kfree operations. This patch avoids that need by moving all free objects onto a lockless_freelist. The regular freelist continues to exist and will be used to free objects. So while we consume the lockless_freelist the regular freelist may build up objects. If we are out of objects on the lockless_freelist then we may check the regular freelist. If it has objects then we move those over to the lockless_freelist and do this again. There is a significant savings in terms of atomic operations that have to be performed. We can even free directly to the lockless_freelist if we know that we are running on the same processor. So this speeds up short lived objects. They may be allocated and freed without taking the slab_lock. This is particular good for netperf. In order to maximize the effect of the new faster hotpath we extract the hottest performance pieces into inlined functions. These are then inlined into kmem_cache_alloc and kmem_cache_free. So hotpath allocation and freeing no longer requires a subroutine call within SLUB. [I am not sure that it is worth doing this because it changes the easy to read structure of slub just to reduce atomic ops. However, there is someone out there with a benchmark on 4 way and 8 way processor systems that seems to show a 5% regression vs. Slab. Seems that the regression is due to increased atomic operations use vs. SLAB in SLUB). I wonder if this is applicable or discernable at all in a real workload?] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-10 18:15:16 +08:00
* Slow path. The lockless freelist is empty or we need to perform
* debugging duties.
*
* Processing is still very fast if new objects have been freed to the
* regular freelist. In that case we simply take over the regular freelist
* as the lockless freelist and zap the regular freelist.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*
slub: support concurrent local and remote frees and allocs on a slab Avoid atomic overhead in slab_alloc and slab_free SLUB needs to use the slab_lock for the per cpu slabs to synchronize with potential kfree operations. This patch avoids that need by moving all free objects onto a lockless_freelist. The regular freelist continues to exist and will be used to free objects. So while we consume the lockless_freelist the regular freelist may build up objects. If we are out of objects on the lockless_freelist then we may check the regular freelist. If it has objects then we move those over to the lockless_freelist and do this again. There is a significant savings in terms of atomic operations that have to be performed. We can even free directly to the lockless_freelist if we know that we are running on the same processor. So this speeds up short lived objects. They may be allocated and freed without taking the slab_lock. This is particular good for netperf. In order to maximize the effect of the new faster hotpath we extract the hottest performance pieces into inlined functions. These are then inlined into kmem_cache_alloc and kmem_cache_free. So hotpath allocation and freeing no longer requires a subroutine call within SLUB. [I am not sure that it is worth doing this because it changes the easy to read structure of slub just to reduce atomic ops. However, there is someone out there with a benchmark on 4 way and 8 way processor systems that seems to show a 5% regression vs. Slab. Seems that the regression is due to increased atomic operations use vs. SLAB in SLUB). I wonder if this is applicable or discernable at all in a real workload?] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-10 18:15:16 +08:00
* If that is not working then we fall back to the partial lists. We take the
* first element of the freelist as the object to allocate now and move the
* rest of the freelist to the lockless freelist.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*
slub: support concurrent local and remote frees and allocs on a slab Avoid atomic overhead in slab_alloc and slab_free SLUB needs to use the slab_lock for the per cpu slabs to synchronize with potential kfree operations. This patch avoids that need by moving all free objects onto a lockless_freelist. The regular freelist continues to exist and will be used to free objects. So while we consume the lockless_freelist the regular freelist may build up objects. If we are out of objects on the lockless_freelist then we may check the regular freelist. If it has objects then we move those over to the lockless_freelist and do this again. There is a significant savings in terms of atomic operations that have to be performed. We can even free directly to the lockless_freelist if we know that we are running on the same processor. So this speeds up short lived objects. They may be allocated and freed without taking the slab_lock. This is particular good for netperf. In order to maximize the effect of the new faster hotpath we extract the hottest performance pieces into inlined functions. These are then inlined into kmem_cache_alloc and kmem_cache_free. So hotpath allocation and freeing no longer requires a subroutine call within SLUB. [I am not sure that it is worth doing this because it changes the easy to read structure of slub just to reduce atomic ops. However, there is someone out there with a benchmark on 4 way and 8 way processor systems that seems to show a 5% regression vs. Slab. Seems that the regression is due to increased atomic operations use vs. SLAB in SLUB). I wonder if this is applicable or discernable at all in a real workload?] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-10 18:15:16 +08:00
* And if we were unable to get a new slab from the partial slab lists then
* we need to allocate a new slab. This is the slowest path since it involves
* a call to the page allocator and the setup of a new slab.
*
* Version of __slab_alloc to use when we know that interrupts are
* already disabled (which is the case for bulk allocation).
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*/
static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
unsigned long addr, struct kmem_cache_cpu *c)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
{
void *freelist;
struct page *page;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
stat(s, ALLOC_SLOWPATH);
page = c->page;
mm, slub: prevent kmalloc_node crashes and memory leaks Sachin reports [1] a crash in SLUB __slab_alloc(): BUG: Kernel NULL pointer dereference on read at 0x000073b0 Faulting instruction address: 0xc0000000003d55f4 Oops: Kernel access of bad area, sig: 11 [#1] LE PAGE_SIZE=64K MMU=Hash SMP NR_CPUS=2048 NUMA pSeries Modules linked in: CPU: 19 PID: 1 Comm: systemd Not tainted 5.6.0-rc2-next-20200218-autotest #1 NIP: c0000000003d55f4 LR: c0000000003d5b94 CTR: 0000000000000000 REGS: c0000008b37836d0 TRAP: 0300 Not tainted (5.6.0-rc2-next-20200218-autotest) MSR: 8000000000009033 <SF,EE,ME,IR,DR,RI,LE> CR: 24004844 XER: 00000000 CFAR: c00000000000dec4 DAR: 00000000000073b0 DSISR: 40000000 IRQMASK: 1 GPR00: c0000000003d5b94 c0000008b3783960 c00000000155d400 c0000008b301f500 GPR04: 0000000000000dc0 0000000000000002 c0000000003443d8 c0000008bb398620 GPR08: 00000008ba2f0000 0000000000000001 0000000000000000 0000000000000000 GPR12: 0000000024004844 c00000001ec52a00 0000000000000000 0000000000000000 GPR16: c0000008a1b20048 c000000001595898 c000000001750c18 0000000000000002 GPR20: c000000001750c28 c000000001624470 0000000fffffffe0 5deadbeef0000122 GPR24: 0000000000000001 0000000000000dc0 0000000000000002 c0000000003443d8 GPR28: c0000008b301f500 c0000008bb398620 0000000000000000 c00c000002287180 NIP ___slab_alloc+0x1f4/0x760 LR __slab_alloc+0x34/0x60 Call Trace: ___slab_alloc+0x334/0x760 (unreliable) __slab_alloc+0x34/0x60 __kmalloc_node+0x110/0x490 kvmalloc_node+0x58/0x110 mem_cgroup_css_online+0x108/0x270 online_css+0x48/0xd0 cgroup_apply_control_enable+0x2ec/0x4d0 cgroup_mkdir+0x228/0x5f0 kernfs_iop_mkdir+0x90/0xf0 vfs_mkdir+0x110/0x230 do_mkdirat+0xb0/0x1a0 system_call+0x5c/0x68 This is a PowerPC platform with following NUMA topology: available: 2 nodes (0-1) node 0 cpus: node 0 size: 0 MB node 0 free: 0 MB node 1 cpus: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 node 1 size: 35247 MB node 1 free: 30907 MB node distances: node 0 1 0: 10 40 1: 40 10 possible numa nodes: 0-31 This only happens with a mmotm patch "mm/memcontrol.c: allocate shrinker_map on appropriate NUMA node" [2] which effectively calls kmalloc_node for each possible node. SLUB however only allocates kmem_cache_node on online N_NORMAL_MEMORY nodes, and relies on node_to_mem_node to return such valid node for other nodes since commit a561ce00b09e ("slub: fall back to node_to_mem_node() node if allocating on memoryless node"). This is however not true in this configuration where the _node_numa_mem_ array is not initialized for nodes 0 and 2-31, thus it contains zeroes and get_partial() ends up accessing non-allocated kmem_cache_node. A related issue was reported by Bharata (originally by Ramachandran) [3] where a similar PowerPC configuration, but with mainline kernel without patch [2] ends up allocating large amounts of pages by kmalloc-1k kmalloc-512. This seems to have the same underlying issue with node_to_mem_node() not behaving as expected, and might probably also lead to an infinite loop with CONFIG_SLUB_CPU_PARTIAL [4]. This patch should fix both issues by not relying on node_to_mem_node() anymore and instead simply falling back to NUMA_NO_NODE, when kmalloc_node(node) is attempted for a node that's not online, or has no usable memory. The "usable memory" condition is also changed from node_present_pages() to N_NORMAL_MEMORY node state, as that is exactly the condition that SLUB uses to allocate kmem_cache_node structures. The check in get_partial() is removed completely, as the checks in ___slab_alloc() are now sufficient to prevent get_partial() being reached with an invalid node. [1] https://lore.kernel.org/linux-next/3381CD91-AB3D-4773-BA04-E7A072A63968@linux.vnet.ibm.com/ [2] https://lore.kernel.org/linux-mm/fff0e636-4c36-ed10-281c-8cdb0687c839@virtuozzo.com/ [3] https://lore.kernel.org/linux-mm/20200317092624.GB22538@in.ibm.com/ [4] https://lore.kernel.org/linux-mm/088b5996-faae-8a56-ef9c-5b567125ae54@suse.cz/ Fixes: a561ce00b09e ("slub: fall back to node_to_mem_node() node if allocating on memoryless node") Reported-by: Sachin Sant <sachinp@linux.vnet.ibm.com> Reported-by: PUVICHAKRAVARTHY RAMACHANDRAN <puvichakravarthy@in.ibm.com> Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Tested-by: Sachin Sant <sachinp@linux.vnet.ibm.com> Tested-by: Bharata B Rao <bharata@linux.ibm.com> Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Michael Ellerman <mpe@ellerman.id.au> Cc: Michal Hocko <mhocko@kernel.org> Cc: Christopher Lameter <cl@linux.com> Cc: linuxppc-dev@lists.ozlabs.org Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Kirill Tkhai <ktkhai@virtuozzo.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Nathan Lynch <nathanl@linux.ibm.com> Cc: <stable@vger.kernel.org> Link: http://lkml.kernel.org/r/20200320115533.9604-1-vbabka@suse.cz Debugged-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-03-22 09:22:37 +08:00
if (!page) {
/*
* if the node is not online or has no normal memory, just
* ignore the node constraint
*/
if (unlikely(node != NUMA_NO_NODE &&
mm, slab, slub: stop taking memory hotplug lock Since commit 03afc0e25f7f ("slab: get_online_mems for kmem_cache_{create,destroy,shrink}") we are taking memory hotplug lock for SLAB and SLUB when creating, destroying or shrinking a cache. It is quite a heavy lock and it's best to avoid it if possible, as we had several issues with lockdep complaining about ordering in the past, see e.g. e4f8e513c3d3 ("mm/slub: fix a deadlock in show_slab_objects()"). The problem scenario in 03afc0e25f7f (solved by the memory hotplug lock) can be summarized as follows: while there's slab_mutex synchronizing new kmem cache creation and SLUB's MEM_GOING_ONLINE callback slab_mem_going_online_callback(), we may miss creation of kmem_cache_node for the hotplugged node in the new kmem cache, because the hotplug callback doesn't yet see the new cache, and cache creation in init_kmem_cache_nodes() only inits kmem_cache_node for nodes in the N_NORMAL_MEMORY nodemask, which however may not yet include the new node, as that happens only later after the MEM_GOING_ONLINE callback. Instead of using get/put_online_mems(), the problem can be solved by SLUB maintaining its own nodemask of nodes for which it has allocated the per-node kmem_cache_node structures. This nodemask would generally mirror the N_NORMAL_MEMORY nodemask, but would be updated only in under SLUB's control in its memory hotplug callbacks under the slab_mutex. This patch adds such nodemask and its handling. Commit 03afc0e25f7f mentiones "issues like [the one above]", but there don't appear to be further issues. All the paths (shared for SLAB and SLUB) taking the memory hotplug locks are also taking the slab_mutex, except kmem_cache_shrink() where 03afc0e25f7f replaced slab_mutex with get/put_online_mems(). We however cannot simply restore slab_mutex in kmem_cache_shrink(), as SLUB can enters the function from a write to sysfs 'shrink' file, thus holding kernfs lock, and in kmem_cache_create() the kernfs lock is nested within slab_mutex. But on closer inspection we don't actually need to protect kmem_cache_shrink() from hotplug callbacks: While SLUB's __kmem_cache_shrink() does for_each_kmem_cache_node(), missing a new node added in parallel hotplug is not fatal, and parallel hotremove does not free kmem_cache_node's anymore after the previous patch, so use-after free cannot happen. The per-node shrinking itself is protected by n->list_lock. Same is true for SLAB, and SLOB is no-op. SLAB also doesn't need the memory hotplug locking, which it only gained by 03afc0e25f7f through the shared paths in slab_common.c. Its memory hotplug callbacks are also protected by slab_mutex against races with these paths. The problem of SLUB relying on N_NORMAL_MEMORY doesn't apply to SLAB, as its setup_kmem_cache_nodes relies on N_ONLINE, and the new node is already set there during the MEM_GOING_ONLINE callback, so no special care is needed for SLAB. As such, this patch removes all get/put_online_mems() usage by the slab subsystem. Link: https://lkml.kernel.org/r/20210113131634.3671-3-vbabka@suse.cz Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Cc: Christoph Lameter <cl@linux.com> Cc: David Hildenbrand <david@redhat.com> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: Qian Cai <cai@redhat.com> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-25 04:01:12 +08:00
!node_isset(node, slab_nodes)))
mm, slub: prevent kmalloc_node crashes and memory leaks Sachin reports [1] a crash in SLUB __slab_alloc(): BUG: Kernel NULL pointer dereference on read at 0x000073b0 Faulting instruction address: 0xc0000000003d55f4 Oops: Kernel access of bad area, sig: 11 [#1] LE PAGE_SIZE=64K MMU=Hash SMP NR_CPUS=2048 NUMA pSeries Modules linked in: CPU: 19 PID: 1 Comm: systemd Not tainted 5.6.0-rc2-next-20200218-autotest #1 NIP: c0000000003d55f4 LR: c0000000003d5b94 CTR: 0000000000000000 REGS: c0000008b37836d0 TRAP: 0300 Not tainted (5.6.0-rc2-next-20200218-autotest) MSR: 8000000000009033 <SF,EE,ME,IR,DR,RI,LE> CR: 24004844 XER: 00000000 CFAR: c00000000000dec4 DAR: 00000000000073b0 DSISR: 40000000 IRQMASK: 1 GPR00: c0000000003d5b94 c0000008b3783960 c00000000155d400 c0000008b301f500 GPR04: 0000000000000dc0 0000000000000002 c0000000003443d8 c0000008bb398620 GPR08: 00000008ba2f0000 0000000000000001 0000000000000000 0000000000000000 GPR12: 0000000024004844 c00000001ec52a00 0000000000000000 0000000000000000 GPR16: c0000008a1b20048 c000000001595898 c000000001750c18 0000000000000002 GPR20: c000000001750c28 c000000001624470 0000000fffffffe0 5deadbeef0000122 GPR24: 0000000000000001 0000000000000dc0 0000000000000002 c0000000003443d8 GPR28: c0000008b301f500 c0000008bb398620 0000000000000000 c00c000002287180 NIP ___slab_alloc+0x1f4/0x760 LR __slab_alloc+0x34/0x60 Call Trace: ___slab_alloc+0x334/0x760 (unreliable) __slab_alloc+0x34/0x60 __kmalloc_node+0x110/0x490 kvmalloc_node+0x58/0x110 mem_cgroup_css_online+0x108/0x270 online_css+0x48/0xd0 cgroup_apply_control_enable+0x2ec/0x4d0 cgroup_mkdir+0x228/0x5f0 kernfs_iop_mkdir+0x90/0xf0 vfs_mkdir+0x110/0x230 do_mkdirat+0xb0/0x1a0 system_call+0x5c/0x68 This is a PowerPC platform with following NUMA topology: available: 2 nodes (0-1) node 0 cpus: node 0 size: 0 MB node 0 free: 0 MB node 1 cpus: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 node 1 size: 35247 MB node 1 free: 30907 MB node distances: node 0 1 0: 10 40 1: 40 10 possible numa nodes: 0-31 This only happens with a mmotm patch "mm/memcontrol.c: allocate shrinker_map on appropriate NUMA node" [2] which effectively calls kmalloc_node for each possible node. SLUB however only allocates kmem_cache_node on online N_NORMAL_MEMORY nodes, and relies on node_to_mem_node to return such valid node for other nodes since commit a561ce00b09e ("slub: fall back to node_to_mem_node() node if allocating on memoryless node"). This is however not true in this configuration where the _node_numa_mem_ array is not initialized for nodes 0 and 2-31, thus it contains zeroes and get_partial() ends up accessing non-allocated kmem_cache_node. A related issue was reported by Bharata (originally by Ramachandran) [3] where a similar PowerPC configuration, but with mainline kernel without patch [2] ends up allocating large amounts of pages by kmalloc-1k kmalloc-512. This seems to have the same underlying issue with node_to_mem_node() not behaving as expected, and might probably also lead to an infinite loop with CONFIG_SLUB_CPU_PARTIAL [4]. This patch should fix both issues by not relying on node_to_mem_node() anymore and instead simply falling back to NUMA_NO_NODE, when kmalloc_node(node) is attempted for a node that's not online, or has no usable memory. The "usable memory" condition is also changed from node_present_pages() to N_NORMAL_MEMORY node state, as that is exactly the condition that SLUB uses to allocate kmem_cache_node structures. The check in get_partial() is removed completely, as the checks in ___slab_alloc() are now sufficient to prevent get_partial() being reached with an invalid node. [1] https://lore.kernel.org/linux-next/3381CD91-AB3D-4773-BA04-E7A072A63968@linux.vnet.ibm.com/ [2] https://lore.kernel.org/linux-mm/fff0e636-4c36-ed10-281c-8cdb0687c839@virtuozzo.com/ [3] https://lore.kernel.org/linux-mm/20200317092624.GB22538@in.ibm.com/ [4] https://lore.kernel.org/linux-mm/088b5996-faae-8a56-ef9c-5b567125ae54@suse.cz/ Fixes: a561ce00b09e ("slub: fall back to node_to_mem_node() node if allocating on memoryless node") Reported-by: Sachin Sant <sachinp@linux.vnet.ibm.com> Reported-by: PUVICHAKRAVARTHY RAMACHANDRAN <puvichakravarthy@in.ibm.com> Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Tested-by: Sachin Sant <sachinp@linux.vnet.ibm.com> Tested-by: Bharata B Rao <bharata@linux.ibm.com> Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Michael Ellerman <mpe@ellerman.id.au> Cc: Michal Hocko <mhocko@kernel.org> Cc: Christopher Lameter <cl@linux.com> Cc: linuxppc-dev@lists.ozlabs.org Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Kirill Tkhai <ktkhai@virtuozzo.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Nathan Lynch <nathanl@linux.ibm.com> Cc: <stable@vger.kernel.org> Link: http://lkml.kernel.org/r/20200320115533.9604-1-vbabka@suse.cz Debugged-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-03-22 09:22:37 +08:00
node = NUMA_NO_NODE;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
goto new_slab;
mm, slub: prevent kmalloc_node crashes and memory leaks Sachin reports [1] a crash in SLUB __slab_alloc(): BUG: Kernel NULL pointer dereference on read at 0x000073b0 Faulting instruction address: 0xc0000000003d55f4 Oops: Kernel access of bad area, sig: 11 [#1] LE PAGE_SIZE=64K MMU=Hash SMP NR_CPUS=2048 NUMA pSeries Modules linked in: CPU: 19 PID: 1 Comm: systemd Not tainted 5.6.0-rc2-next-20200218-autotest #1 NIP: c0000000003d55f4 LR: c0000000003d5b94 CTR: 0000000000000000 REGS: c0000008b37836d0 TRAP: 0300 Not tainted (5.6.0-rc2-next-20200218-autotest) MSR: 8000000000009033 <SF,EE,ME,IR,DR,RI,LE> CR: 24004844 XER: 00000000 CFAR: c00000000000dec4 DAR: 00000000000073b0 DSISR: 40000000 IRQMASK: 1 GPR00: c0000000003d5b94 c0000008b3783960 c00000000155d400 c0000008b301f500 GPR04: 0000000000000dc0 0000000000000002 c0000000003443d8 c0000008bb398620 GPR08: 00000008ba2f0000 0000000000000001 0000000000000000 0000000000000000 GPR12: 0000000024004844 c00000001ec52a00 0000000000000000 0000000000000000 GPR16: c0000008a1b20048 c000000001595898 c000000001750c18 0000000000000002 GPR20: c000000001750c28 c000000001624470 0000000fffffffe0 5deadbeef0000122 GPR24: 0000000000000001 0000000000000dc0 0000000000000002 c0000000003443d8 GPR28: c0000008b301f500 c0000008bb398620 0000000000000000 c00c000002287180 NIP ___slab_alloc+0x1f4/0x760 LR __slab_alloc+0x34/0x60 Call Trace: ___slab_alloc+0x334/0x760 (unreliable) __slab_alloc+0x34/0x60 __kmalloc_node+0x110/0x490 kvmalloc_node+0x58/0x110 mem_cgroup_css_online+0x108/0x270 online_css+0x48/0xd0 cgroup_apply_control_enable+0x2ec/0x4d0 cgroup_mkdir+0x228/0x5f0 kernfs_iop_mkdir+0x90/0xf0 vfs_mkdir+0x110/0x230 do_mkdirat+0xb0/0x1a0 system_call+0x5c/0x68 This is a PowerPC platform with following NUMA topology: available: 2 nodes (0-1) node 0 cpus: node 0 size: 0 MB node 0 free: 0 MB node 1 cpus: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 node 1 size: 35247 MB node 1 free: 30907 MB node distances: node 0 1 0: 10 40 1: 40 10 possible numa nodes: 0-31 This only happens with a mmotm patch "mm/memcontrol.c: allocate shrinker_map on appropriate NUMA node" [2] which effectively calls kmalloc_node for each possible node. SLUB however only allocates kmem_cache_node on online N_NORMAL_MEMORY nodes, and relies on node_to_mem_node to return such valid node for other nodes since commit a561ce00b09e ("slub: fall back to node_to_mem_node() node if allocating on memoryless node"). This is however not true in this configuration where the _node_numa_mem_ array is not initialized for nodes 0 and 2-31, thus it contains zeroes and get_partial() ends up accessing non-allocated kmem_cache_node. A related issue was reported by Bharata (originally by Ramachandran) [3] where a similar PowerPC configuration, but with mainline kernel without patch [2] ends up allocating large amounts of pages by kmalloc-1k kmalloc-512. This seems to have the same underlying issue with node_to_mem_node() not behaving as expected, and might probably also lead to an infinite loop with CONFIG_SLUB_CPU_PARTIAL [4]. This patch should fix both issues by not relying on node_to_mem_node() anymore and instead simply falling back to NUMA_NO_NODE, when kmalloc_node(node) is attempted for a node that's not online, or has no usable memory. The "usable memory" condition is also changed from node_present_pages() to N_NORMAL_MEMORY node state, as that is exactly the condition that SLUB uses to allocate kmem_cache_node structures. The check in get_partial() is removed completely, as the checks in ___slab_alloc() are now sufficient to prevent get_partial() being reached with an invalid node. [1] https://lore.kernel.org/linux-next/3381CD91-AB3D-4773-BA04-E7A072A63968@linux.vnet.ibm.com/ [2] https://lore.kernel.org/linux-mm/fff0e636-4c36-ed10-281c-8cdb0687c839@virtuozzo.com/ [3] https://lore.kernel.org/linux-mm/20200317092624.GB22538@in.ibm.com/ [4] https://lore.kernel.org/linux-mm/088b5996-faae-8a56-ef9c-5b567125ae54@suse.cz/ Fixes: a561ce00b09e ("slub: fall back to node_to_mem_node() node if allocating on memoryless node") Reported-by: Sachin Sant <sachinp@linux.vnet.ibm.com> Reported-by: PUVICHAKRAVARTHY RAMACHANDRAN <puvichakravarthy@in.ibm.com> Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Tested-by: Sachin Sant <sachinp@linux.vnet.ibm.com> Tested-by: Bharata B Rao <bharata@linux.ibm.com> Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Michael Ellerman <mpe@ellerman.id.au> Cc: Michal Hocko <mhocko@kernel.org> Cc: Christopher Lameter <cl@linux.com> Cc: linuxppc-dev@lists.ozlabs.org Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Kirill Tkhai <ktkhai@virtuozzo.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Nathan Lynch <nathanl@linux.ibm.com> Cc: <stable@vger.kernel.org> Link: http://lkml.kernel.org/r/20200320115533.9604-1-vbabka@suse.cz Debugged-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-03-22 09:22:37 +08:00
}
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
redo:
if (unlikely(!node_match(page, node))) {
mm, slub: prevent kmalloc_node crashes and memory leaks Sachin reports [1] a crash in SLUB __slab_alloc(): BUG: Kernel NULL pointer dereference on read at 0x000073b0 Faulting instruction address: 0xc0000000003d55f4 Oops: Kernel access of bad area, sig: 11 [#1] LE PAGE_SIZE=64K MMU=Hash SMP NR_CPUS=2048 NUMA pSeries Modules linked in: CPU: 19 PID: 1 Comm: systemd Not tainted 5.6.0-rc2-next-20200218-autotest #1 NIP: c0000000003d55f4 LR: c0000000003d5b94 CTR: 0000000000000000 REGS: c0000008b37836d0 TRAP: 0300 Not tainted (5.6.0-rc2-next-20200218-autotest) MSR: 8000000000009033 <SF,EE,ME,IR,DR,RI,LE> CR: 24004844 XER: 00000000 CFAR: c00000000000dec4 DAR: 00000000000073b0 DSISR: 40000000 IRQMASK: 1 GPR00: c0000000003d5b94 c0000008b3783960 c00000000155d400 c0000008b301f500 GPR04: 0000000000000dc0 0000000000000002 c0000000003443d8 c0000008bb398620 GPR08: 00000008ba2f0000 0000000000000001 0000000000000000 0000000000000000 GPR12: 0000000024004844 c00000001ec52a00 0000000000000000 0000000000000000 GPR16: c0000008a1b20048 c000000001595898 c000000001750c18 0000000000000002 GPR20: c000000001750c28 c000000001624470 0000000fffffffe0 5deadbeef0000122 GPR24: 0000000000000001 0000000000000dc0 0000000000000002 c0000000003443d8 GPR28: c0000008b301f500 c0000008bb398620 0000000000000000 c00c000002287180 NIP ___slab_alloc+0x1f4/0x760 LR __slab_alloc+0x34/0x60 Call Trace: ___slab_alloc+0x334/0x760 (unreliable) __slab_alloc+0x34/0x60 __kmalloc_node+0x110/0x490 kvmalloc_node+0x58/0x110 mem_cgroup_css_online+0x108/0x270 online_css+0x48/0xd0 cgroup_apply_control_enable+0x2ec/0x4d0 cgroup_mkdir+0x228/0x5f0 kernfs_iop_mkdir+0x90/0xf0 vfs_mkdir+0x110/0x230 do_mkdirat+0xb0/0x1a0 system_call+0x5c/0x68 This is a PowerPC platform with following NUMA topology: available: 2 nodes (0-1) node 0 cpus: node 0 size: 0 MB node 0 free: 0 MB node 1 cpus: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 node 1 size: 35247 MB node 1 free: 30907 MB node distances: node 0 1 0: 10 40 1: 40 10 possible numa nodes: 0-31 This only happens with a mmotm patch "mm/memcontrol.c: allocate shrinker_map on appropriate NUMA node" [2] which effectively calls kmalloc_node for each possible node. SLUB however only allocates kmem_cache_node on online N_NORMAL_MEMORY nodes, and relies on node_to_mem_node to return such valid node for other nodes since commit a561ce00b09e ("slub: fall back to node_to_mem_node() node if allocating on memoryless node"). This is however not true in this configuration where the _node_numa_mem_ array is not initialized for nodes 0 and 2-31, thus it contains zeroes and get_partial() ends up accessing non-allocated kmem_cache_node. A related issue was reported by Bharata (originally by Ramachandran) [3] where a similar PowerPC configuration, but with mainline kernel without patch [2] ends up allocating large amounts of pages by kmalloc-1k kmalloc-512. This seems to have the same underlying issue with node_to_mem_node() not behaving as expected, and might probably also lead to an infinite loop with CONFIG_SLUB_CPU_PARTIAL [4]. This patch should fix both issues by not relying on node_to_mem_node() anymore and instead simply falling back to NUMA_NO_NODE, when kmalloc_node(node) is attempted for a node that's not online, or has no usable memory. The "usable memory" condition is also changed from node_present_pages() to N_NORMAL_MEMORY node state, as that is exactly the condition that SLUB uses to allocate kmem_cache_node structures. The check in get_partial() is removed completely, as the checks in ___slab_alloc() are now sufficient to prevent get_partial() being reached with an invalid node. [1] https://lore.kernel.org/linux-next/3381CD91-AB3D-4773-BA04-E7A072A63968@linux.vnet.ibm.com/ [2] https://lore.kernel.org/linux-mm/fff0e636-4c36-ed10-281c-8cdb0687c839@virtuozzo.com/ [3] https://lore.kernel.org/linux-mm/20200317092624.GB22538@in.ibm.com/ [4] https://lore.kernel.org/linux-mm/088b5996-faae-8a56-ef9c-5b567125ae54@suse.cz/ Fixes: a561ce00b09e ("slub: fall back to node_to_mem_node() node if allocating on memoryless node") Reported-by: Sachin Sant <sachinp@linux.vnet.ibm.com> Reported-by: PUVICHAKRAVARTHY RAMACHANDRAN <puvichakravarthy@in.ibm.com> Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Tested-by: Sachin Sant <sachinp@linux.vnet.ibm.com> Tested-by: Bharata B Rao <bharata@linux.ibm.com> Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Michael Ellerman <mpe@ellerman.id.au> Cc: Michal Hocko <mhocko@kernel.org> Cc: Christopher Lameter <cl@linux.com> Cc: linuxppc-dev@lists.ozlabs.org Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Kirill Tkhai <ktkhai@virtuozzo.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Nathan Lynch <nathanl@linux.ibm.com> Cc: <stable@vger.kernel.org> Link: http://lkml.kernel.org/r/20200320115533.9604-1-vbabka@suse.cz Debugged-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-03-22 09:22:37 +08:00
/*
* same as above but node_match() being false already
* implies node != NUMA_NO_NODE
*/
mm, slab, slub: stop taking memory hotplug lock Since commit 03afc0e25f7f ("slab: get_online_mems for kmem_cache_{create,destroy,shrink}") we are taking memory hotplug lock for SLAB and SLUB when creating, destroying or shrinking a cache. It is quite a heavy lock and it's best to avoid it if possible, as we had several issues with lockdep complaining about ordering in the past, see e.g. e4f8e513c3d3 ("mm/slub: fix a deadlock in show_slab_objects()"). The problem scenario in 03afc0e25f7f (solved by the memory hotplug lock) can be summarized as follows: while there's slab_mutex synchronizing new kmem cache creation and SLUB's MEM_GOING_ONLINE callback slab_mem_going_online_callback(), we may miss creation of kmem_cache_node for the hotplugged node in the new kmem cache, because the hotplug callback doesn't yet see the new cache, and cache creation in init_kmem_cache_nodes() only inits kmem_cache_node for nodes in the N_NORMAL_MEMORY nodemask, which however may not yet include the new node, as that happens only later after the MEM_GOING_ONLINE callback. Instead of using get/put_online_mems(), the problem can be solved by SLUB maintaining its own nodemask of nodes for which it has allocated the per-node kmem_cache_node structures. This nodemask would generally mirror the N_NORMAL_MEMORY nodemask, but would be updated only in under SLUB's control in its memory hotplug callbacks under the slab_mutex. This patch adds such nodemask and its handling. Commit 03afc0e25f7f mentiones "issues like [the one above]", but there don't appear to be further issues. All the paths (shared for SLAB and SLUB) taking the memory hotplug locks are also taking the slab_mutex, except kmem_cache_shrink() where 03afc0e25f7f replaced slab_mutex with get/put_online_mems(). We however cannot simply restore slab_mutex in kmem_cache_shrink(), as SLUB can enters the function from a write to sysfs 'shrink' file, thus holding kernfs lock, and in kmem_cache_create() the kernfs lock is nested within slab_mutex. But on closer inspection we don't actually need to protect kmem_cache_shrink() from hotplug callbacks: While SLUB's __kmem_cache_shrink() does for_each_kmem_cache_node(), missing a new node added in parallel hotplug is not fatal, and parallel hotremove does not free kmem_cache_node's anymore after the previous patch, so use-after free cannot happen. The per-node shrinking itself is protected by n->list_lock. Same is true for SLAB, and SLOB is no-op. SLAB also doesn't need the memory hotplug locking, which it only gained by 03afc0e25f7f through the shared paths in slab_common.c. Its memory hotplug callbacks are also protected by slab_mutex against races with these paths. The problem of SLUB relying on N_NORMAL_MEMORY doesn't apply to SLAB, as its setup_kmem_cache_nodes relies on N_ONLINE, and the new node is already set there during the MEM_GOING_ONLINE callback, so no special care is needed for SLAB. As such, this patch removes all get/put_online_mems() usage by the slab subsystem. Link: https://lkml.kernel.org/r/20210113131634.3671-3-vbabka@suse.cz Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Cc: Christoph Lameter <cl@linux.com> Cc: David Hildenbrand <david@redhat.com> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: Qian Cai <cai@redhat.com> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-25 04:01:12 +08:00
if (!node_isset(node, slab_nodes)) {
mm, slub: prevent kmalloc_node crashes and memory leaks Sachin reports [1] a crash in SLUB __slab_alloc(): BUG: Kernel NULL pointer dereference on read at 0x000073b0 Faulting instruction address: 0xc0000000003d55f4 Oops: Kernel access of bad area, sig: 11 [#1] LE PAGE_SIZE=64K MMU=Hash SMP NR_CPUS=2048 NUMA pSeries Modules linked in: CPU: 19 PID: 1 Comm: systemd Not tainted 5.6.0-rc2-next-20200218-autotest #1 NIP: c0000000003d55f4 LR: c0000000003d5b94 CTR: 0000000000000000 REGS: c0000008b37836d0 TRAP: 0300 Not tainted (5.6.0-rc2-next-20200218-autotest) MSR: 8000000000009033 <SF,EE,ME,IR,DR,RI,LE> CR: 24004844 XER: 00000000 CFAR: c00000000000dec4 DAR: 00000000000073b0 DSISR: 40000000 IRQMASK: 1 GPR00: c0000000003d5b94 c0000008b3783960 c00000000155d400 c0000008b301f500 GPR04: 0000000000000dc0 0000000000000002 c0000000003443d8 c0000008bb398620 GPR08: 00000008ba2f0000 0000000000000001 0000000000000000 0000000000000000 GPR12: 0000000024004844 c00000001ec52a00 0000000000000000 0000000000000000 GPR16: c0000008a1b20048 c000000001595898 c000000001750c18 0000000000000002 GPR20: c000000001750c28 c000000001624470 0000000fffffffe0 5deadbeef0000122 GPR24: 0000000000000001 0000000000000dc0 0000000000000002 c0000000003443d8 GPR28: c0000008b301f500 c0000008bb398620 0000000000000000 c00c000002287180 NIP ___slab_alloc+0x1f4/0x760 LR __slab_alloc+0x34/0x60 Call Trace: ___slab_alloc+0x334/0x760 (unreliable) __slab_alloc+0x34/0x60 __kmalloc_node+0x110/0x490 kvmalloc_node+0x58/0x110 mem_cgroup_css_online+0x108/0x270 online_css+0x48/0xd0 cgroup_apply_control_enable+0x2ec/0x4d0 cgroup_mkdir+0x228/0x5f0 kernfs_iop_mkdir+0x90/0xf0 vfs_mkdir+0x110/0x230 do_mkdirat+0xb0/0x1a0 system_call+0x5c/0x68 This is a PowerPC platform with following NUMA topology: available: 2 nodes (0-1) node 0 cpus: node 0 size: 0 MB node 0 free: 0 MB node 1 cpus: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 node 1 size: 35247 MB node 1 free: 30907 MB node distances: node 0 1 0: 10 40 1: 40 10 possible numa nodes: 0-31 This only happens with a mmotm patch "mm/memcontrol.c: allocate shrinker_map on appropriate NUMA node" [2] which effectively calls kmalloc_node for each possible node. SLUB however only allocates kmem_cache_node on online N_NORMAL_MEMORY nodes, and relies on node_to_mem_node to return such valid node for other nodes since commit a561ce00b09e ("slub: fall back to node_to_mem_node() node if allocating on memoryless node"). This is however not true in this configuration where the _node_numa_mem_ array is not initialized for nodes 0 and 2-31, thus it contains zeroes and get_partial() ends up accessing non-allocated kmem_cache_node. A related issue was reported by Bharata (originally by Ramachandran) [3] where a similar PowerPC configuration, but with mainline kernel without patch [2] ends up allocating large amounts of pages by kmalloc-1k kmalloc-512. This seems to have the same underlying issue with node_to_mem_node() not behaving as expected, and might probably also lead to an infinite loop with CONFIG_SLUB_CPU_PARTIAL [4]. This patch should fix both issues by not relying on node_to_mem_node() anymore and instead simply falling back to NUMA_NO_NODE, when kmalloc_node(node) is attempted for a node that's not online, or has no usable memory. The "usable memory" condition is also changed from node_present_pages() to N_NORMAL_MEMORY node state, as that is exactly the condition that SLUB uses to allocate kmem_cache_node structures. The check in get_partial() is removed completely, as the checks in ___slab_alloc() are now sufficient to prevent get_partial() being reached with an invalid node. [1] https://lore.kernel.org/linux-next/3381CD91-AB3D-4773-BA04-E7A072A63968@linux.vnet.ibm.com/ [2] https://lore.kernel.org/linux-mm/fff0e636-4c36-ed10-281c-8cdb0687c839@virtuozzo.com/ [3] https://lore.kernel.org/linux-mm/20200317092624.GB22538@in.ibm.com/ [4] https://lore.kernel.org/linux-mm/088b5996-faae-8a56-ef9c-5b567125ae54@suse.cz/ Fixes: a561ce00b09e ("slub: fall back to node_to_mem_node() node if allocating on memoryless node") Reported-by: Sachin Sant <sachinp@linux.vnet.ibm.com> Reported-by: PUVICHAKRAVARTHY RAMACHANDRAN <puvichakravarthy@in.ibm.com> Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Tested-by: Sachin Sant <sachinp@linux.vnet.ibm.com> Tested-by: Bharata B Rao <bharata@linux.ibm.com> Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Michael Ellerman <mpe@ellerman.id.au> Cc: Michal Hocko <mhocko@kernel.org> Cc: Christopher Lameter <cl@linux.com> Cc: linuxppc-dev@lists.ozlabs.org Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Kirill Tkhai <ktkhai@virtuozzo.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Nathan Lynch <nathanl@linux.ibm.com> Cc: <stable@vger.kernel.org> Link: http://lkml.kernel.org/r/20200320115533.9604-1-vbabka@suse.cz Debugged-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-03-22 09:22:37 +08:00
node = NUMA_NO_NODE;
goto redo;
} else {
stat(s, ALLOC_NODE_MISMATCH);
deactivate_slab(s, page, c->freelist, c);
goto new_slab;
}
}
mm: sl[au]b: add knowledge of PFMEMALLOC reserve pages When a user or administrator requires swap for their application, they create a swap partition and file, format it with mkswap and activate it with swapon. Swap over the network is considered as an option in diskless systems. The two likely scenarios are when blade servers are used as part of a cluster where the form factor or maintenance costs do not allow the use of disks and thin clients. The Linux Terminal Server Project recommends the use of the Network Block Device (NBD) for swap according to the manual at https://sourceforge.net/projects/ltsp/files/Docs-Admin-Guide/LTSPManual.pdf/download There is also documentation and tutorials on how to setup swap over NBD at places like https://help.ubuntu.com/community/UbuntuLTSP/EnableNBDSWAP The nbd-client also documents the use of NBD as swap. Despite this, the fact is that a machine using NBD for swap can deadlock within minutes if swap is used intensively. This patch series addresses the problem. The core issue is that network block devices do not use mempools like normal block devices do. As the host cannot control where they receive packets from, they cannot reliably work out in advance how much memory they might need. Some years ago, Peter Zijlstra developed a series of patches that supported swap over an NFS that at least one distribution is carrying within their kernels. This patch series borrows very heavily from Peter's work to support swapping over NBD as a pre-requisite to supporting swap-over-NFS. The bulk of the complexity is concerned with preserving memory that is allocated from the PFMEMALLOC reserves for use by the network layer which is needed for both NBD and NFS. Patch 1 adds knowledge of the PFMEMALLOC reserves to SLAB and SLUB to preserve access to pages allocated under low memory situations to callers that are freeing memory. Patch 2 optimises the SLUB fast path to avoid pfmemalloc checks Patch 3 introduces __GFP_MEMALLOC to allow access to the PFMEMALLOC reserves without setting PFMEMALLOC. Patch 4 opens the possibility for softirqs to use PFMEMALLOC reserves for later use by network packet processing. Patch 5 only sets page->pfmemalloc when ALLOC_NO_WATERMARKS was required Patch 6 ignores memory policies when ALLOC_NO_WATERMARKS is set. Patches 7-12 allows network processing to use PFMEMALLOC reserves when the socket has been marked as being used by the VM to clean pages. If packets are received and stored in pages that were allocated under low-memory situations and are unrelated to the VM, the packets are dropped. Patch 11 reintroduces __skb_alloc_page which the networking folk may object to but is needed in some cases to propogate pfmemalloc from a newly allocated page to an skb. If there is a strong objection, this patch can be dropped with the impact being that swap-over-network will be slower in some cases but it should not fail. Patch 13 is a micro-optimisation to avoid a function call in the common case. Patch 14 tags NBD sockets as being SOCK_MEMALLOC so they can use PFMEMALLOC if necessary. Patch 15 notes that it is still possible for the PFMEMALLOC reserve to be depleted. To prevent this, direct reclaimers get throttled on a waitqueue if 50% of the PFMEMALLOC reserves are depleted. It is expected that kswapd and the direct reclaimers already running will clean enough pages for the low watermark to be reached and the throttled processes are woken up. Patch 16 adds a statistic to track how often processes get throttled Some basic performance testing was run using kernel builds, netperf on loopback for UDP and TCP, hackbench (pipes and sockets), iozone and sysbench. Each of them were expected to use the sl*b allocators reasonably heavily but there did not appear to be significant performance variances. For testing swap-over-NBD, a machine was booted with 2G of RAM with a swapfile backed by NBD. 8*NUM_CPU processes were started that create anonymous memory mappings and read them linearly in a loop. The total size of the mappings were 4*PHYSICAL_MEMORY to use swap heavily under memory pressure. Without the patches and using SLUB, the machine locks up within minutes and runs to completion with them applied. With SLAB, the story is different as an unpatched kernel run to completion. However, the patched kernel completed the test 45% faster. MICRO 3.5.0-rc2 3.5.0-rc2 vanilla swapnbd Unrecognised test vmscan-anon-mmap-write MMTests Statistics: duration Sys Time Running Test (seconds) 197.80 173.07 User+Sys Time Running Test (seconds) 206.96 182.03 Total Elapsed Time (seconds) 3240.70 1762.09 This patch: mm: sl[au]b: add knowledge of PFMEMALLOC reserve pages Allocations of pages below the min watermark run a risk of the machine hanging due to a lack of memory. To prevent this, only callers who have PF_MEMALLOC or TIF_MEMDIE set and are not processing an interrupt are allowed to allocate with ALLOC_NO_WATERMARKS. Once they are allocated to a slab though, nothing prevents other callers consuming free objects within those slabs. This patch limits access to slab pages that were alloced from the PFMEMALLOC reserves. When this patch is applied, pages allocated from below the low watermark are returned with page->pfmemalloc set and it is up to the caller to determine how the page should be protected. SLAB restricts access to any page with page->pfmemalloc set to callers which are known to able to access the PFMEMALLOC reserve. If one is not available, an attempt is made to allocate a new page rather than use a reserve. SLUB is a bit more relaxed in that it only records if the current per-CPU page was allocated from PFMEMALLOC reserve and uses another partial slab if the caller does not have the necessary GFP or process flags. This was found to be sufficient in tests to avoid hangs due to SLUB generally maintaining smaller lists than SLAB. In low-memory conditions it does mean that !PFMEMALLOC allocators can fail a slab allocation even though free objects are available because they are being preserved for callers that are freeing pages. [a.p.zijlstra@chello.nl: Original implementation] [sebastian@breakpoint.cc: Correct order of page flag clearing] Signed-off-by: Mel Gorman <mgorman@suse.de> Cc: David Miller <davem@davemloft.net> Cc: Neil Brown <neilb@suse.de> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Christie <michaelc@cs.wisc.edu> Cc: Eric B Munson <emunson@mgebm.net> Cc: Eric Dumazet <eric.dumazet@gmail.com> Cc: Sebastian Andrzej Siewior <sebastian@breakpoint.cc> Cc: Mel Gorman <mgorman@suse.de> Cc: Christoph Lameter <cl@linux.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-08-01 07:43:58 +08:00
/*
* By rights, we should be searching for a slab page that was
* PFMEMALLOC but right now, we are losing the pfmemalloc
* information when the page leaves the per-cpu allocator
*/
if (unlikely(!pfmemalloc_match(page, gfpflags))) {
deactivate_slab(s, page, c->freelist, c);
mm: sl[au]b: add knowledge of PFMEMALLOC reserve pages When a user or administrator requires swap for their application, they create a swap partition and file, format it with mkswap and activate it with swapon. Swap over the network is considered as an option in diskless systems. The two likely scenarios are when blade servers are used as part of a cluster where the form factor or maintenance costs do not allow the use of disks and thin clients. The Linux Terminal Server Project recommends the use of the Network Block Device (NBD) for swap according to the manual at https://sourceforge.net/projects/ltsp/files/Docs-Admin-Guide/LTSPManual.pdf/download There is also documentation and tutorials on how to setup swap over NBD at places like https://help.ubuntu.com/community/UbuntuLTSP/EnableNBDSWAP The nbd-client also documents the use of NBD as swap. Despite this, the fact is that a machine using NBD for swap can deadlock within minutes if swap is used intensively. This patch series addresses the problem. The core issue is that network block devices do not use mempools like normal block devices do. As the host cannot control where they receive packets from, they cannot reliably work out in advance how much memory they might need. Some years ago, Peter Zijlstra developed a series of patches that supported swap over an NFS that at least one distribution is carrying within their kernels. This patch series borrows very heavily from Peter's work to support swapping over NBD as a pre-requisite to supporting swap-over-NFS. The bulk of the complexity is concerned with preserving memory that is allocated from the PFMEMALLOC reserves for use by the network layer which is needed for both NBD and NFS. Patch 1 adds knowledge of the PFMEMALLOC reserves to SLAB and SLUB to preserve access to pages allocated under low memory situations to callers that are freeing memory. Patch 2 optimises the SLUB fast path to avoid pfmemalloc checks Patch 3 introduces __GFP_MEMALLOC to allow access to the PFMEMALLOC reserves without setting PFMEMALLOC. Patch 4 opens the possibility for softirqs to use PFMEMALLOC reserves for later use by network packet processing. Patch 5 only sets page->pfmemalloc when ALLOC_NO_WATERMARKS was required Patch 6 ignores memory policies when ALLOC_NO_WATERMARKS is set. Patches 7-12 allows network processing to use PFMEMALLOC reserves when the socket has been marked as being used by the VM to clean pages. If packets are received and stored in pages that were allocated under low-memory situations and are unrelated to the VM, the packets are dropped. Patch 11 reintroduces __skb_alloc_page which the networking folk may object to but is needed in some cases to propogate pfmemalloc from a newly allocated page to an skb. If there is a strong objection, this patch can be dropped with the impact being that swap-over-network will be slower in some cases but it should not fail. Patch 13 is a micro-optimisation to avoid a function call in the common case. Patch 14 tags NBD sockets as being SOCK_MEMALLOC so they can use PFMEMALLOC if necessary. Patch 15 notes that it is still possible for the PFMEMALLOC reserve to be depleted. To prevent this, direct reclaimers get throttled on a waitqueue if 50% of the PFMEMALLOC reserves are depleted. It is expected that kswapd and the direct reclaimers already running will clean enough pages for the low watermark to be reached and the throttled processes are woken up. Patch 16 adds a statistic to track how often processes get throttled Some basic performance testing was run using kernel builds, netperf on loopback for UDP and TCP, hackbench (pipes and sockets), iozone and sysbench. Each of them were expected to use the sl*b allocators reasonably heavily but there did not appear to be significant performance variances. For testing swap-over-NBD, a machine was booted with 2G of RAM with a swapfile backed by NBD. 8*NUM_CPU processes were started that create anonymous memory mappings and read them linearly in a loop. The total size of the mappings were 4*PHYSICAL_MEMORY to use swap heavily under memory pressure. Without the patches and using SLUB, the machine locks up within minutes and runs to completion with them applied. With SLAB, the story is different as an unpatched kernel run to completion. However, the patched kernel completed the test 45% faster. MICRO 3.5.0-rc2 3.5.0-rc2 vanilla swapnbd Unrecognised test vmscan-anon-mmap-write MMTests Statistics: duration Sys Time Running Test (seconds) 197.80 173.07 User+Sys Time Running Test (seconds) 206.96 182.03 Total Elapsed Time (seconds) 3240.70 1762.09 This patch: mm: sl[au]b: add knowledge of PFMEMALLOC reserve pages Allocations of pages below the min watermark run a risk of the machine hanging due to a lack of memory. To prevent this, only callers who have PF_MEMALLOC or TIF_MEMDIE set and are not processing an interrupt are allowed to allocate with ALLOC_NO_WATERMARKS. Once they are allocated to a slab though, nothing prevents other callers consuming free objects within those slabs. This patch limits access to slab pages that were alloced from the PFMEMALLOC reserves. When this patch is applied, pages allocated from below the low watermark are returned with page->pfmemalloc set and it is up to the caller to determine how the page should be protected. SLAB restricts access to any page with page->pfmemalloc set to callers which are known to able to access the PFMEMALLOC reserve. If one is not available, an attempt is made to allocate a new page rather than use a reserve. SLUB is a bit more relaxed in that it only records if the current per-CPU page was allocated from PFMEMALLOC reserve and uses another partial slab if the caller does not have the necessary GFP or process flags. This was found to be sufficient in tests to avoid hangs due to SLUB generally maintaining smaller lists than SLAB. In low-memory conditions it does mean that !PFMEMALLOC allocators can fail a slab allocation even though free objects are available because they are being preserved for callers that are freeing pages. [a.p.zijlstra@chello.nl: Original implementation] [sebastian@breakpoint.cc: Correct order of page flag clearing] Signed-off-by: Mel Gorman <mgorman@suse.de> Cc: David Miller <davem@davemloft.net> Cc: Neil Brown <neilb@suse.de> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Christie <michaelc@cs.wisc.edu> Cc: Eric B Munson <emunson@mgebm.net> Cc: Eric Dumazet <eric.dumazet@gmail.com> Cc: Sebastian Andrzej Siewior <sebastian@breakpoint.cc> Cc: Mel Gorman <mgorman@suse.de> Cc: Christoph Lameter <cl@linux.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-08-01 07:43:58 +08:00
goto new_slab;
}
/* must check again c->freelist in case of cpu migration or IRQ */
freelist = c->freelist;
if (freelist)
goto load_freelist;
freelist = get_freelist(s, page);
if (!freelist) {
c->page = NULL;
stat(s, DEACTIVATE_BYPASS);
goto new_slab;
}
stat(s, ALLOC_REFILL);
slub: support concurrent local and remote frees and allocs on a slab Avoid atomic overhead in slab_alloc and slab_free SLUB needs to use the slab_lock for the per cpu slabs to synchronize with potential kfree operations. This patch avoids that need by moving all free objects onto a lockless_freelist. The regular freelist continues to exist and will be used to free objects. So while we consume the lockless_freelist the regular freelist may build up objects. If we are out of objects on the lockless_freelist then we may check the regular freelist. If it has objects then we move those over to the lockless_freelist and do this again. There is a significant savings in terms of atomic operations that have to be performed. We can even free directly to the lockless_freelist if we know that we are running on the same processor. So this speeds up short lived objects. They may be allocated and freed without taking the slab_lock. This is particular good for netperf. In order to maximize the effect of the new faster hotpath we extract the hottest performance pieces into inlined functions. These are then inlined into kmem_cache_alloc and kmem_cache_free. So hotpath allocation and freeing no longer requires a subroutine call within SLUB. [I am not sure that it is worth doing this because it changes the easy to read structure of slub just to reduce atomic ops. However, there is someone out there with a benchmark on 4 way and 8 way processor systems that seems to show a 5% regression vs. Slab. Seems that the regression is due to increased atomic operations use vs. SLAB in SLUB). I wonder if this is applicable or discernable at all in a real workload?] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-10 18:15:16 +08:00
load_freelist:
/*
* freelist is pointing to the list of objects to be used.
* page is pointing to the page from which the objects are obtained.
* That page must be frozen for per cpu allocations to work.
*/
VM_BUG_ON(!c->page->frozen);
c->freelist = get_freepointer(s, freelist);
Lockless (and preemptless) fastpaths for slub Use the this_cpu_cmpxchg_double functionality to implement a lockless allocation algorithm on arches that support fast this_cpu_ops. Each of the per cpu pointers is paired with a transaction id that ensures that updates of the per cpu information can only occur in sequence on a certain cpu. A transaction id is a "long" integer that is comprised of an event number and the cpu number. The event number is incremented for every change to the per cpu state. This means that the cmpxchg instruction can verify for an update that nothing interfered and that we are updating the percpu structure for the processor where we picked up the information and that we are also currently on that processor when we update the information. This results in a significant decrease of the overhead in the fastpaths. It also makes it easy to adopt the fast path for realtime kernels since this is lockless and does not require the use of the current per cpu area over the critical section. It is only important that the per cpu area is current at the beginning of the critical section and at the end. So there is no need even to disable preemption. Test results show that the fastpath cycle count is reduced by up to ~ 40% (alloc/free test goes from ~140 cycles down to ~80). The slowpath for kfree adds a few cycles. Sadly this does nothing for the slowpath which is where the main issues with performance in slub are but the best case performance rises significantly. (For that see the more complex slub patches that require cmpxchg_double) Kmalloc: alloc/free test Before: 10000 times kmalloc(8)/kfree -> 134 cycles 10000 times kmalloc(16)/kfree -> 152 cycles 10000 times kmalloc(32)/kfree -> 144 cycles 10000 times kmalloc(64)/kfree -> 142 cycles 10000 times kmalloc(128)/kfree -> 142 cycles 10000 times kmalloc(256)/kfree -> 132 cycles 10000 times kmalloc(512)/kfree -> 132 cycles 10000 times kmalloc(1024)/kfree -> 135 cycles 10000 times kmalloc(2048)/kfree -> 135 cycles 10000 times kmalloc(4096)/kfree -> 135 cycles 10000 times kmalloc(8192)/kfree -> 144 cycles 10000 times kmalloc(16384)/kfree -> 754 cycles After: 10000 times kmalloc(8)/kfree -> 78 cycles 10000 times kmalloc(16)/kfree -> 78 cycles 10000 times kmalloc(32)/kfree -> 82 cycles 10000 times kmalloc(64)/kfree -> 88 cycles 10000 times kmalloc(128)/kfree -> 79 cycles 10000 times kmalloc(256)/kfree -> 79 cycles 10000 times kmalloc(512)/kfree -> 85 cycles 10000 times kmalloc(1024)/kfree -> 82 cycles 10000 times kmalloc(2048)/kfree -> 82 cycles 10000 times kmalloc(4096)/kfree -> 85 cycles 10000 times kmalloc(8192)/kfree -> 82 cycles 10000 times kmalloc(16384)/kfree -> 706 cycles Kmalloc: Repeatedly allocate then free test Before: 10000 times kmalloc(8) -> 211 cycles kfree -> 113 cycles 10000 times kmalloc(16) -> 174 cycles kfree -> 115 cycles 10000 times kmalloc(32) -> 235 cycles kfree -> 129 cycles 10000 times kmalloc(64) -> 222 cycles kfree -> 120 cycles 10000 times kmalloc(128) -> 343 cycles kfree -> 139 cycles 10000 times kmalloc(256) -> 827 cycles kfree -> 147 cycles 10000 times kmalloc(512) -> 1048 cycles kfree -> 272 cycles 10000 times kmalloc(1024) -> 2043 cycles kfree -> 528 cycles 10000 times kmalloc(2048) -> 4002 cycles kfree -> 571 cycles 10000 times kmalloc(4096) -> 7740 cycles kfree -> 628 cycles 10000 times kmalloc(8192) -> 8062 cycles kfree -> 850 cycles 10000 times kmalloc(16384) -> 8895 cycles kfree -> 1249 cycles After: 10000 times kmalloc(8) -> 190 cycles kfree -> 129 cycles 10000 times kmalloc(16) -> 76 cycles kfree -> 123 cycles 10000 times kmalloc(32) -> 126 cycles kfree -> 124 cycles 10000 times kmalloc(64) -> 181 cycles kfree -> 128 cycles 10000 times kmalloc(128) -> 310 cycles kfree -> 140 cycles 10000 times kmalloc(256) -> 809 cycles kfree -> 165 cycles 10000 times kmalloc(512) -> 1005 cycles kfree -> 269 cycles 10000 times kmalloc(1024) -> 1999 cycles kfree -> 527 cycles 10000 times kmalloc(2048) -> 3967 cycles kfree -> 570 cycles 10000 times kmalloc(4096) -> 7658 cycles kfree -> 637 cycles 10000 times kmalloc(8192) -> 8111 cycles kfree -> 859 cycles 10000 times kmalloc(16384) -> 8791 cycles kfree -> 1173 cycles Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-02-26 01:38:54 +08:00
c->tid = next_tid(c->tid);
return freelist;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
new_slab:
if (slub_percpu_partial(c)) {
page = c->page = slub_percpu_partial(c);
slub_set_percpu_partial(c, page);
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
stat(s, CPU_PARTIAL_ALLOC);
goto redo;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
freelist = new_slab_objects(s, gfpflags, node, &c);
if (unlikely(!freelist)) {
slab_out_of_memory(s, gfpflags, node);
return NULL;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
page = c->page;
mm: slub: optimise the SLUB fast path to avoid pfmemalloc checks This patch removes the check for pfmemalloc from the alloc hotpath and puts the logic after the election of a new per cpu slab. For a pfmemalloc page we do not use the fast path but force the use of the slow path which is also used for the debug case. This has the side-effect of weakening pfmemalloc processing in the following way; 1. A process that is allocating for network swap calls __slab_alloc. pfmemalloc_match is true so the freelist is loaded and c->freelist is now pointing to a pfmemalloc page. 2. A process that is attempting normal allocations calls slab_alloc, finds the pfmemalloc page on the freelist and uses it because it did not check pfmemalloc_match() The patch allows non-pfmemalloc allocations to use pfmemalloc pages with the kmalloc slabs being the most vunerable caches on the grounds they are most likely to have a mix of pfmemalloc and !pfmemalloc requests. A later patch will still protect the system as processes will get throttled if the pfmemalloc reserves get depleted but performance will not degrade as smoothly. [mgorman@suse.de: Expanded changelog] Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Mel Gorman <mgorman@suse.de> Cc: David Miller <davem@davemloft.net> Cc: Neil Brown <neilb@suse.de> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Christie <michaelc@cs.wisc.edu> Cc: Eric B Munson <emunson@mgebm.net> Cc: Eric Dumazet <eric.dumazet@gmail.com> Cc: Sebastian Andrzej Siewior <sebastian@breakpoint.cc> Cc: Mel Gorman <mgorman@suse.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-08-01 07:44:00 +08:00
if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
goto load_freelist;
/* Only entered in the debug case */
if (kmem_cache_debug(s) &&
!alloc_debug_processing(s, page, freelist, addr))
goto new_slab; /* Slab failed checks. Next slab needed */
slub: support concurrent local and remote frees and allocs on a slab Avoid atomic overhead in slab_alloc and slab_free SLUB needs to use the slab_lock for the per cpu slabs to synchronize with potential kfree operations. This patch avoids that need by moving all free objects onto a lockless_freelist. The regular freelist continues to exist and will be used to free objects. So while we consume the lockless_freelist the regular freelist may build up objects. If we are out of objects on the lockless_freelist then we may check the regular freelist. If it has objects then we move those over to the lockless_freelist and do this again. There is a significant savings in terms of atomic operations that have to be performed. We can even free directly to the lockless_freelist if we know that we are running on the same processor. So this speeds up short lived objects. They may be allocated and freed without taking the slab_lock. This is particular good for netperf. In order to maximize the effect of the new faster hotpath we extract the hottest performance pieces into inlined functions. These are then inlined into kmem_cache_alloc and kmem_cache_free. So hotpath allocation and freeing no longer requires a subroutine call within SLUB. [I am not sure that it is worth doing this because it changes the easy to read structure of slub just to reduce atomic ops. However, there is someone out there with a benchmark on 4 way and 8 way processor systems that seems to show a 5% regression vs. Slab. Seems that the regression is due to increased atomic operations use vs. SLAB in SLUB). I wonder if this is applicable or discernable at all in a real workload?] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-10 18:15:16 +08:00
deactivate_slab(s, page, get_freepointer(s, freelist), c);
return freelist;
slub: support concurrent local and remote frees and allocs on a slab Avoid atomic overhead in slab_alloc and slab_free SLUB needs to use the slab_lock for the per cpu slabs to synchronize with potential kfree operations. This patch avoids that need by moving all free objects onto a lockless_freelist. The regular freelist continues to exist and will be used to free objects. So while we consume the lockless_freelist the regular freelist may build up objects. If we are out of objects on the lockless_freelist then we may check the regular freelist. If it has objects then we move those over to the lockless_freelist and do this again. There is a significant savings in terms of atomic operations that have to be performed. We can even free directly to the lockless_freelist if we know that we are running on the same processor. So this speeds up short lived objects. They may be allocated and freed without taking the slab_lock. This is particular good for netperf. In order to maximize the effect of the new faster hotpath we extract the hottest performance pieces into inlined functions. These are then inlined into kmem_cache_alloc and kmem_cache_free. So hotpath allocation and freeing no longer requires a subroutine call within SLUB. [I am not sure that it is worth doing this because it changes the easy to read structure of slub just to reduce atomic ops. However, there is someone out there with a benchmark on 4 way and 8 way processor systems that seems to show a 5% regression vs. Slab. Seems that the regression is due to increased atomic operations use vs. SLAB in SLUB). I wonder if this is applicable or discernable at all in a real workload?] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-10 18:15:16 +08:00
}
/*
* Another one that disabled interrupt and compensates for possible
* cpu changes by refetching the per cpu area pointer.
*/
static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
unsigned long addr, struct kmem_cache_cpu *c)
{
void *p;
unsigned long flags;
local_irq_save(flags);
#ifdef CONFIG_PREEMPTION
/*
* We may have been preempted and rescheduled on a different
* cpu before disabling interrupts. Need to reload cpu area
* pointer.
*/
c = this_cpu_ptr(s->cpu_slab);
#endif
p = ___slab_alloc(s, gfpflags, node, addr, c);
local_irq_restore(flags);
return p;
}
/*
* If the object has been wiped upon free, make sure it's fully initialized by
* zeroing out freelist pointer.
*/
static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s,
void *obj)
{
if (unlikely(slab_want_init_on_free(s)) && obj)
memset((void *)((char *)kasan_reset_tag(obj) + s->offset),
0, sizeof(void *));
}
slub: support concurrent local and remote frees and allocs on a slab Avoid atomic overhead in slab_alloc and slab_free SLUB needs to use the slab_lock for the per cpu slabs to synchronize with potential kfree operations. This patch avoids that need by moving all free objects onto a lockless_freelist. The regular freelist continues to exist and will be used to free objects. So while we consume the lockless_freelist the regular freelist may build up objects. If we are out of objects on the lockless_freelist then we may check the regular freelist. If it has objects then we move those over to the lockless_freelist and do this again. There is a significant savings in terms of atomic operations that have to be performed. We can even free directly to the lockless_freelist if we know that we are running on the same processor. So this speeds up short lived objects. They may be allocated and freed without taking the slab_lock. This is particular good for netperf. In order to maximize the effect of the new faster hotpath we extract the hottest performance pieces into inlined functions. These are then inlined into kmem_cache_alloc and kmem_cache_free. So hotpath allocation and freeing no longer requires a subroutine call within SLUB. [I am not sure that it is worth doing this because it changes the easy to read structure of slub just to reduce atomic ops. However, there is someone out there with a benchmark on 4 way and 8 way processor systems that seems to show a 5% regression vs. Slab. Seems that the regression is due to increased atomic operations use vs. SLAB in SLUB). I wonder if this is applicable or discernable at all in a real workload?] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-10 18:15:16 +08:00
/*
* Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
* have the fastpath folded into their functions. So no function call
* overhead for requests that can be satisfied on the fastpath.
*
* The fastpath works by first checking if the lockless freelist can be used.
* If not then __slab_alloc is called for slow processing.
*
* Otherwise we can simply pick the next object from the lockless free list.
*/
static __always_inline void *slab_alloc_node(struct kmem_cache *s,
mm, kfence: insert KFENCE hooks for SLUB Inserts KFENCE hooks into the SLUB allocator. To pass the originally requested size to KFENCE, add an argument 'orig_size' to slab_alloc*(). The additional argument is required to preserve the requested original size for kmalloc() allocations, which uses size classes (e.g. an allocation of 272 bytes will return an object of size 512). Therefore, kmem_cache::size does not represent the kmalloc-caller's requested size, and we must introduce the argument 'orig_size' to propagate the originally requested size to KFENCE. Without the originally requested size, we would not be able to detect out-of-bounds accesses for objects placed at the end of a KFENCE object page if that object is not equal to the kmalloc-size class it was bucketed into. When KFENCE is disabled, there is no additional overhead, since slab_alloc*() functions are __always_inline. Link: https://lkml.kernel.org/r/20201103175841.3495947-6-elver@google.com Signed-off-by: Marco Elver <elver@google.com> Signed-off-by: Alexander Potapenko <glider@google.com> Reviewed-by: Dmitry Vyukov <dvyukov@google.com> Reviewed-by: Jann Horn <jannh@google.com> Co-developed-by: Marco Elver <elver@google.com> Cc: Andrey Konovalov <andreyknvl@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christopher Lameter <cl@linux.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: David Rientjes <rientjes@google.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Hillf Danton <hdanton@sina.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Joern Engel <joern@purestorage.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kees Cook <keescook@chromium.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Paul E. McKenney <paulmck@kernel.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: SeongJae Park <sjpark@amazon.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-26 09:19:16 +08:00
gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
slub: support concurrent local and remote frees and allocs on a slab Avoid atomic overhead in slab_alloc and slab_free SLUB needs to use the slab_lock for the per cpu slabs to synchronize with potential kfree operations. This patch avoids that need by moving all free objects onto a lockless_freelist. The regular freelist continues to exist and will be used to free objects. So while we consume the lockless_freelist the regular freelist may build up objects. If we are out of objects on the lockless_freelist then we may check the regular freelist. If it has objects then we move those over to the lockless_freelist and do this again. There is a significant savings in terms of atomic operations that have to be performed. We can even free directly to the lockless_freelist if we know that we are running on the same processor. So this speeds up short lived objects. They may be allocated and freed without taking the slab_lock. This is particular good for netperf. In order to maximize the effect of the new faster hotpath we extract the hottest performance pieces into inlined functions. These are then inlined into kmem_cache_alloc and kmem_cache_free. So hotpath allocation and freeing no longer requires a subroutine call within SLUB. [I am not sure that it is worth doing this because it changes the easy to read structure of slub just to reduce atomic ops. However, there is someone out there with a benchmark on 4 way and 8 way processor systems that seems to show a 5% regression vs. Slab. Seems that the regression is due to increased atomic operations use vs. SLAB in SLUB). I wonder if this is applicable or discernable at all in a real workload?] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-10 18:15:16 +08:00
{
void *object;
SLUB: Avoid page struct cacheline bouncing due to remote frees to cpu slab A remote free may access the same page struct that also contains the lockless freelist for the cpu slab. If objects have a short lifetime and are freed by a different processor then remote frees back to the slab from which we are currently allocating are frequent. The cacheline with the page struct needs to be repeately acquired in exclusive mode by both the allocating thread and the freeing thread. If this is frequent enough then performance will suffer because of cacheline bouncing. This patchset puts the lockless_freelist pointer in its own cacheline. In order to make that happen we introduce a per cpu structure called kmem_cache_cpu. Instead of keeping an array of pointers to page structs we now keep an array to a per cpu structure that--among other things--contains the pointer to the lockless freelist. The freeing thread can then keep possession of exclusive access to the page struct cacheline while the allocating thread keeps its exclusive access to the cacheline containing the per cpu structure. This works as long as the allocating cpu is able to service its request from the lockless freelist. If the lockless freelist runs empty then the allocating thread needs to acquire exclusive access to the cacheline with the page struct lock the slab. The allocating thread will then check if new objects were freed to the per cpu slab. If so it will keep the slab as the cpu slab and continue with the recently remote freed objects. So the allocating thread can take a series of just freed remote pages and dish them out again. Ideally allocations could be just recycling objects in the same slab this way which will lead to an ideal allocation / remote free pattern. The number of objects that can be handled in this way is limited by the capacity of one slab. Increasing slab size via slub_min_objects/ slub_max_order may increase the number of objects and therefore performance. If the allocating thread runs out of objects and finds that no objects were put back by the remote processor then it will retrieve a new slab (from the partial lists or from the page allocator) and start with a whole new set of objects while the remote thread may still be freeing objects to the old cpu slab. This may then repeat until the new slab is also exhausted. If remote freeing has freed objects in the earlier slab then that earlier slab will now be on the partial freelist and the allocating thread will pick that slab next for allocation. So the loop is extended. However, both threads need to take the list_lock to make the swizzling via the partial list happen. It is likely that this kind of scheme will keep the objects being passed around to a small set that can be kept in the cpu caches leading to increased performance. More code cleanups become possible: - Instead of passing a cpu we can now pass a kmem_cache_cpu structure around. Allows reducing the number of parameters to various functions. - Can define a new node_match() function for NUMA to encapsulate locality checks. Effect on allocations: Cachelines touched before this patch: Write: page cache struct and first cacheline of object Cachelines touched after this patch: Write: kmem_cache_cpu cacheline and first cacheline of object Read: page cache struct (but see later patch that avoids touching that cacheline) The handling when the lockless alloc list runs empty gets to be a bit more complicated since another cacheline has now to be written to. But that is halfway out of the hot path. Effect on freeing: Cachelines touched before this patch: Write: page_struct and first cacheline of object Cachelines touched after this patch depending on how we free: Write(to cpu_slab): kmem_cache_cpu struct and first cacheline of object Write(to other): page struct and first cacheline of object Read(to cpu_slab): page struct to id slab etc. (but see later patch that avoids touching the page struct on free) Read(to other): cpu local kmem_cache_cpu struct to verify its not the cpu slab. Summary: Pro: - Distinct cachelines so that concurrent remote frees and local allocs on a cpuslab can occur without cacheline bouncing. - Avoids potential bouncing cachelines because of neighboring per cpu pointer updates in kmem_cache's cpu_slab structure since it now grows to a cacheline (Therefore remove the comment that talks about that concern). Cons: - Freeing objects now requires the reading of one additional cacheline. That can be mitigated for some cases by the following patches but its not possible to completely eliminate these references. - Memory usage grows slightly. The size of each per cpu object is blown up from one word (pointing to the page_struct) to one cacheline with various data. So this is NR_CPUS*NR_SLABS*L1_BYTES more memory use. Lets say NR_SLABS is 100 and a cache line size of 128 then we have just increased SLAB metadata requirements by 12.8k per cpu. (Another later patch reduces these requirements) Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 16:26:05 +08:00
struct kmem_cache_cpu *c;
struct page *page;
Lockless (and preemptless) fastpaths for slub Use the this_cpu_cmpxchg_double functionality to implement a lockless allocation algorithm on arches that support fast this_cpu_ops. Each of the per cpu pointers is paired with a transaction id that ensures that updates of the per cpu information can only occur in sequence on a certain cpu. A transaction id is a "long" integer that is comprised of an event number and the cpu number. The event number is incremented for every change to the per cpu state. This means that the cmpxchg instruction can verify for an update that nothing interfered and that we are updating the percpu structure for the processor where we picked up the information and that we are also currently on that processor when we update the information. This results in a significant decrease of the overhead in the fastpaths. It also makes it easy to adopt the fast path for realtime kernels since this is lockless and does not require the use of the current per cpu area over the critical section. It is only important that the per cpu area is current at the beginning of the critical section and at the end. So there is no need even to disable preemption. Test results show that the fastpath cycle count is reduced by up to ~ 40% (alloc/free test goes from ~140 cycles down to ~80). The slowpath for kfree adds a few cycles. Sadly this does nothing for the slowpath which is where the main issues with performance in slub are but the best case performance rises significantly. (For that see the more complex slub patches that require cmpxchg_double) Kmalloc: alloc/free test Before: 10000 times kmalloc(8)/kfree -> 134 cycles 10000 times kmalloc(16)/kfree -> 152 cycles 10000 times kmalloc(32)/kfree -> 144 cycles 10000 times kmalloc(64)/kfree -> 142 cycles 10000 times kmalloc(128)/kfree -> 142 cycles 10000 times kmalloc(256)/kfree -> 132 cycles 10000 times kmalloc(512)/kfree -> 132 cycles 10000 times kmalloc(1024)/kfree -> 135 cycles 10000 times kmalloc(2048)/kfree -> 135 cycles 10000 times kmalloc(4096)/kfree -> 135 cycles 10000 times kmalloc(8192)/kfree -> 144 cycles 10000 times kmalloc(16384)/kfree -> 754 cycles After: 10000 times kmalloc(8)/kfree -> 78 cycles 10000 times kmalloc(16)/kfree -> 78 cycles 10000 times kmalloc(32)/kfree -> 82 cycles 10000 times kmalloc(64)/kfree -> 88 cycles 10000 times kmalloc(128)/kfree -> 79 cycles 10000 times kmalloc(256)/kfree -> 79 cycles 10000 times kmalloc(512)/kfree -> 85 cycles 10000 times kmalloc(1024)/kfree -> 82 cycles 10000 times kmalloc(2048)/kfree -> 82 cycles 10000 times kmalloc(4096)/kfree -> 85 cycles 10000 times kmalloc(8192)/kfree -> 82 cycles 10000 times kmalloc(16384)/kfree -> 706 cycles Kmalloc: Repeatedly allocate then free test Before: 10000 times kmalloc(8) -> 211 cycles kfree -> 113 cycles 10000 times kmalloc(16) -> 174 cycles kfree -> 115 cycles 10000 times kmalloc(32) -> 235 cycles kfree -> 129 cycles 10000 times kmalloc(64) -> 222 cycles kfree -> 120 cycles 10000 times kmalloc(128) -> 343 cycles kfree -> 139 cycles 10000 times kmalloc(256) -> 827 cycles kfree -> 147 cycles 10000 times kmalloc(512) -> 1048 cycles kfree -> 272 cycles 10000 times kmalloc(1024) -> 2043 cycles kfree -> 528 cycles 10000 times kmalloc(2048) -> 4002 cycles kfree -> 571 cycles 10000 times kmalloc(4096) -> 7740 cycles kfree -> 628 cycles 10000 times kmalloc(8192) -> 8062 cycles kfree -> 850 cycles 10000 times kmalloc(16384) -> 8895 cycles kfree -> 1249 cycles After: 10000 times kmalloc(8) -> 190 cycles kfree -> 129 cycles 10000 times kmalloc(16) -> 76 cycles kfree -> 123 cycles 10000 times kmalloc(32) -> 126 cycles kfree -> 124 cycles 10000 times kmalloc(64) -> 181 cycles kfree -> 128 cycles 10000 times kmalloc(128) -> 310 cycles kfree -> 140 cycles 10000 times kmalloc(256) -> 809 cycles kfree -> 165 cycles 10000 times kmalloc(512) -> 1005 cycles kfree -> 269 cycles 10000 times kmalloc(1024) -> 1999 cycles kfree -> 527 cycles 10000 times kmalloc(2048) -> 3967 cycles kfree -> 570 cycles 10000 times kmalloc(4096) -> 7658 cycles kfree -> 637 cycles 10000 times kmalloc(8192) -> 8111 cycles kfree -> 859 cycles 10000 times kmalloc(16384) -> 8791 cycles kfree -> 1173 cycles Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-02-26 01:38:54 +08:00
unsigned long tid;
struct obj_cgroup *objcg = NULL;
kasan, mm: integrate slab init_on_alloc with HW_TAGS This change uses the previously added memory initialization feature of HW_TAGS KASAN routines for slab memory when init_on_alloc is enabled. With this change, memory initialization memset() is no longer called when both HW_TAGS KASAN and init_on_alloc are enabled. Instead, memory is initialized in KASAN runtime. The memory initialization memset() is moved into slab_post_alloc_hook() that currently directly follows the initialization loop. A new argument is added to slab_post_alloc_hook() that indicates whether to initialize the memory or not. To avoid discrepancies with which memory gets initialized that can be caused by future changes, both KASAN hook and initialization memset() are put together and a warning comment is added. Combining setting allocation tags with memory initialization improves HW_TAGS KASAN performance when init_on_alloc is enabled. Link: https://lkml.kernel.org/r/c1292aeb5d519da221ec74a0684a949b027d7720.1615296150.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Reviewed-by: Marco Elver <elver@google.com> Cc: Alexander Potapenko <glider@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Branislav Rankov <Branislav.Rankov@arm.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christoph Lameter <cl@linux.com> Cc: David Rientjes <rientjes@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Evgenii Stepanov <eugenis@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kevin Brodsky <kevin.brodsky@arm.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Collingbourne <pcc@google.com> Cc: Vincenzo Frascino <vincenzo.frascino@arm.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-30 14:00:06 +08:00
bool init = false;
SLUB: Alternate fast paths using cmpxchg_local Provide an alternate implementation of the SLUB fast paths for alloc and free using cmpxchg_local. The cmpxchg_local fast path is selected for arches that have CONFIG_FAST_CMPXCHG_LOCAL set. An arch should only set CONFIG_FAST_CMPXCHG_LOCAL if the cmpxchg_local is faster than an interrupt enable/disable sequence. This is known to be true for both x86 platforms so set FAST_CMPXCHG_LOCAL for both arches. Currently another requirement for the fastpath is that the kernel is compiled without preemption. The restriction will go away with the introduction of a new per cpu allocator and new per cpu operations. The advantages of a cmpxchg_local based fast path are: 1. Potentially lower cycle count (30%-60% faster) 2. There is no need to disable and enable interrupts on the fast path. Currently interrupts have to be disabled and enabled on every slab operation. This is likely avoiding a significant percentage of interrupt off / on sequences in the kernel. 3. The disposal of freed slabs can occur with interrupts enabled. The alternate path is realized using #ifdef's. Several attempts to do the same with macros and inline functions resulted in a mess (in particular due to the strange way that local_interrupt_save() handles its argument and due to the need to define macros/functions that sometimes disable interrupts and sometimes do something else). [clameter: Stripped preempt bits and disabled fastpath if preempt is enabled] Signed-off-by: Christoph Lameter <clameter@sgi.com> Reviewed-by: Pekka Enberg <penberg@cs.helsinki.fi> Cc: <linux-arch@vger.kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2008-01-08 15:20:30 +08:00
s = slab_pre_alloc_hook(s, &objcg, 1, gfpflags);
memcg: fix possible use-after-free in memcg_kmem_get_cache() Suppose task @t that belongs to a memory cgroup @memcg is going to allocate an object from a kmem cache @c. The copy of @c corresponding to @memcg, @mc, is empty. Then if kmem_cache_alloc races with the memory cgroup destruction we can access the memory cgroup's copy of the cache after it was destroyed: CPU0 CPU1 ---- ---- [ current=@t @mc->memcg_params->nr_pages=0 ] kmem_cache_alloc(@c): call memcg_kmem_get_cache(@c); proceed to allocation from @mc: alloc a page for @mc: ... move @t from @memcg destroy @memcg: mem_cgroup_css_offline(@memcg): memcg_unregister_all_caches(@memcg): kmem_cache_destroy(@mc) add page to @mc We could fix this issue by taking a reference to a per-memcg cache, but that would require adding a per-cpu reference counter to per-memcg caches, which would look cumbersome. Instead, let's take a reference to a memory cgroup, which already has a per-cpu reference counter, in the beginning of kmem_cache_alloc to be dropped in the end, and move per memcg caches destruction from css offline to css free. As a side effect, per-memcg caches will be destroyed not one by one, but all at once when the last page accounted to the memory cgroup is freed. This doesn't sound as a high price for code readability though. Note, this patch does add some overhead to the kmem_cache_alloc hot path, but it is pretty negligible - it's just a function call plus a per cpu counter decrement, which is comparable to what we already have in memcg_kmem_get_cache. Besides, it's only relevant if there are memory cgroups with kmem accounting enabled. I don't think we can find a way to handle this race w/o it, because alloc_page called from kmem_cache_alloc may sleep so we can't flush all pending kmallocs w/o reference counting. Signed-off-by: Vladimir Davydov <vdavydov@parallels.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-12-13 08:56:38 +08:00
if (!s)
return NULL;
mm, kfence: insert KFENCE hooks for SLUB Inserts KFENCE hooks into the SLUB allocator. To pass the originally requested size to KFENCE, add an argument 'orig_size' to slab_alloc*(). The additional argument is required to preserve the requested original size for kmalloc() allocations, which uses size classes (e.g. an allocation of 272 bytes will return an object of size 512). Therefore, kmem_cache::size does not represent the kmalloc-caller's requested size, and we must introduce the argument 'orig_size' to propagate the originally requested size to KFENCE. Without the originally requested size, we would not be able to detect out-of-bounds accesses for objects placed at the end of a KFENCE object page if that object is not equal to the kmalloc-size class it was bucketed into. When KFENCE is disabled, there is no additional overhead, since slab_alloc*() functions are __always_inline. Link: https://lkml.kernel.org/r/20201103175841.3495947-6-elver@google.com Signed-off-by: Marco Elver <elver@google.com> Signed-off-by: Alexander Potapenko <glider@google.com> Reviewed-by: Dmitry Vyukov <dvyukov@google.com> Reviewed-by: Jann Horn <jannh@google.com> Co-developed-by: Marco Elver <elver@google.com> Cc: Andrey Konovalov <andreyknvl@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christopher Lameter <cl@linux.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: David Rientjes <rientjes@google.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Hillf Danton <hdanton@sina.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Joern Engel <joern@purestorage.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kees Cook <keescook@chromium.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Paul E. McKenney <paulmck@kernel.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: SeongJae Park <sjpark@amazon.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-26 09:19:16 +08:00
object = kfence_alloc(s, orig_size, gfpflags);
if (unlikely(object))
goto out;
Lockless (and preemptless) fastpaths for slub Use the this_cpu_cmpxchg_double functionality to implement a lockless allocation algorithm on arches that support fast this_cpu_ops. Each of the per cpu pointers is paired with a transaction id that ensures that updates of the per cpu information can only occur in sequence on a certain cpu. A transaction id is a "long" integer that is comprised of an event number and the cpu number. The event number is incremented for every change to the per cpu state. This means that the cmpxchg instruction can verify for an update that nothing interfered and that we are updating the percpu structure for the processor where we picked up the information and that we are also currently on that processor when we update the information. This results in a significant decrease of the overhead in the fastpaths. It also makes it easy to adopt the fast path for realtime kernels since this is lockless and does not require the use of the current per cpu area over the critical section. It is only important that the per cpu area is current at the beginning of the critical section and at the end. So there is no need even to disable preemption. Test results show that the fastpath cycle count is reduced by up to ~ 40% (alloc/free test goes from ~140 cycles down to ~80). The slowpath for kfree adds a few cycles. Sadly this does nothing for the slowpath which is where the main issues with performance in slub are but the best case performance rises significantly. (For that see the more complex slub patches that require cmpxchg_double) Kmalloc: alloc/free test Before: 10000 times kmalloc(8)/kfree -> 134 cycles 10000 times kmalloc(16)/kfree -> 152 cycles 10000 times kmalloc(32)/kfree -> 144 cycles 10000 times kmalloc(64)/kfree -> 142 cycles 10000 times kmalloc(128)/kfree -> 142 cycles 10000 times kmalloc(256)/kfree -> 132 cycles 10000 times kmalloc(512)/kfree -> 132 cycles 10000 times kmalloc(1024)/kfree -> 135 cycles 10000 times kmalloc(2048)/kfree -> 135 cycles 10000 times kmalloc(4096)/kfree -> 135 cycles 10000 times kmalloc(8192)/kfree -> 144 cycles 10000 times kmalloc(16384)/kfree -> 754 cycles After: 10000 times kmalloc(8)/kfree -> 78 cycles 10000 times kmalloc(16)/kfree -> 78 cycles 10000 times kmalloc(32)/kfree -> 82 cycles 10000 times kmalloc(64)/kfree -> 88 cycles 10000 times kmalloc(128)/kfree -> 79 cycles 10000 times kmalloc(256)/kfree -> 79 cycles 10000 times kmalloc(512)/kfree -> 85 cycles 10000 times kmalloc(1024)/kfree -> 82 cycles 10000 times kmalloc(2048)/kfree -> 82 cycles 10000 times kmalloc(4096)/kfree -> 85 cycles 10000 times kmalloc(8192)/kfree -> 82 cycles 10000 times kmalloc(16384)/kfree -> 706 cycles Kmalloc: Repeatedly allocate then free test Before: 10000 times kmalloc(8) -> 211 cycles kfree -> 113 cycles 10000 times kmalloc(16) -> 174 cycles kfree -> 115 cycles 10000 times kmalloc(32) -> 235 cycles kfree -> 129 cycles 10000 times kmalloc(64) -> 222 cycles kfree -> 120 cycles 10000 times kmalloc(128) -> 343 cycles kfree -> 139 cycles 10000 times kmalloc(256) -> 827 cycles kfree -> 147 cycles 10000 times kmalloc(512) -> 1048 cycles kfree -> 272 cycles 10000 times kmalloc(1024) -> 2043 cycles kfree -> 528 cycles 10000 times kmalloc(2048) -> 4002 cycles kfree -> 571 cycles 10000 times kmalloc(4096) -> 7740 cycles kfree -> 628 cycles 10000 times kmalloc(8192) -> 8062 cycles kfree -> 850 cycles 10000 times kmalloc(16384) -> 8895 cycles kfree -> 1249 cycles After: 10000 times kmalloc(8) -> 190 cycles kfree -> 129 cycles 10000 times kmalloc(16) -> 76 cycles kfree -> 123 cycles 10000 times kmalloc(32) -> 126 cycles kfree -> 124 cycles 10000 times kmalloc(64) -> 181 cycles kfree -> 128 cycles 10000 times kmalloc(128) -> 310 cycles kfree -> 140 cycles 10000 times kmalloc(256) -> 809 cycles kfree -> 165 cycles 10000 times kmalloc(512) -> 1005 cycles kfree -> 269 cycles 10000 times kmalloc(1024) -> 1999 cycles kfree -> 527 cycles 10000 times kmalloc(2048) -> 3967 cycles kfree -> 570 cycles 10000 times kmalloc(4096) -> 7658 cycles kfree -> 637 cycles 10000 times kmalloc(8192) -> 8111 cycles kfree -> 859 cycles 10000 times kmalloc(16384) -> 8791 cycles kfree -> 1173 cycles Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-02-26 01:38:54 +08:00
redo:
/*
* Must read kmem_cache cpu data via this cpu ptr. Preemption is
* enabled. We may switch back and forth between cpus while
* reading from one cpu area. That does not matter as long
* as we end up on the original cpu again when doing the cmpxchg.
*
mm/slub: optimize alloc/free fastpath by removing preemption on/off We had to insert a preempt enable/disable in the fastpath a while ago in order to guarantee that tid and kmem_cache_cpu are retrieved on the same cpu. It is the problem only for CONFIG_PREEMPT in which scheduler can move the process to other cpu during retrieving data. Now, I reach the solution to remove preempt enable/disable in the fastpath. If tid is matched with kmem_cache_cpu's tid after tid and kmem_cache_cpu are retrieved by separate this_cpu operation, it means that they are retrieved on the same cpu. If not matched, we just have to retry it. With this guarantee, preemption enable/disable isn't need at all even if CONFIG_PREEMPT, so this patch removes it. I saw roughly 5% win in a fast-path loop over kmem_cache_alloc/free in CONFIG_PREEMPT. (14.821 ns -> 14.049 ns) Below is the result of Christoph's slab_test reported by Jesper Dangaard Brouer. * Before Single thread testing ===================== 1. Kmalloc: Repeatedly allocate then free test 10000 times kmalloc(8) -> 49 cycles kfree -> 62 cycles 10000 times kmalloc(16) -> 48 cycles kfree -> 64 cycles 10000 times kmalloc(32) -> 53 cycles kfree -> 70 cycles 10000 times kmalloc(64) -> 64 cycles kfree -> 77 cycles 10000 times kmalloc(128) -> 74 cycles kfree -> 84 cycles 10000 times kmalloc(256) -> 84 cycles kfree -> 114 cycles 10000 times kmalloc(512) -> 83 cycles kfree -> 116 cycles 10000 times kmalloc(1024) -> 81 cycles kfree -> 120 cycles 10000 times kmalloc(2048) -> 104 cycles kfree -> 136 cycles 10000 times kmalloc(4096) -> 142 cycles kfree -> 165 cycles 10000 times kmalloc(8192) -> 238 cycles kfree -> 226 cycles 10000 times kmalloc(16384) -> 403 cycles kfree -> 264 cycles 2. Kmalloc: alloc/free test 10000 times kmalloc(8)/kfree -> 68 cycles 10000 times kmalloc(16)/kfree -> 68 cycles 10000 times kmalloc(32)/kfree -> 69 cycles 10000 times kmalloc(64)/kfree -> 68 cycles 10000 times kmalloc(128)/kfree -> 68 cycles 10000 times kmalloc(256)/kfree -> 68 cycles 10000 times kmalloc(512)/kfree -> 74 cycles 10000 times kmalloc(1024)/kfree -> 75 cycles 10000 times kmalloc(2048)/kfree -> 74 cycles 10000 times kmalloc(4096)/kfree -> 74 cycles 10000 times kmalloc(8192)/kfree -> 75 cycles 10000 times kmalloc(16384)/kfree -> 510 cycles * After Single thread testing ===================== 1. Kmalloc: Repeatedly allocate then free test 10000 times kmalloc(8) -> 46 cycles kfree -> 61 cycles 10000 times kmalloc(16) -> 46 cycles kfree -> 63 cycles 10000 times kmalloc(32) -> 49 cycles kfree -> 69 cycles 10000 times kmalloc(64) -> 57 cycles kfree -> 76 cycles 10000 times kmalloc(128) -> 66 cycles kfree -> 83 cycles 10000 times kmalloc(256) -> 84 cycles kfree -> 110 cycles 10000 times kmalloc(512) -> 77 cycles kfree -> 114 cycles 10000 times kmalloc(1024) -> 80 cycles kfree -> 116 cycles 10000 times kmalloc(2048) -> 102 cycles kfree -> 131 cycles 10000 times kmalloc(4096) -> 135 cycles kfree -> 163 cycles 10000 times kmalloc(8192) -> 238 cycles kfree -> 218 cycles 10000 times kmalloc(16384) -> 399 cycles kfree -> 262 cycles 2. Kmalloc: alloc/free test 10000 times kmalloc(8)/kfree -> 65 cycles 10000 times kmalloc(16)/kfree -> 66 cycles 10000 times kmalloc(32)/kfree -> 65 cycles 10000 times kmalloc(64)/kfree -> 66 cycles 10000 times kmalloc(128)/kfree -> 66 cycles 10000 times kmalloc(256)/kfree -> 71 cycles 10000 times kmalloc(512)/kfree -> 72 cycles 10000 times kmalloc(1024)/kfree -> 71 cycles 10000 times kmalloc(2048)/kfree -> 71 cycles 10000 times kmalloc(4096)/kfree -> 71 cycles 10000 times kmalloc(8192)/kfree -> 65 cycles 10000 times kmalloc(16384)/kfree -> 511 cycles Most of the results are better than before. Note that this change slightly worses performance in !CONFIG_PREEMPT, roughly 0.3%. Implementing each case separately would help performance, but, since it's so marginal, I didn't do that. This would help maintanance since we have same code for all cases. Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com> Acked-by: Christoph Lameter <cl@linux.com> Tested-by: Jesper Dangaard Brouer <brouer@redhat.com> Acked-by: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 06:09:32 +08:00
* We should guarantee that tid and kmem_cache are retrieved on
* the same cpu. It could be different if CONFIG_PREEMPTION so we need
mm/slub: optimize alloc/free fastpath by removing preemption on/off We had to insert a preempt enable/disable in the fastpath a while ago in order to guarantee that tid and kmem_cache_cpu are retrieved on the same cpu. It is the problem only for CONFIG_PREEMPT in which scheduler can move the process to other cpu during retrieving data. Now, I reach the solution to remove preempt enable/disable in the fastpath. If tid is matched with kmem_cache_cpu's tid after tid and kmem_cache_cpu are retrieved by separate this_cpu operation, it means that they are retrieved on the same cpu. If not matched, we just have to retry it. With this guarantee, preemption enable/disable isn't need at all even if CONFIG_PREEMPT, so this patch removes it. I saw roughly 5% win in a fast-path loop over kmem_cache_alloc/free in CONFIG_PREEMPT. (14.821 ns -> 14.049 ns) Below is the result of Christoph's slab_test reported by Jesper Dangaard Brouer. * Before Single thread testing ===================== 1. Kmalloc: Repeatedly allocate then free test 10000 times kmalloc(8) -> 49 cycles kfree -> 62 cycles 10000 times kmalloc(16) -> 48 cycles kfree -> 64 cycles 10000 times kmalloc(32) -> 53 cycles kfree -> 70 cycles 10000 times kmalloc(64) -> 64 cycles kfree -> 77 cycles 10000 times kmalloc(128) -> 74 cycles kfree -> 84 cycles 10000 times kmalloc(256) -> 84 cycles kfree -> 114 cycles 10000 times kmalloc(512) -> 83 cycles kfree -> 116 cycles 10000 times kmalloc(1024) -> 81 cycles kfree -> 120 cycles 10000 times kmalloc(2048) -> 104 cycles kfree -> 136 cycles 10000 times kmalloc(4096) -> 142 cycles kfree -> 165 cycles 10000 times kmalloc(8192) -> 238 cycles kfree -> 226 cycles 10000 times kmalloc(16384) -> 403 cycles kfree -> 264 cycles 2. Kmalloc: alloc/free test 10000 times kmalloc(8)/kfree -> 68 cycles 10000 times kmalloc(16)/kfree -> 68 cycles 10000 times kmalloc(32)/kfree -> 69 cycles 10000 times kmalloc(64)/kfree -> 68 cycles 10000 times kmalloc(128)/kfree -> 68 cycles 10000 times kmalloc(256)/kfree -> 68 cycles 10000 times kmalloc(512)/kfree -> 74 cycles 10000 times kmalloc(1024)/kfree -> 75 cycles 10000 times kmalloc(2048)/kfree -> 74 cycles 10000 times kmalloc(4096)/kfree -> 74 cycles 10000 times kmalloc(8192)/kfree -> 75 cycles 10000 times kmalloc(16384)/kfree -> 510 cycles * After Single thread testing ===================== 1. Kmalloc: Repeatedly allocate then free test 10000 times kmalloc(8) -> 46 cycles kfree -> 61 cycles 10000 times kmalloc(16) -> 46 cycles kfree -> 63 cycles 10000 times kmalloc(32) -> 49 cycles kfree -> 69 cycles 10000 times kmalloc(64) -> 57 cycles kfree -> 76 cycles 10000 times kmalloc(128) -> 66 cycles kfree -> 83 cycles 10000 times kmalloc(256) -> 84 cycles kfree -> 110 cycles 10000 times kmalloc(512) -> 77 cycles kfree -> 114 cycles 10000 times kmalloc(1024) -> 80 cycles kfree -> 116 cycles 10000 times kmalloc(2048) -> 102 cycles kfree -> 131 cycles 10000 times kmalloc(4096) -> 135 cycles kfree -> 163 cycles 10000 times kmalloc(8192) -> 238 cycles kfree -> 218 cycles 10000 times kmalloc(16384) -> 399 cycles kfree -> 262 cycles 2. Kmalloc: alloc/free test 10000 times kmalloc(8)/kfree -> 65 cycles 10000 times kmalloc(16)/kfree -> 66 cycles 10000 times kmalloc(32)/kfree -> 65 cycles 10000 times kmalloc(64)/kfree -> 66 cycles 10000 times kmalloc(128)/kfree -> 66 cycles 10000 times kmalloc(256)/kfree -> 71 cycles 10000 times kmalloc(512)/kfree -> 72 cycles 10000 times kmalloc(1024)/kfree -> 71 cycles 10000 times kmalloc(2048)/kfree -> 71 cycles 10000 times kmalloc(4096)/kfree -> 71 cycles 10000 times kmalloc(8192)/kfree -> 65 cycles 10000 times kmalloc(16384)/kfree -> 511 cycles Most of the results are better than before. Note that this change slightly worses performance in !CONFIG_PREEMPT, roughly 0.3%. Implementing each case separately would help performance, but, since it's so marginal, I didn't do that. This would help maintanance since we have same code for all cases. Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com> Acked-by: Christoph Lameter <cl@linux.com> Tested-by: Jesper Dangaard Brouer <brouer@redhat.com> Acked-by: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 06:09:32 +08:00
* to check if it is matched or not.
Lockless (and preemptless) fastpaths for slub Use the this_cpu_cmpxchg_double functionality to implement a lockless allocation algorithm on arches that support fast this_cpu_ops. Each of the per cpu pointers is paired with a transaction id that ensures that updates of the per cpu information can only occur in sequence on a certain cpu. A transaction id is a "long" integer that is comprised of an event number and the cpu number. The event number is incremented for every change to the per cpu state. This means that the cmpxchg instruction can verify for an update that nothing interfered and that we are updating the percpu structure for the processor where we picked up the information and that we are also currently on that processor when we update the information. This results in a significant decrease of the overhead in the fastpaths. It also makes it easy to adopt the fast path for realtime kernels since this is lockless and does not require the use of the current per cpu area over the critical section. It is only important that the per cpu area is current at the beginning of the critical section and at the end. So there is no need even to disable preemption. Test results show that the fastpath cycle count is reduced by up to ~ 40% (alloc/free test goes from ~140 cycles down to ~80). The slowpath for kfree adds a few cycles. Sadly this does nothing for the slowpath which is where the main issues with performance in slub are but the best case performance rises significantly. (For that see the more complex slub patches that require cmpxchg_double) Kmalloc: alloc/free test Before: 10000 times kmalloc(8)/kfree -> 134 cycles 10000 times kmalloc(16)/kfree -> 152 cycles 10000 times kmalloc(32)/kfree -> 144 cycles 10000 times kmalloc(64)/kfree -> 142 cycles 10000 times kmalloc(128)/kfree -> 142 cycles 10000 times kmalloc(256)/kfree -> 132 cycles 10000 times kmalloc(512)/kfree -> 132 cycles 10000 times kmalloc(1024)/kfree -> 135 cycles 10000 times kmalloc(2048)/kfree -> 135 cycles 10000 times kmalloc(4096)/kfree -> 135 cycles 10000 times kmalloc(8192)/kfree -> 144 cycles 10000 times kmalloc(16384)/kfree -> 754 cycles After: 10000 times kmalloc(8)/kfree -> 78 cycles 10000 times kmalloc(16)/kfree -> 78 cycles 10000 times kmalloc(32)/kfree -> 82 cycles 10000 times kmalloc(64)/kfree -> 88 cycles 10000 times kmalloc(128)/kfree -> 79 cycles 10000 times kmalloc(256)/kfree -> 79 cycles 10000 times kmalloc(512)/kfree -> 85 cycles 10000 times kmalloc(1024)/kfree -> 82 cycles 10000 times kmalloc(2048)/kfree -> 82 cycles 10000 times kmalloc(4096)/kfree -> 85 cycles 10000 times kmalloc(8192)/kfree -> 82 cycles 10000 times kmalloc(16384)/kfree -> 706 cycles Kmalloc: Repeatedly allocate then free test Before: 10000 times kmalloc(8) -> 211 cycles kfree -> 113 cycles 10000 times kmalloc(16) -> 174 cycles kfree -> 115 cycles 10000 times kmalloc(32) -> 235 cycles kfree -> 129 cycles 10000 times kmalloc(64) -> 222 cycles kfree -> 120 cycles 10000 times kmalloc(128) -> 343 cycles kfree -> 139 cycles 10000 times kmalloc(256) -> 827 cycles kfree -> 147 cycles 10000 times kmalloc(512) -> 1048 cycles kfree -> 272 cycles 10000 times kmalloc(1024) -> 2043 cycles kfree -> 528 cycles 10000 times kmalloc(2048) -> 4002 cycles kfree -> 571 cycles 10000 times kmalloc(4096) -> 7740 cycles kfree -> 628 cycles 10000 times kmalloc(8192) -> 8062 cycles kfree -> 850 cycles 10000 times kmalloc(16384) -> 8895 cycles kfree -> 1249 cycles After: 10000 times kmalloc(8) -> 190 cycles kfree -> 129 cycles 10000 times kmalloc(16) -> 76 cycles kfree -> 123 cycles 10000 times kmalloc(32) -> 126 cycles kfree -> 124 cycles 10000 times kmalloc(64) -> 181 cycles kfree -> 128 cycles 10000 times kmalloc(128) -> 310 cycles kfree -> 140 cycles 10000 times kmalloc(256) -> 809 cycles kfree -> 165 cycles 10000 times kmalloc(512) -> 1005 cycles kfree -> 269 cycles 10000 times kmalloc(1024) -> 1999 cycles kfree -> 527 cycles 10000 times kmalloc(2048) -> 3967 cycles kfree -> 570 cycles 10000 times kmalloc(4096) -> 7658 cycles kfree -> 637 cycles 10000 times kmalloc(8192) -> 8111 cycles kfree -> 859 cycles 10000 times kmalloc(16384) -> 8791 cycles kfree -> 1173 cycles Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-02-26 01:38:54 +08:00
*/
mm/slub: optimize alloc/free fastpath by removing preemption on/off We had to insert a preempt enable/disable in the fastpath a while ago in order to guarantee that tid and kmem_cache_cpu are retrieved on the same cpu. It is the problem only for CONFIG_PREEMPT in which scheduler can move the process to other cpu during retrieving data. Now, I reach the solution to remove preempt enable/disable in the fastpath. If tid is matched with kmem_cache_cpu's tid after tid and kmem_cache_cpu are retrieved by separate this_cpu operation, it means that they are retrieved on the same cpu. If not matched, we just have to retry it. With this guarantee, preemption enable/disable isn't need at all even if CONFIG_PREEMPT, so this patch removes it. I saw roughly 5% win in a fast-path loop over kmem_cache_alloc/free in CONFIG_PREEMPT. (14.821 ns -> 14.049 ns) Below is the result of Christoph's slab_test reported by Jesper Dangaard Brouer. * Before Single thread testing ===================== 1. Kmalloc: Repeatedly allocate then free test 10000 times kmalloc(8) -> 49 cycles kfree -> 62 cycles 10000 times kmalloc(16) -> 48 cycles kfree -> 64 cycles 10000 times kmalloc(32) -> 53 cycles kfree -> 70 cycles 10000 times kmalloc(64) -> 64 cycles kfree -> 77 cycles 10000 times kmalloc(128) -> 74 cycles kfree -> 84 cycles 10000 times kmalloc(256) -> 84 cycles kfree -> 114 cycles 10000 times kmalloc(512) -> 83 cycles kfree -> 116 cycles 10000 times kmalloc(1024) -> 81 cycles kfree -> 120 cycles 10000 times kmalloc(2048) -> 104 cycles kfree -> 136 cycles 10000 times kmalloc(4096) -> 142 cycles kfree -> 165 cycles 10000 times kmalloc(8192) -> 238 cycles kfree -> 226 cycles 10000 times kmalloc(16384) -> 403 cycles kfree -> 264 cycles 2. Kmalloc: alloc/free test 10000 times kmalloc(8)/kfree -> 68 cycles 10000 times kmalloc(16)/kfree -> 68 cycles 10000 times kmalloc(32)/kfree -> 69 cycles 10000 times kmalloc(64)/kfree -> 68 cycles 10000 times kmalloc(128)/kfree -> 68 cycles 10000 times kmalloc(256)/kfree -> 68 cycles 10000 times kmalloc(512)/kfree -> 74 cycles 10000 times kmalloc(1024)/kfree -> 75 cycles 10000 times kmalloc(2048)/kfree -> 74 cycles 10000 times kmalloc(4096)/kfree -> 74 cycles 10000 times kmalloc(8192)/kfree -> 75 cycles 10000 times kmalloc(16384)/kfree -> 510 cycles * After Single thread testing ===================== 1. Kmalloc: Repeatedly allocate then free test 10000 times kmalloc(8) -> 46 cycles kfree -> 61 cycles 10000 times kmalloc(16) -> 46 cycles kfree -> 63 cycles 10000 times kmalloc(32) -> 49 cycles kfree -> 69 cycles 10000 times kmalloc(64) -> 57 cycles kfree -> 76 cycles 10000 times kmalloc(128) -> 66 cycles kfree -> 83 cycles 10000 times kmalloc(256) -> 84 cycles kfree -> 110 cycles 10000 times kmalloc(512) -> 77 cycles kfree -> 114 cycles 10000 times kmalloc(1024) -> 80 cycles kfree -> 116 cycles 10000 times kmalloc(2048) -> 102 cycles kfree -> 131 cycles 10000 times kmalloc(4096) -> 135 cycles kfree -> 163 cycles 10000 times kmalloc(8192) -> 238 cycles kfree -> 218 cycles 10000 times kmalloc(16384) -> 399 cycles kfree -> 262 cycles 2. Kmalloc: alloc/free test 10000 times kmalloc(8)/kfree -> 65 cycles 10000 times kmalloc(16)/kfree -> 66 cycles 10000 times kmalloc(32)/kfree -> 65 cycles 10000 times kmalloc(64)/kfree -> 66 cycles 10000 times kmalloc(128)/kfree -> 66 cycles 10000 times kmalloc(256)/kfree -> 71 cycles 10000 times kmalloc(512)/kfree -> 72 cycles 10000 times kmalloc(1024)/kfree -> 71 cycles 10000 times kmalloc(2048)/kfree -> 71 cycles 10000 times kmalloc(4096)/kfree -> 71 cycles 10000 times kmalloc(8192)/kfree -> 65 cycles 10000 times kmalloc(16384)/kfree -> 511 cycles Most of the results are better than before. Note that this change slightly worses performance in !CONFIG_PREEMPT, roughly 0.3%. Implementing each case separately would help performance, but, since it's so marginal, I didn't do that. This would help maintanance since we have same code for all cases. Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com> Acked-by: Christoph Lameter <cl@linux.com> Tested-by: Jesper Dangaard Brouer <brouer@redhat.com> Acked-by: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 06:09:32 +08:00
do {
tid = this_cpu_read(s->cpu_slab->tid);
c = raw_cpu_ptr(s->cpu_slab);
} while (IS_ENABLED(CONFIG_PREEMPTION) &&
mm/slub: fix lockups on PREEMPT && !SMP kernels Commit 9aabf810a67c ("mm/slub: optimize alloc/free fastpath by removing preemption on/off") introduced an occasional hang for kernels built with CONFIG_PREEMPT && !CONFIG_SMP. The problem is the following loop the patch introduced to slab_alloc_node and slab_free: do { tid = this_cpu_read(s->cpu_slab->tid); c = raw_cpu_ptr(s->cpu_slab); } while (IS_ENABLED(CONFIG_PREEMPT) && unlikely(tid != c->tid)); GCC 4.9 has been observed to hoist the load of c and c->tid above the loop for !SMP kernels (as in this case raw_cpu_ptr(x) is compile-time constant and does not force a reload). On arm64 the generated assembly looks like: ldr x4, [x0,#8] loop: ldr x1, [x0,#8] cmp x1, x4 b.ne loop If the thread is preempted between the load of c->tid (into x1) and tid (into x4), and an allocation or free occurs in another thread (bumping the cpu_slab's tid), the thread will be stuck in the loop until s->cpu_slab->tid wraps, which may be forever in the absence of allocations/frees on the same CPU. This patch changes the loop condition to access c->tid with READ_ONCE. This ensures that the value is reloaded even when the compiler would otherwise assume it could cache the value, and also ensures that the load will not be torn. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: David Rientjes <rientjes@google.com> Cc: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: Steve Capper <steve.capper@linaro.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-03-26 06:55:23 +08:00
unlikely(tid != READ_ONCE(c->tid)));
mm/slub: optimize alloc/free fastpath by removing preemption on/off We had to insert a preempt enable/disable in the fastpath a while ago in order to guarantee that tid and kmem_cache_cpu are retrieved on the same cpu. It is the problem only for CONFIG_PREEMPT in which scheduler can move the process to other cpu during retrieving data. Now, I reach the solution to remove preempt enable/disable in the fastpath. If tid is matched with kmem_cache_cpu's tid after tid and kmem_cache_cpu are retrieved by separate this_cpu operation, it means that they are retrieved on the same cpu. If not matched, we just have to retry it. With this guarantee, preemption enable/disable isn't need at all even if CONFIG_PREEMPT, so this patch removes it. I saw roughly 5% win in a fast-path loop over kmem_cache_alloc/free in CONFIG_PREEMPT. (14.821 ns -> 14.049 ns) Below is the result of Christoph's slab_test reported by Jesper Dangaard Brouer. * Before Single thread testing ===================== 1. Kmalloc: Repeatedly allocate then free test 10000 times kmalloc(8) -> 49 cycles kfree -> 62 cycles 10000 times kmalloc(16) -> 48 cycles kfree -> 64 cycles 10000 times kmalloc(32) -> 53 cycles kfree -> 70 cycles 10000 times kmalloc(64) -> 64 cycles kfree -> 77 cycles 10000 times kmalloc(128) -> 74 cycles kfree -> 84 cycles 10000 times kmalloc(256) -> 84 cycles kfree -> 114 cycles 10000 times kmalloc(512) -> 83 cycles kfree -> 116 cycles 10000 times kmalloc(1024) -> 81 cycles kfree -> 120 cycles 10000 times kmalloc(2048) -> 104 cycles kfree -> 136 cycles 10000 times kmalloc(4096) -> 142 cycles kfree -> 165 cycles 10000 times kmalloc(8192) -> 238 cycles kfree -> 226 cycles 10000 times kmalloc(16384) -> 403 cycles kfree -> 264 cycles 2. Kmalloc: alloc/free test 10000 times kmalloc(8)/kfree -> 68 cycles 10000 times kmalloc(16)/kfree -> 68 cycles 10000 times kmalloc(32)/kfree -> 69 cycles 10000 times kmalloc(64)/kfree -> 68 cycles 10000 times kmalloc(128)/kfree -> 68 cycles 10000 times kmalloc(256)/kfree -> 68 cycles 10000 times kmalloc(512)/kfree -> 74 cycles 10000 times kmalloc(1024)/kfree -> 75 cycles 10000 times kmalloc(2048)/kfree -> 74 cycles 10000 times kmalloc(4096)/kfree -> 74 cycles 10000 times kmalloc(8192)/kfree -> 75 cycles 10000 times kmalloc(16384)/kfree -> 510 cycles * After Single thread testing ===================== 1. Kmalloc: Repeatedly allocate then free test 10000 times kmalloc(8) -> 46 cycles kfree -> 61 cycles 10000 times kmalloc(16) -> 46 cycles kfree -> 63 cycles 10000 times kmalloc(32) -> 49 cycles kfree -> 69 cycles 10000 times kmalloc(64) -> 57 cycles kfree -> 76 cycles 10000 times kmalloc(128) -> 66 cycles kfree -> 83 cycles 10000 times kmalloc(256) -> 84 cycles kfree -> 110 cycles 10000 times kmalloc(512) -> 77 cycles kfree -> 114 cycles 10000 times kmalloc(1024) -> 80 cycles kfree -> 116 cycles 10000 times kmalloc(2048) -> 102 cycles kfree -> 131 cycles 10000 times kmalloc(4096) -> 135 cycles kfree -> 163 cycles 10000 times kmalloc(8192) -> 238 cycles kfree -> 218 cycles 10000 times kmalloc(16384) -> 399 cycles kfree -> 262 cycles 2. Kmalloc: alloc/free test 10000 times kmalloc(8)/kfree -> 65 cycles 10000 times kmalloc(16)/kfree -> 66 cycles 10000 times kmalloc(32)/kfree -> 65 cycles 10000 times kmalloc(64)/kfree -> 66 cycles 10000 times kmalloc(128)/kfree -> 66 cycles 10000 times kmalloc(256)/kfree -> 71 cycles 10000 times kmalloc(512)/kfree -> 72 cycles 10000 times kmalloc(1024)/kfree -> 71 cycles 10000 times kmalloc(2048)/kfree -> 71 cycles 10000 times kmalloc(4096)/kfree -> 71 cycles 10000 times kmalloc(8192)/kfree -> 65 cycles 10000 times kmalloc(16384)/kfree -> 511 cycles Most of the results are better than before. Note that this change slightly worses performance in !CONFIG_PREEMPT, roughly 0.3%. Implementing each case separately would help performance, but, since it's so marginal, I didn't do that. This would help maintanance since we have same code for all cases. Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com> Acked-by: Christoph Lameter <cl@linux.com> Tested-by: Jesper Dangaard Brouer <brouer@redhat.com> Acked-by: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 06:09:32 +08:00
/*
* Irqless object alloc/free algorithm used here depends on sequence
* of fetching cpu_slab's data. tid should be fetched before anything
* on c to guarantee that object and page associated with previous tid
* won't be used with current tid. If we fetch tid first, object and
* page could be one associated with next tid and our alloc/free
* request will be failed. In this case, we will retry. So, no problem.
*/
barrier();
Lockless (and preemptless) fastpaths for slub Use the this_cpu_cmpxchg_double functionality to implement a lockless allocation algorithm on arches that support fast this_cpu_ops. Each of the per cpu pointers is paired with a transaction id that ensures that updates of the per cpu information can only occur in sequence on a certain cpu. A transaction id is a "long" integer that is comprised of an event number and the cpu number. The event number is incremented for every change to the per cpu state. This means that the cmpxchg instruction can verify for an update that nothing interfered and that we are updating the percpu structure for the processor where we picked up the information and that we are also currently on that processor when we update the information. This results in a significant decrease of the overhead in the fastpaths. It also makes it easy to adopt the fast path for realtime kernels since this is lockless and does not require the use of the current per cpu area over the critical section. It is only important that the per cpu area is current at the beginning of the critical section and at the end. So there is no need even to disable preemption. Test results show that the fastpath cycle count is reduced by up to ~ 40% (alloc/free test goes from ~140 cycles down to ~80). The slowpath for kfree adds a few cycles. Sadly this does nothing for the slowpath which is where the main issues with performance in slub are but the best case performance rises significantly. (For that see the more complex slub patches that require cmpxchg_double) Kmalloc: alloc/free test Before: 10000 times kmalloc(8)/kfree -> 134 cycles 10000 times kmalloc(16)/kfree -> 152 cycles 10000 times kmalloc(32)/kfree -> 144 cycles 10000 times kmalloc(64)/kfree -> 142 cycles 10000 times kmalloc(128)/kfree -> 142 cycles 10000 times kmalloc(256)/kfree -> 132 cycles 10000 times kmalloc(512)/kfree -> 132 cycles 10000 times kmalloc(1024)/kfree -> 135 cycles 10000 times kmalloc(2048)/kfree -> 135 cycles 10000 times kmalloc(4096)/kfree -> 135 cycles 10000 times kmalloc(8192)/kfree -> 144 cycles 10000 times kmalloc(16384)/kfree -> 754 cycles After: 10000 times kmalloc(8)/kfree -> 78 cycles 10000 times kmalloc(16)/kfree -> 78 cycles 10000 times kmalloc(32)/kfree -> 82 cycles 10000 times kmalloc(64)/kfree -> 88 cycles 10000 times kmalloc(128)/kfree -> 79 cycles 10000 times kmalloc(256)/kfree -> 79 cycles 10000 times kmalloc(512)/kfree -> 85 cycles 10000 times kmalloc(1024)/kfree -> 82 cycles 10000 times kmalloc(2048)/kfree -> 82 cycles 10000 times kmalloc(4096)/kfree -> 85 cycles 10000 times kmalloc(8192)/kfree -> 82 cycles 10000 times kmalloc(16384)/kfree -> 706 cycles Kmalloc: Repeatedly allocate then free test Before: 10000 times kmalloc(8) -> 211 cycles kfree -> 113 cycles 10000 times kmalloc(16) -> 174 cycles kfree -> 115 cycles 10000 times kmalloc(32) -> 235 cycles kfree -> 129 cycles 10000 times kmalloc(64) -> 222 cycles kfree -> 120 cycles 10000 times kmalloc(128) -> 343 cycles kfree -> 139 cycles 10000 times kmalloc(256) -> 827 cycles kfree -> 147 cycles 10000 times kmalloc(512) -> 1048 cycles kfree -> 272 cycles 10000 times kmalloc(1024) -> 2043 cycles kfree -> 528 cycles 10000 times kmalloc(2048) -> 4002 cycles kfree -> 571 cycles 10000 times kmalloc(4096) -> 7740 cycles kfree -> 628 cycles 10000 times kmalloc(8192) -> 8062 cycles kfree -> 850 cycles 10000 times kmalloc(16384) -> 8895 cycles kfree -> 1249 cycles After: 10000 times kmalloc(8) -> 190 cycles kfree -> 129 cycles 10000 times kmalloc(16) -> 76 cycles kfree -> 123 cycles 10000 times kmalloc(32) -> 126 cycles kfree -> 124 cycles 10000 times kmalloc(64) -> 181 cycles kfree -> 128 cycles 10000 times kmalloc(128) -> 310 cycles kfree -> 140 cycles 10000 times kmalloc(256) -> 809 cycles kfree -> 165 cycles 10000 times kmalloc(512) -> 1005 cycles kfree -> 269 cycles 10000 times kmalloc(1024) -> 1999 cycles kfree -> 527 cycles 10000 times kmalloc(2048) -> 3967 cycles kfree -> 570 cycles 10000 times kmalloc(4096) -> 7658 cycles kfree -> 637 cycles 10000 times kmalloc(8192) -> 8111 cycles kfree -> 859 cycles 10000 times kmalloc(16384) -> 8791 cycles kfree -> 1173 cycles Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-02-26 01:38:54 +08:00
/*
* The transaction ids are globally unique per cpu and per operation on
* a per cpu queue. Thus they can be guarantee that the cmpxchg_double
* occurs on the right processor and that there was no operation on the
* linked list in between.
*/
object = c->freelist;
page = c->page;
mm/slub: fix panic in slab_alloc_node() While doing memory hot-unplug operation on a PowerPC VM running 1024 CPUs with 11TB of ram, I hit the following panic: BUG: Kernel NULL pointer dereference on read at 0x00000007 Faulting instruction address: 0xc000000000456048 Oops: Kernel access of bad area, sig: 11 [#2] LE PAGE_SIZE=64K MMU=Hash SMP NR_CPUS= 2048 NUMA pSeries Modules linked in: rpadlpar_io rpaphp CPU: 160 PID: 1 Comm: systemd Tainted: G D 5.9.0 #1 NIP: c000000000456048 LR: c000000000455fd4 CTR: c00000000047b350 REGS: c00006028d1b77a0 TRAP: 0300 Tainted: G D (5.9.0) MSR: 8000000000009033 <SF,EE,ME,IR,DR,RI,LE> CR: 24004228 XER: 00000000 CFAR: c00000000000f1b0 DAR: 0000000000000007 DSISR: 40000000 IRQMASK: 0 GPR00: c000000000455fd4 c00006028d1b7a30 c000000001bec800 0000000000000000 GPR04: 0000000000000dc0 0000000000000000 00000000000374ef c00007c53df99320 GPR08: 000007c53c980000 0000000000000000 000007c53c980000 0000000000000000 GPR12: 0000000000004400 c00000001e8e4400 0000000000000000 0000000000000f6a GPR16: 0000000000000000 c000000001c25930 c000000001d62528 00000000000000c1 GPR20: c000000001d62538 c00006be469e9000 0000000fffffffe0 c0000000003c0ff8 GPR24: 0000000000000018 0000000000000000 0000000000000dc0 0000000000000000 GPR28: c00007c513755700 c000000001c236a4 c00007bc4001f800 0000000000000001 NIP [c000000000456048] __kmalloc_node+0x108/0x790 LR [c000000000455fd4] __kmalloc_node+0x94/0x790 Call Trace: kvmalloc_node+0x58/0x110 mem_cgroup_css_online+0x10c/0x270 online_css+0x48/0xd0 cgroup_apply_control_enable+0x2c4/0x470 cgroup_mkdir+0x408/0x5f0 kernfs_iop_mkdir+0x90/0x100 vfs_mkdir+0x138/0x250 do_mkdirat+0x154/0x1c0 system_call_exception+0xf8/0x200 system_call_common+0xf0/0x27c Instruction dump: e93e0000 e90d0030 39290008 7cc9402a e94d0030 e93e0000 7ce95214 7f89502a 2fbc0000 419e0018 41920230 e9270010 <89290007> 7f994800 419e0220 7ee6bb78 This pointing to the following code: mm/slub.c:2851 if (unlikely(!object || !node_match(page, node))) { c000000000456038: 00 00 bc 2f cmpdi cr7,r28,0 c00000000045603c: 18 00 9e 41 beq cr7,c000000000456054 <__kmalloc_node+0x114> node_match(): mm/slub.c:2491 if (node != NUMA_NO_NODE && page_to_nid(page) != node) c000000000456040: 30 02 92 41 beq cr4,c000000000456270 <__kmalloc_node+0x330> page_to_nid(): include/linux/mm.h:1294 c000000000456044: 10 00 27 e9 ld r9,16(r7) c000000000456048: 07 00 29 89 lbz r9,7(r9) <<<< r9 = NULL node_match(): mm/slub.c:2491 c00000000045604c: 00 48 99 7f cmpw cr7,r25,r9 c000000000456050: 20 02 9e 41 beq cr7,c000000000456270 <__kmalloc_node+0x330> The panic occurred in slab_alloc_node() when checking for the page's node: object = c->freelist; page = c->page; if (unlikely(!object || !node_match(page, node))) { object = __slab_alloc(s, gfpflags, node, addr, c); stat(s, ALLOC_SLOWPATH); The issue is that object is not NULL while page is NULL which is odd but may happen if the cache flush happened after loading object but before loading page. Thus checking for the page pointer is required too. The cache flush is done through an inter processor interrupt when a piece of memory is off-lined. That interrupt is triggered when a memory hot-unplug operation is initiated and offline_pages() is calling the slub's MEM_GOING_OFFLINE callback slab_mem_going_offline_callback() which is calling flush_cpu_slab(). If that interrupt is caught between the reading of c->freelist and the reading of c->page, this could lead to such a situation. That situation is expected and the later call to this_cpu_cmpxchg_double() will detect the change to c->freelist and redo the whole operation. In commit 6159d0f5c03e ("mm/slub.c: page is always non-NULL in node_match()") check on the page pointer has been removed assuming that page is always valid when it is called. It happens that this is not true in that particular case, so check for page before calling node_match() here. Fixes: 6159d0f5c03e ("mm/slub.c: page is always non-NULL in node_match()") Signed-off-by: Laurent Dufour <ldufour@linux.ibm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Acked-by: Vlastimil Babka <vbabka@suse.cz> Acked-by: Christoph Lameter <cl@linux.com> Cc: Wei Yang <richard.weiyang@gmail.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Nathan Lynch <nathanl@linux.ibm.com> Cc: Scott Cheloha <cheloha@linux.ibm.com> Cc: Michal Hocko <mhocko@suse.com> Cc: <stable@vger.kernel.org> Link: https://lkml.kernel.org/r/20201027190406.33283-1-ldufour@linux.ibm.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-11-14 14:51:53 +08:00
if (unlikely(!object || !page || !node_match(page, node))) {
SLUB: Avoid page struct cacheline bouncing due to remote frees to cpu slab A remote free may access the same page struct that also contains the lockless freelist for the cpu slab. If objects have a short lifetime and are freed by a different processor then remote frees back to the slab from which we are currently allocating are frequent. The cacheline with the page struct needs to be repeately acquired in exclusive mode by both the allocating thread and the freeing thread. If this is frequent enough then performance will suffer because of cacheline bouncing. This patchset puts the lockless_freelist pointer in its own cacheline. In order to make that happen we introduce a per cpu structure called kmem_cache_cpu. Instead of keeping an array of pointers to page structs we now keep an array to a per cpu structure that--among other things--contains the pointer to the lockless freelist. The freeing thread can then keep possession of exclusive access to the page struct cacheline while the allocating thread keeps its exclusive access to the cacheline containing the per cpu structure. This works as long as the allocating cpu is able to service its request from the lockless freelist. If the lockless freelist runs empty then the allocating thread needs to acquire exclusive access to the cacheline with the page struct lock the slab. The allocating thread will then check if new objects were freed to the per cpu slab. If so it will keep the slab as the cpu slab and continue with the recently remote freed objects. So the allocating thread can take a series of just freed remote pages and dish them out again. Ideally allocations could be just recycling objects in the same slab this way which will lead to an ideal allocation / remote free pattern. The number of objects that can be handled in this way is limited by the capacity of one slab. Increasing slab size via slub_min_objects/ slub_max_order may increase the number of objects and therefore performance. If the allocating thread runs out of objects and finds that no objects were put back by the remote processor then it will retrieve a new slab (from the partial lists or from the page allocator) and start with a whole new set of objects while the remote thread may still be freeing objects to the old cpu slab. This may then repeat until the new slab is also exhausted. If remote freeing has freed objects in the earlier slab then that earlier slab will now be on the partial freelist and the allocating thread will pick that slab next for allocation. So the loop is extended. However, both threads need to take the list_lock to make the swizzling via the partial list happen. It is likely that this kind of scheme will keep the objects being passed around to a small set that can be kept in the cpu caches leading to increased performance. More code cleanups become possible: - Instead of passing a cpu we can now pass a kmem_cache_cpu structure around. Allows reducing the number of parameters to various functions. - Can define a new node_match() function for NUMA to encapsulate locality checks. Effect on allocations: Cachelines touched before this patch: Write: page cache struct and first cacheline of object Cachelines touched after this patch: Write: kmem_cache_cpu cacheline and first cacheline of object Read: page cache struct (but see later patch that avoids touching that cacheline) The handling when the lockless alloc list runs empty gets to be a bit more complicated since another cacheline has now to be written to. But that is halfway out of the hot path. Effect on freeing: Cachelines touched before this patch: Write: page_struct and first cacheline of object Cachelines touched after this patch depending on how we free: Write(to cpu_slab): kmem_cache_cpu struct and first cacheline of object Write(to other): page struct and first cacheline of object Read(to cpu_slab): page struct to id slab etc. (but see later patch that avoids touching the page struct on free) Read(to other): cpu local kmem_cache_cpu struct to verify its not the cpu slab. Summary: Pro: - Distinct cachelines so that concurrent remote frees and local allocs on a cpuslab can occur without cacheline bouncing. - Avoids potential bouncing cachelines because of neighboring per cpu pointer updates in kmem_cache's cpu_slab structure since it now grows to a cacheline (Therefore remove the comment that talks about that concern). Cons: - Freeing objects now requires the reading of one additional cacheline. That can be mitigated for some cases by the following patches but its not possible to completely eliminate these references. - Memory usage grows slightly. The size of each per cpu object is blown up from one word (pointing to the page_struct) to one cacheline with various data. So this is NR_CPUS*NR_SLABS*L1_BYTES more memory use. Lets say NR_SLABS is 100 and a cache line size of 128 then we have just increased SLAB metadata requirements by 12.8k per cpu. (Another later patch reduces these requirements) Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 16:26:05 +08:00
object = __slab_alloc(s, gfpflags, node, addr, c);
} else {
slub: prefetch next freelist pointer in slab_alloc() Recycling a page is a problem, since freelist link chain is hot on cpu(s) which freed objects, and possibly very cold on cpu currently owning slab. Adding a prefetch of cache line containing the pointer to next object in slab_alloc() helps a lot in many workloads, in particular on assymetric ones (allocations done on one cpu, frees on another cpus). Added cost is three machine instructions only. Examples on my dual socket quad core ht machine (Intel CPU E5540 @2.53GHz) (16 logical cpus, 2 memory nodes), 64bit kernel. Before patch : # perf stat -r 32 hackbench 50 process 4000 >/dev/null Performance counter stats for 'hackbench 50 process 4000' (32 runs): 327577,471718 task-clock # 15,821 CPUs utilized ( +- 0,64% ) 28 866 491 context-switches # 0,088 M/sec ( +- 1,80% ) 1 506 929 CPU-migrations # 0,005 M/sec ( +- 3,24% ) 127 151 page-faults # 0,000 M/sec ( +- 0,16% ) 829 399 813 448 cycles # 2,532 GHz ( +- 0,64% ) 580 664 691 740 stalled-cycles-frontend # 70,01% frontend cycles idle ( +- 0,71% ) 197 431 700 448 stalled-cycles-backend # 23,80% backend cycles idle ( +- 1,03% ) 503 548 648 975 instructions # 0,61 insns per cycle # 1,15 stalled cycles per insn ( +- 0,46% ) 95 780 068 471 branches # 292,389 M/sec ( +- 0,48% ) 1 426 407 916 branch-misses # 1,49% of all branches ( +- 1,35% ) 20,705679994 seconds time elapsed ( +- 0,64% ) After patch : # perf stat -r 32 hackbench 50 process 4000 >/dev/null Performance counter stats for 'hackbench 50 process 4000' (32 runs): 286236,542804 task-clock # 15,786 CPUs utilized ( +- 1,32% ) 19 703 372 context-switches # 0,069 M/sec ( +- 4,99% ) 1 658 249 CPU-migrations # 0,006 M/sec ( +- 6,62% ) 126 776 page-faults # 0,000 M/sec ( +- 0,12% ) 724 636 593 213 cycles # 2,532 GHz ( +- 1,32% ) 499 320 714 837 stalled-cycles-frontend # 68,91% frontend cycles idle ( +- 1,47% ) 156 555 126 809 stalled-cycles-backend # 21,60% backend cycles idle ( +- 2,22% ) 463 897 792 661 instructions # 0,64 insns per cycle # 1,08 stalled cycles per insn ( +- 0,94% ) 87 717 352 563 branches # 306,451 M/sec ( +- 0,99% ) 941 738 280 branch-misses # 1,07% of all branches ( +- 3,35% ) 18,132070670 seconds time elapsed ( +- 1,30% ) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Acked-by: Christoph Lameter <cl@linux.com> CC: Matt Mackall <mpm@selenic.com> CC: David Rientjes <rientjes@google.com> CC: "Alex,Shi" <alex.shi@intel.com> CC: Shaohua Li <shaohua.li@intel.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-12-16 23:25:34 +08:00
void *next_object = get_freepointer_safe(s, object);
Lockless (and preemptless) fastpaths for slub Use the this_cpu_cmpxchg_double functionality to implement a lockless allocation algorithm on arches that support fast this_cpu_ops. Each of the per cpu pointers is paired with a transaction id that ensures that updates of the per cpu information can only occur in sequence on a certain cpu. A transaction id is a "long" integer that is comprised of an event number and the cpu number. The event number is incremented for every change to the per cpu state. This means that the cmpxchg instruction can verify for an update that nothing interfered and that we are updating the percpu structure for the processor where we picked up the information and that we are also currently on that processor when we update the information. This results in a significant decrease of the overhead in the fastpaths. It also makes it easy to adopt the fast path for realtime kernels since this is lockless and does not require the use of the current per cpu area over the critical section. It is only important that the per cpu area is current at the beginning of the critical section and at the end. So there is no need even to disable preemption. Test results show that the fastpath cycle count is reduced by up to ~ 40% (alloc/free test goes from ~140 cycles down to ~80). The slowpath for kfree adds a few cycles. Sadly this does nothing for the slowpath which is where the main issues with performance in slub are but the best case performance rises significantly. (For that see the more complex slub patches that require cmpxchg_double) Kmalloc: alloc/free test Before: 10000 times kmalloc(8)/kfree -> 134 cycles 10000 times kmalloc(16)/kfree -> 152 cycles 10000 times kmalloc(32)/kfree -> 144 cycles 10000 times kmalloc(64)/kfree -> 142 cycles 10000 times kmalloc(128)/kfree -> 142 cycles 10000 times kmalloc(256)/kfree -> 132 cycles 10000 times kmalloc(512)/kfree -> 132 cycles 10000 times kmalloc(1024)/kfree -> 135 cycles 10000 times kmalloc(2048)/kfree -> 135 cycles 10000 times kmalloc(4096)/kfree -> 135 cycles 10000 times kmalloc(8192)/kfree -> 144 cycles 10000 times kmalloc(16384)/kfree -> 754 cycles After: 10000 times kmalloc(8)/kfree -> 78 cycles 10000 times kmalloc(16)/kfree -> 78 cycles 10000 times kmalloc(32)/kfree -> 82 cycles 10000 times kmalloc(64)/kfree -> 88 cycles 10000 times kmalloc(128)/kfree -> 79 cycles 10000 times kmalloc(256)/kfree -> 79 cycles 10000 times kmalloc(512)/kfree -> 85 cycles 10000 times kmalloc(1024)/kfree -> 82 cycles 10000 times kmalloc(2048)/kfree -> 82 cycles 10000 times kmalloc(4096)/kfree -> 85 cycles 10000 times kmalloc(8192)/kfree -> 82 cycles 10000 times kmalloc(16384)/kfree -> 706 cycles Kmalloc: Repeatedly allocate then free test Before: 10000 times kmalloc(8) -> 211 cycles kfree -> 113 cycles 10000 times kmalloc(16) -> 174 cycles kfree -> 115 cycles 10000 times kmalloc(32) -> 235 cycles kfree -> 129 cycles 10000 times kmalloc(64) -> 222 cycles kfree -> 120 cycles 10000 times kmalloc(128) -> 343 cycles kfree -> 139 cycles 10000 times kmalloc(256) -> 827 cycles kfree -> 147 cycles 10000 times kmalloc(512) -> 1048 cycles kfree -> 272 cycles 10000 times kmalloc(1024) -> 2043 cycles kfree -> 528 cycles 10000 times kmalloc(2048) -> 4002 cycles kfree -> 571 cycles 10000 times kmalloc(4096) -> 7740 cycles kfree -> 628 cycles 10000 times kmalloc(8192) -> 8062 cycles kfree -> 850 cycles 10000 times kmalloc(16384) -> 8895 cycles kfree -> 1249 cycles After: 10000 times kmalloc(8) -> 190 cycles kfree -> 129 cycles 10000 times kmalloc(16) -> 76 cycles kfree -> 123 cycles 10000 times kmalloc(32) -> 126 cycles kfree -> 124 cycles 10000 times kmalloc(64) -> 181 cycles kfree -> 128 cycles 10000 times kmalloc(128) -> 310 cycles kfree -> 140 cycles 10000 times kmalloc(256) -> 809 cycles kfree -> 165 cycles 10000 times kmalloc(512) -> 1005 cycles kfree -> 269 cycles 10000 times kmalloc(1024) -> 1999 cycles kfree -> 527 cycles 10000 times kmalloc(2048) -> 3967 cycles kfree -> 570 cycles 10000 times kmalloc(4096) -> 7658 cycles kfree -> 637 cycles 10000 times kmalloc(8192) -> 8111 cycles kfree -> 859 cycles 10000 times kmalloc(16384) -> 8791 cycles kfree -> 1173 cycles Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-02-26 01:38:54 +08:00
/*
* The cmpxchg will only match if there was no additional
Lockless (and preemptless) fastpaths for slub Use the this_cpu_cmpxchg_double functionality to implement a lockless allocation algorithm on arches that support fast this_cpu_ops. Each of the per cpu pointers is paired with a transaction id that ensures that updates of the per cpu information can only occur in sequence on a certain cpu. A transaction id is a "long" integer that is comprised of an event number and the cpu number. The event number is incremented for every change to the per cpu state. This means that the cmpxchg instruction can verify for an update that nothing interfered and that we are updating the percpu structure for the processor where we picked up the information and that we are also currently on that processor when we update the information. This results in a significant decrease of the overhead in the fastpaths. It also makes it easy to adopt the fast path for realtime kernels since this is lockless and does not require the use of the current per cpu area over the critical section. It is only important that the per cpu area is current at the beginning of the critical section and at the end. So there is no need even to disable preemption. Test results show that the fastpath cycle count is reduced by up to ~ 40% (alloc/free test goes from ~140 cycles down to ~80). The slowpath for kfree adds a few cycles. Sadly this does nothing for the slowpath which is where the main issues with performance in slub are but the best case performance rises significantly. (For that see the more complex slub patches that require cmpxchg_double) Kmalloc: alloc/free test Before: 10000 times kmalloc(8)/kfree -> 134 cycles 10000 times kmalloc(16)/kfree -> 152 cycles 10000 times kmalloc(32)/kfree -> 144 cycles 10000 times kmalloc(64)/kfree -> 142 cycles 10000 times kmalloc(128)/kfree -> 142 cycles 10000 times kmalloc(256)/kfree -> 132 cycles 10000 times kmalloc(512)/kfree -> 132 cycles 10000 times kmalloc(1024)/kfree -> 135 cycles 10000 times kmalloc(2048)/kfree -> 135 cycles 10000 times kmalloc(4096)/kfree -> 135 cycles 10000 times kmalloc(8192)/kfree -> 144 cycles 10000 times kmalloc(16384)/kfree -> 754 cycles After: 10000 times kmalloc(8)/kfree -> 78 cycles 10000 times kmalloc(16)/kfree -> 78 cycles 10000 times kmalloc(32)/kfree -> 82 cycles 10000 times kmalloc(64)/kfree -> 88 cycles 10000 times kmalloc(128)/kfree -> 79 cycles 10000 times kmalloc(256)/kfree -> 79 cycles 10000 times kmalloc(512)/kfree -> 85 cycles 10000 times kmalloc(1024)/kfree -> 82 cycles 10000 times kmalloc(2048)/kfree -> 82 cycles 10000 times kmalloc(4096)/kfree -> 85 cycles 10000 times kmalloc(8192)/kfree -> 82 cycles 10000 times kmalloc(16384)/kfree -> 706 cycles Kmalloc: Repeatedly allocate then free test Before: 10000 times kmalloc(8) -> 211 cycles kfree -> 113 cycles 10000 times kmalloc(16) -> 174 cycles kfree -> 115 cycles 10000 times kmalloc(32) -> 235 cycles kfree -> 129 cycles 10000 times kmalloc(64) -> 222 cycles kfree -> 120 cycles 10000 times kmalloc(128) -> 343 cycles kfree -> 139 cycles 10000 times kmalloc(256) -> 827 cycles kfree -> 147 cycles 10000 times kmalloc(512) -> 1048 cycles kfree -> 272 cycles 10000 times kmalloc(1024) -> 2043 cycles kfree -> 528 cycles 10000 times kmalloc(2048) -> 4002 cycles kfree -> 571 cycles 10000 times kmalloc(4096) -> 7740 cycles kfree -> 628 cycles 10000 times kmalloc(8192) -> 8062 cycles kfree -> 850 cycles 10000 times kmalloc(16384) -> 8895 cycles kfree -> 1249 cycles After: 10000 times kmalloc(8) -> 190 cycles kfree -> 129 cycles 10000 times kmalloc(16) -> 76 cycles kfree -> 123 cycles 10000 times kmalloc(32) -> 126 cycles kfree -> 124 cycles 10000 times kmalloc(64) -> 181 cycles kfree -> 128 cycles 10000 times kmalloc(128) -> 310 cycles kfree -> 140 cycles 10000 times kmalloc(256) -> 809 cycles kfree -> 165 cycles 10000 times kmalloc(512) -> 1005 cycles kfree -> 269 cycles 10000 times kmalloc(1024) -> 1999 cycles kfree -> 527 cycles 10000 times kmalloc(2048) -> 3967 cycles kfree -> 570 cycles 10000 times kmalloc(4096) -> 7658 cycles kfree -> 637 cycles 10000 times kmalloc(8192) -> 8111 cycles kfree -> 859 cycles 10000 times kmalloc(16384) -> 8791 cycles kfree -> 1173 cycles Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-02-26 01:38:54 +08:00
* operation and if we are on the right processor.
*
* The cmpxchg does the following atomically (without lock
* semantics!)
Lockless (and preemptless) fastpaths for slub Use the this_cpu_cmpxchg_double functionality to implement a lockless allocation algorithm on arches that support fast this_cpu_ops. Each of the per cpu pointers is paired with a transaction id that ensures that updates of the per cpu information can only occur in sequence on a certain cpu. A transaction id is a "long" integer that is comprised of an event number and the cpu number. The event number is incremented for every change to the per cpu state. This means that the cmpxchg instruction can verify for an update that nothing interfered and that we are updating the percpu structure for the processor where we picked up the information and that we are also currently on that processor when we update the information. This results in a significant decrease of the overhead in the fastpaths. It also makes it easy to adopt the fast path for realtime kernels since this is lockless and does not require the use of the current per cpu area over the critical section. It is only important that the per cpu area is current at the beginning of the critical section and at the end. So there is no need even to disable preemption. Test results show that the fastpath cycle count is reduced by up to ~ 40% (alloc/free test goes from ~140 cycles down to ~80). The slowpath for kfree adds a few cycles. Sadly this does nothing for the slowpath which is where the main issues with performance in slub are but the best case performance rises significantly. (For that see the more complex slub patches that require cmpxchg_double) Kmalloc: alloc/free test Before: 10000 times kmalloc(8)/kfree -> 134 cycles 10000 times kmalloc(16)/kfree -> 152 cycles 10000 times kmalloc(32)/kfree -> 144 cycles 10000 times kmalloc(64)/kfree -> 142 cycles 10000 times kmalloc(128)/kfree -> 142 cycles 10000 times kmalloc(256)/kfree -> 132 cycles 10000 times kmalloc(512)/kfree -> 132 cycles 10000 times kmalloc(1024)/kfree -> 135 cycles 10000 times kmalloc(2048)/kfree -> 135 cycles 10000 times kmalloc(4096)/kfree -> 135 cycles 10000 times kmalloc(8192)/kfree -> 144 cycles 10000 times kmalloc(16384)/kfree -> 754 cycles After: 10000 times kmalloc(8)/kfree -> 78 cycles 10000 times kmalloc(16)/kfree -> 78 cycles 10000 times kmalloc(32)/kfree -> 82 cycles 10000 times kmalloc(64)/kfree -> 88 cycles 10000 times kmalloc(128)/kfree -> 79 cycles 10000 times kmalloc(256)/kfree -> 79 cycles 10000 times kmalloc(512)/kfree -> 85 cycles 10000 times kmalloc(1024)/kfree -> 82 cycles 10000 times kmalloc(2048)/kfree -> 82 cycles 10000 times kmalloc(4096)/kfree -> 85 cycles 10000 times kmalloc(8192)/kfree -> 82 cycles 10000 times kmalloc(16384)/kfree -> 706 cycles Kmalloc: Repeatedly allocate then free test Before: 10000 times kmalloc(8) -> 211 cycles kfree -> 113 cycles 10000 times kmalloc(16) -> 174 cycles kfree -> 115 cycles 10000 times kmalloc(32) -> 235 cycles kfree -> 129 cycles 10000 times kmalloc(64) -> 222 cycles kfree -> 120 cycles 10000 times kmalloc(128) -> 343 cycles kfree -> 139 cycles 10000 times kmalloc(256) -> 827 cycles kfree -> 147 cycles 10000 times kmalloc(512) -> 1048 cycles kfree -> 272 cycles 10000 times kmalloc(1024) -> 2043 cycles kfree -> 528 cycles 10000 times kmalloc(2048) -> 4002 cycles kfree -> 571 cycles 10000 times kmalloc(4096) -> 7740 cycles kfree -> 628 cycles 10000 times kmalloc(8192) -> 8062 cycles kfree -> 850 cycles 10000 times kmalloc(16384) -> 8895 cycles kfree -> 1249 cycles After: 10000 times kmalloc(8) -> 190 cycles kfree -> 129 cycles 10000 times kmalloc(16) -> 76 cycles kfree -> 123 cycles 10000 times kmalloc(32) -> 126 cycles kfree -> 124 cycles 10000 times kmalloc(64) -> 181 cycles kfree -> 128 cycles 10000 times kmalloc(128) -> 310 cycles kfree -> 140 cycles 10000 times kmalloc(256) -> 809 cycles kfree -> 165 cycles 10000 times kmalloc(512) -> 1005 cycles kfree -> 269 cycles 10000 times kmalloc(1024) -> 1999 cycles kfree -> 527 cycles 10000 times kmalloc(2048) -> 3967 cycles kfree -> 570 cycles 10000 times kmalloc(4096) -> 7658 cycles kfree -> 637 cycles 10000 times kmalloc(8192) -> 8111 cycles kfree -> 859 cycles 10000 times kmalloc(16384) -> 8791 cycles kfree -> 1173 cycles Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-02-26 01:38:54 +08:00
* 1. Relocate first pointer to the current per cpu area.
* 2. Verify that tid and freelist have not been changed
* 3. If they were not changed replace tid and freelist
*
* Since this is without lock semantics the protection is only
* against code executing on this cpu *not* from access by
* other cpus.
Lockless (and preemptless) fastpaths for slub Use the this_cpu_cmpxchg_double functionality to implement a lockless allocation algorithm on arches that support fast this_cpu_ops. Each of the per cpu pointers is paired with a transaction id that ensures that updates of the per cpu information can only occur in sequence on a certain cpu. A transaction id is a "long" integer that is comprised of an event number and the cpu number. The event number is incremented for every change to the per cpu state. This means that the cmpxchg instruction can verify for an update that nothing interfered and that we are updating the percpu structure for the processor where we picked up the information and that we are also currently on that processor when we update the information. This results in a significant decrease of the overhead in the fastpaths. It also makes it easy to adopt the fast path for realtime kernels since this is lockless and does not require the use of the current per cpu area over the critical section. It is only important that the per cpu area is current at the beginning of the critical section and at the end. So there is no need even to disable preemption. Test results show that the fastpath cycle count is reduced by up to ~ 40% (alloc/free test goes from ~140 cycles down to ~80). The slowpath for kfree adds a few cycles. Sadly this does nothing for the slowpath which is where the main issues with performance in slub are but the best case performance rises significantly. (For that see the more complex slub patches that require cmpxchg_double) Kmalloc: alloc/free test Before: 10000 times kmalloc(8)/kfree -> 134 cycles 10000 times kmalloc(16)/kfree -> 152 cycles 10000 times kmalloc(32)/kfree -> 144 cycles 10000 times kmalloc(64)/kfree -> 142 cycles 10000 times kmalloc(128)/kfree -> 142 cycles 10000 times kmalloc(256)/kfree -> 132 cycles 10000 times kmalloc(512)/kfree -> 132 cycles 10000 times kmalloc(1024)/kfree -> 135 cycles 10000 times kmalloc(2048)/kfree -> 135 cycles 10000 times kmalloc(4096)/kfree -> 135 cycles 10000 times kmalloc(8192)/kfree -> 144 cycles 10000 times kmalloc(16384)/kfree -> 754 cycles After: 10000 times kmalloc(8)/kfree -> 78 cycles 10000 times kmalloc(16)/kfree -> 78 cycles 10000 times kmalloc(32)/kfree -> 82 cycles 10000 times kmalloc(64)/kfree -> 88 cycles 10000 times kmalloc(128)/kfree -> 79 cycles 10000 times kmalloc(256)/kfree -> 79 cycles 10000 times kmalloc(512)/kfree -> 85 cycles 10000 times kmalloc(1024)/kfree -> 82 cycles 10000 times kmalloc(2048)/kfree -> 82 cycles 10000 times kmalloc(4096)/kfree -> 85 cycles 10000 times kmalloc(8192)/kfree -> 82 cycles 10000 times kmalloc(16384)/kfree -> 706 cycles Kmalloc: Repeatedly allocate then free test Before: 10000 times kmalloc(8) -> 211 cycles kfree -> 113 cycles 10000 times kmalloc(16) -> 174 cycles kfree -> 115 cycles 10000 times kmalloc(32) -> 235 cycles kfree -> 129 cycles 10000 times kmalloc(64) -> 222 cycles kfree -> 120 cycles 10000 times kmalloc(128) -> 343 cycles kfree -> 139 cycles 10000 times kmalloc(256) -> 827 cycles kfree -> 147 cycles 10000 times kmalloc(512) -> 1048 cycles kfree -> 272 cycles 10000 times kmalloc(1024) -> 2043 cycles kfree -> 528 cycles 10000 times kmalloc(2048) -> 4002 cycles kfree -> 571 cycles 10000 times kmalloc(4096) -> 7740 cycles kfree -> 628 cycles 10000 times kmalloc(8192) -> 8062 cycles kfree -> 850 cycles 10000 times kmalloc(16384) -> 8895 cycles kfree -> 1249 cycles After: 10000 times kmalloc(8) -> 190 cycles kfree -> 129 cycles 10000 times kmalloc(16) -> 76 cycles kfree -> 123 cycles 10000 times kmalloc(32) -> 126 cycles kfree -> 124 cycles 10000 times kmalloc(64) -> 181 cycles kfree -> 128 cycles 10000 times kmalloc(128) -> 310 cycles kfree -> 140 cycles 10000 times kmalloc(256) -> 809 cycles kfree -> 165 cycles 10000 times kmalloc(512) -> 1005 cycles kfree -> 269 cycles 10000 times kmalloc(1024) -> 1999 cycles kfree -> 527 cycles 10000 times kmalloc(2048) -> 3967 cycles kfree -> 570 cycles 10000 times kmalloc(4096) -> 7658 cycles kfree -> 637 cycles 10000 times kmalloc(8192) -> 8111 cycles kfree -> 859 cycles 10000 times kmalloc(16384) -> 8791 cycles kfree -> 1173 cycles Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-02-26 01:38:54 +08:00
*/
if (unlikely(!this_cpu_cmpxchg_double(
Lockless (and preemptless) fastpaths for slub Use the this_cpu_cmpxchg_double functionality to implement a lockless allocation algorithm on arches that support fast this_cpu_ops. Each of the per cpu pointers is paired with a transaction id that ensures that updates of the per cpu information can only occur in sequence on a certain cpu. A transaction id is a "long" integer that is comprised of an event number and the cpu number. The event number is incremented for every change to the per cpu state. This means that the cmpxchg instruction can verify for an update that nothing interfered and that we are updating the percpu structure for the processor where we picked up the information and that we are also currently on that processor when we update the information. This results in a significant decrease of the overhead in the fastpaths. It also makes it easy to adopt the fast path for realtime kernels since this is lockless and does not require the use of the current per cpu area over the critical section. It is only important that the per cpu area is current at the beginning of the critical section and at the end. So there is no need even to disable preemption. Test results show that the fastpath cycle count is reduced by up to ~ 40% (alloc/free test goes from ~140 cycles down to ~80). The slowpath for kfree adds a few cycles. Sadly this does nothing for the slowpath which is where the main issues with performance in slub are but the best case performance rises significantly. (For that see the more complex slub patches that require cmpxchg_double) Kmalloc: alloc/free test Before: 10000 times kmalloc(8)/kfree -> 134 cycles 10000 times kmalloc(16)/kfree -> 152 cycles 10000 times kmalloc(32)/kfree -> 144 cycles 10000 times kmalloc(64)/kfree -> 142 cycles 10000 times kmalloc(128)/kfree -> 142 cycles 10000 times kmalloc(256)/kfree -> 132 cycles 10000 times kmalloc(512)/kfree -> 132 cycles 10000 times kmalloc(1024)/kfree -> 135 cycles 10000 times kmalloc(2048)/kfree -> 135 cycles 10000 times kmalloc(4096)/kfree -> 135 cycles 10000 times kmalloc(8192)/kfree -> 144 cycles 10000 times kmalloc(16384)/kfree -> 754 cycles After: 10000 times kmalloc(8)/kfree -> 78 cycles 10000 times kmalloc(16)/kfree -> 78 cycles 10000 times kmalloc(32)/kfree -> 82 cycles 10000 times kmalloc(64)/kfree -> 88 cycles 10000 times kmalloc(128)/kfree -> 79 cycles 10000 times kmalloc(256)/kfree -> 79 cycles 10000 times kmalloc(512)/kfree -> 85 cycles 10000 times kmalloc(1024)/kfree -> 82 cycles 10000 times kmalloc(2048)/kfree -> 82 cycles 10000 times kmalloc(4096)/kfree -> 85 cycles 10000 times kmalloc(8192)/kfree -> 82 cycles 10000 times kmalloc(16384)/kfree -> 706 cycles Kmalloc: Repeatedly allocate then free test Before: 10000 times kmalloc(8) -> 211 cycles kfree -> 113 cycles 10000 times kmalloc(16) -> 174 cycles kfree -> 115 cycles 10000 times kmalloc(32) -> 235 cycles kfree -> 129 cycles 10000 times kmalloc(64) -> 222 cycles kfree -> 120 cycles 10000 times kmalloc(128) -> 343 cycles kfree -> 139 cycles 10000 times kmalloc(256) -> 827 cycles kfree -> 147 cycles 10000 times kmalloc(512) -> 1048 cycles kfree -> 272 cycles 10000 times kmalloc(1024) -> 2043 cycles kfree -> 528 cycles 10000 times kmalloc(2048) -> 4002 cycles kfree -> 571 cycles 10000 times kmalloc(4096) -> 7740 cycles kfree -> 628 cycles 10000 times kmalloc(8192) -> 8062 cycles kfree -> 850 cycles 10000 times kmalloc(16384) -> 8895 cycles kfree -> 1249 cycles After: 10000 times kmalloc(8) -> 190 cycles kfree -> 129 cycles 10000 times kmalloc(16) -> 76 cycles kfree -> 123 cycles 10000 times kmalloc(32) -> 126 cycles kfree -> 124 cycles 10000 times kmalloc(64) -> 181 cycles kfree -> 128 cycles 10000 times kmalloc(128) -> 310 cycles kfree -> 140 cycles 10000 times kmalloc(256) -> 809 cycles kfree -> 165 cycles 10000 times kmalloc(512) -> 1005 cycles kfree -> 269 cycles 10000 times kmalloc(1024) -> 1999 cycles kfree -> 527 cycles 10000 times kmalloc(2048) -> 3967 cycles kfree -> 570 cycles 10000 times kmalloc(4096) -> 7658 cycles kfree -> 637 cycles 10000 times kmalloc(8192) -> 8111 cycles kfree -> 859 cycles 10000 times kmalloc(16384) -> 8791 cycles kfree -> 1173 cycles Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-02-26 01:38:54 +08:00
s->cpu_slab->freelist, s->cpu_slab->tid,
object, tid,
slub: prefetch next freelist pointer in slab_alloc() Recycling a page is a problem, since freelist link chain is hot on cpu(s) which freed objects, and possibly very cold on cpu currently owning slab. Adding a prefetch of cache line containing the pointer to next object in slab_alloc() helps a lot in many workloads, in particular on assymetric ones (allocations done on one cpu, frees on another cpus). Added cost is three machine instructions only. Examples on my dual socket quad core ht machine (Intel CPU E5540 @2.53GHz) (16 logical cpus, 2 memory nodes), 64bit kernel. Before patch : # perf stat -r 32 hackbench 50 process 4000 >/dev/null Performance counter stats for 'hackbench 50 process 4000' (32 runs): 327577,471718 task-clock # 15,821 CPUs utilized ( +- 0,64% ) 28 866 491 context-switches # 0,088 M/sec ( +- 1,80% ) 1 506 929 CPU-migrations # 0,005 M/sec ( +- 3,24% ) 127 151 page-faults # 0,000 M/sec ( +- 0,16% ) 829 399 813 448 cycles # 2,532 GHz ( +- 0,64% ) 580 664 691 740 stalled-cycles-frontend # 70,01% frontend cycles idle ( +- 0,71% ) 197 431 700 448 stalled-cycles-backend # 23,80% backend cycles idle ( +- 1,03% ) 503 548 648 975 instructions # 0,61 insns per cycle # 1,15 stalled cycles per insn ( +- 0,46% ) 95 780 068 471 branches # 292,389 M/sec ( +- 0,48% ) 1 426 407 916 branch-misses # 1,49% of all branches ( +- 1,35% ) 20,705679994 seconds time elapsed ( +- 0,64% ) After patch : # perf stat -r 32 hackbench 50 process 4000 >/dev/null Performance counter stats for 'hackbench 50 process 4000' (32 runs): 286236,542804 task-clock # 15,786 CPUs utilized ( +- 1,32% ) 19 703 372 context-switches # 0,069 M/sec ( +- 4,99% ) 1 658 249 CPU-migrations # 0,006 M/sec ( +- 6,62% ) 126 776 page-faults # 0,000 M/sec ( +- 0,12% ) 724 636 593 213 cycles # 2,532 GHz ( +- 1,32% ) 499 320 714 837 stalled-cycles-frontend # 68,91% frontend cycles idle ( +- 1,47% ) 156 555 126 809 stalled-cycles-backend # 21,60% backend cycles idle ( +- 2,22% ) 463 897 792 661 instructions # 0,64 insns per cycle # 1,08 stalled cycles per insn ( +- 0,94% ) 87 717 352 563 branches # 306,451 M/sec ( +- 0,99% ) 941 738 280 branch-misses # 1,07% of all branches ( +- 3,35% ) 18,132070670 seconds time elapsed ( +- 1,30% ) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Acked-by: Christoph Lameter <cl@linux.com> CC: Matt Mackall <mpm@selenic.com> CC: David Rientjes <rientjes@google.com> CC: "Alex,Shi" <alex.shi@intel.com> CC: Shaohua Li <shaohua.li@intel.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-12-16 23:25:34 +08:00
next_object, next_tid(tid)))) {
Lockless (and preemptless) fastpaths for slub Use the this_cpu_cmpxchg_double functionality to implement a lockless allocation algorithm on arches that support fast this_cpu_ops. Each of the per cpu pointers is paired with a transaction id that ensures that updates of the per cpu information can only occur in sequence on a certain cpu. A transaction id is a "long" integer that is comprised of an event number and the cpu number. The event number is incremented for every change to the per cpu state. This means that the cmpxchg instruction can verify for an update that nothing interfered and that we are updating the percpu structure for the processor where we picked up the information and that we are also currently on that processor when we update the information. This results in a significant decrease of the overhead in the fastpaths. It also makes it easy to adopt the fast path for realtime kernels since this is lockless and does not require the use of the current per cpu area over the critical section. It is only important that the per cpu area is current at the beginning of the critical section and at the end. So there is no need even to disable preemption. Test results show that the fastpath cycle count is reduced by up to ~ 40% (alloc/free test goes from ~140 cycles down to ~80). The slowpath for kfree adds a few cycles. Sadly this does nothing for the slowpath which is where the main issues with performance in slub are but the best case performance rises significantly. (For that see the more complex slub patches that require cmpxchg_double) Kmalloc: alloc/free test Before: 10000 times kmalloc(8)/kfree -> 134 cycles 10000 times kmalloc(16)/kfree -> 152 cycles 10000 times kmalloc(32)/kfree -> 144 cycles 10000 times kmalloc(64)/kfree -> 142 cycles 10000 times kmalloc(128)/kfree -> 142 cycles 10000 times kmalloc(256)/kfree -> 132 cycles 10000 times kmalloc(512)/kfree -> 132 cycles 10000 times kmalloc(1024)/kfree -> 135 cycles 10000 times kmalloc(2048)/kfree -> 135 cycles 10000 times kmalloc(4096)/kfree -> 135 cycles 10000 times kmalloc(8192)/kfree -> 144 cycles 10000 times kmalloc(16384)/kfree -> 754 cycles After: 10000 times kmalloc(8)/kfree -> 78 cycles 10000 times kmalloc(16)/kfree -> 78 cycles 10000 times kmalloc(32)/kfree -> 82 cycles 10000 times kmalloc(64)/kfree -> 88 cycles 10000 times kmalloc(128)/kfree -> 79 cycles 10000 times kmalloc(256)/kfree -> 79 cycles 10000 times kmalloc(512)/kfree -> 85 cycles 10000 times kmalloc(1024)/kfree -> 82 cycles 10000 times kmalloc(2048)/kfree -> 82 cycles 10000 times kmalloc(4096)/kfree -> 85 cycles 10000 times kmalloc(8192)/kfree -> 82 cycles 10000 times kmalloc(16384)/kfree -> 706 cycles Kmalloc: Repeatedly allocate then free test Before: 10000 times kmalloc(8) -> 211 cycles kfree -> 113 cycles 10000 times kmalloc(16) -> 174 cycles kfree -> 115 cycles 10000 times kmalloc(32) -> 235 cycles kfree -> 129 cycles 10000 times kmalloc(64) -> 222 cycles kfree -> 120 cycles 10000 times kmalloc(128) -> 343 cycles kfree -> 139 cycles 10000 times kmalloc(256) -> 827 cycles kfree -> 147 cycles 10000 times kmalloc(512) -> 1048 cycles kfree -> 272 cycles 10000 times kmalloc(1024) -> 2043 cycles kfree -> 528 cycles 10000 times kmalloc(2048) -> 4002 cycles kfree -> 571 cycles 10000 times kmalloc(4096) -> 7740 cycles kfree -> 628 cycles 10000 times kmalloc(8192) -> 8062 cycles kfree -> 850 cycles 10000 times kmalloc(16384) -> 8895 cycles kfree -> 1249 cycles After: 10000 times kmalloc(8) -> 190 cycles kfree -> 129 cycles 10000 times kmalloc(16) -> 76 cycles kfree -> 123 cycles 10000 times kmalloc(32) -> 126 cycles kfree -> 124 cycles 10000 times kmalloc(64) -> 181 cycles kfree -> 128 cycles 10000 times kmalloc(128) -> 310 cycles kfree -> 140 cycles 10000 times kmalloc(256) -> 809 cycles kfree -> 165 cycles 10000 times kmalloc(512) -> 1005 cycles kfree -> 269 cycles 10000 times kmalloc(1024) -> 1999 cycles kfree -> 527 cycles 10000 times kmalloc(2048) -> 3967 cycles kfree -> 570 cycles 10000 times kmalloc(4096) -> 7658 cycles kfree -> 637 cycles 10000 times kmalloc(8192) -> 8111 cycles kfree -> 859 cycles 10000 times kmalloc(16384) -> 8791 cycles kfree -> 1173 cycles Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-02-26 01:38:54 +08:00
note_cmpxchg_failure("slab_alloc", s, tid);
goto redo;
}
slub: prefetch next freelist pointer in slab_alloc() Recycling a page is a problem, since freelist link chain is hot on cpu(s) which freed objects, and possibly very cold on cpu currently owning slab. Adding a prefetch of cache line containing the pointer to next object in slab_alloc() helps a lot in many workloads, in particular on assymetric ones (allocations done on one cpu, frees on another cpus). Added cost is three machine instructions only. Examples on my dual socket quad core ht machine (Intel CPU E5540 @2.53GHz) (16 logical cpus, 2 memory nodes), 64bit kernel. Before patch : # perf stat -r 32 hackbench 50 process 4000 >/dev/null Performance counter stats for 'hackbench 50 process 4000' (32 runs): 327577,471718 task-clock # 15,821 CPUs utilized ( +- 0,64% ) 28 866 491 context-switches # 0,088 M/sec ( +- 1,80% ) 1 506 929 CPU-migrations # 0,005 M/sec ( +- 3,24% ) 127 151 page-faults # 0,000 M/sec ( +- 0,16% ) 829 399 813 448 cycles # 2,532 GHz ( +- 0,64% ) 580 664 691 740 stalled-cycles-frontend # 70,01% frontend cycles idle ( +- 0,71% ) 197 431 700 448 stalled-cycles-backend # 23,80% backend cycles idle ( +- 1,03% ) 503 548 648 975 instructions # 0,61 insns per cycle # 1,15 stalled cycles per insn ( +- 0,46% ) 95 780 068 471 branches # 292,389 M/sec ( +- 0,48% ) 1 426 407 916 branch-misses # 1,49% of all branches ( +- 1,35% ) 20,705679994 seconds time elapsed ( +- 0,64% ) After patch : # perf stat -r 32 hackbench 50 process 4000 >/dev/null Performance counter stats for 'hackbench 50 process 4000' (32 runs): 286236,542804 task-clock # 15,786 CPUs utilized ( +- 1,32% ) 19 703 372 context-switches # 0,069 M/sec ( +- 4,99% ) 1 658 249 CPU-migrations # 0,006 M/sec ( +- 6,62% ) 126 776 page-faults # 0,000 M/sec ( +- 0,12% ) 724 636 593 213 cycles # 2,532 GHz ( +- 1,32% ) 499 320 714 837 stalled-cycles-frontend # 68,91% frontend cycles idle ( +- 1,47% ) 156 555 126 809 stalled-cycles-backend # 21,60% backend cycles idle ( +- 2,22% ) 463 897 792 661 instructions # 0,64 insns per cycle # 1,08 stalled cycles per insn ( +- 0,94% ) 87 717 352 563 branches # 306,451 M/sec ( +- 0,99% ) 941 738 280 branch-misses # 1,07% of all branches ( +- 3,35% ) 18,132070670 seconds time elapsed ( +- 1,30% ) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Acked-by: Christoph Lameter <cl@linux.com> CC: Matt Mackall <mpm@selenic.com> CC: David Rientjes <rientjes@google.com> CC: "Alex,Shi" <alex.shi@intel.com> CC: Shaohua Li <shaohua.li@intel.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-12-16 23:25:34 +08:00
prefetch_freepointer(s, next_object);
stat(s, ALLOC_FASTPATH);
slub: support concurrent local and remote frees and allocs on a slab Avoid atomic overhead in slab_alloc and slab_free SLUB needs to use the slab_lock for the per cpu slabs to synchronize with potential kfree operations. This patch avoids that need by moving all free objects onto a lockless_freelist. The regular freelist continues to exist and will be used to free objects. So while we consume the lockless_freelist the regular freelist may build up objects. If we are out of objects on the lockless_freelist then we may check the regular freelist. If it has objects then we move those over to the lockless_freelist and do this again. There is a significant savings in terms of atomic operations that have to be performed. We can even free directly to the lockless_freelist if we know that we are running on the same processor. So this speeds up short lived objects. They may be allocated and freed without taking the slab_lock. This is particular good for netperf. In order to maximize the effect of the new faster hotpath we extract the hottest performance pieces into inlined functions. These are then inlined into kmem_cache_alloc and kmem_cache_free. So hotpath allocation and freeing no longer requires a subroutine call within SLUB. [I am not sure that it is worth doing this because it changes the easy to read structure of slub just to reduce atomic ops. However, there is someone out there with a benchmark on 4 way and 8 way processor systems that seems to show a 5% regression vs. Slab. Seems that the regression is due to increased atomic operations use vs. SLAB in SLUB). I wonder if this is applicable or discernable at all in a real workload?] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-10 18:15:16 +08:00
}
maybe_wipe_obj_freeptr(s, object);
kasan, mm: integrate slab init_on_alloc with HW_TAGS This change uses the previously added memory initialization feature of HW_TAGS KASAN routines for slab memory when init_on_alloc is enabled. With this change, memory initialization memset() is no longer called when both HW_TAGS KASAN and init_on_alloc are enabled. Instead, memory is initialized in KASAN runtime. The memory initialization memset() is moved into slab_post_alloc_hook() that currently directly follows the initialization loop. A new argument is added to slab_post_alloc_hook() that indicates whether to initialize the memory or not. To avoid discrepancies with which memory gets initialized that can be caused by future changes, both KASAN hook and initialization memset() are put together and a warning comment is added. Combining setting allocation tags with memory initialization improves HW_TAGS KASAN performance when init_on_alloc is enabled. Link: https://lkml.kernel.org/r/c1292aeb5d519da221ec74a0684a949b027d7720.1615296150.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Reviewed-by: Marco Elver <elver@google.com> Cc: Alexander Potapenko <glider@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Branislav Rankov <Branislav.Rankov@arm.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christoph Lameter <cl@linux.com> Cc: David Rientjes <rientjes@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Evgenii Stepanov <eugenis@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kevin Brodsky <kevin.brodsky@arm.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Collingbourne <pcc@google.com> Cc: Vincenzo Frascino <vincenzo.frascino@arm.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-30 14:00:06 +08:00
init = slab_want_init_on_alloc(gfpflags, s);
mm, kfence: insert KFENCE hooks for SLUB Inserts KFENCE hooks into the SLUB allocator. To pass the originally requested size to KFENCE, add an argument 'orig_size' to slab_alloc*(). The additional argument is required to preserve the requested original size for kmalloc() allocations, which uses size classes (e.g. an allocation of 272 bytes will return an object of size 512). Therefore, kmem_cache::size does not represent the kmalloc-caller's requested size, and we must introduce the argument 'orig_size' to propagate the originally requested size to KFENCE. Without the originally requested size, we would not be able to detect out-of-bounds accesses for objects placed at the end of a KFENCE object page if that object is not equal to the kmalloc-size class it was bucketed into. When KFENCE is disabled, there is no additional overhead, since slab_alloc*() functions are __always_inline. Link: https://lkml.kernel.org/r/20201103175841.3495947-6-elver@google.com Signed-off-by: Marco Elver <elver@google.com> Signed-off-by: Alexander Potapenko <glider@google.com> Reviewed-by: Dmitry Vyukov <dvyukov@google.com> Reviewed-by: Jann Horn <jannh@google.com> Co-developed-by: Marco Elver <elver@google.com> Cc: Andrey Konovalov <andreyknvl@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christopher Lameter <cl@linux.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: David Rientjes <rientjes@google.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Hillf Danton <hdanton@sina.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Joern Engel <joern@purestorage.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kees Cook <keescook@chromium.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Paul E. McKenney <paulmck@kernel.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: SeongJae Park <sjpark@amazon.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-26 09:19:16 +08:00
out:
kasan, mm: integrate slab init_on_alloc with HW_TAGS This change uses the previously added memory initialization feature of HW_TAGS KASAN routines for slab memory when init_on_alloc is enabled. With this change, memory initialization memset() is no longer called when both HW_TAGS KASAN and init_on_alloc are enabled. Instead, memory is initialized in KASAN runtime. The memory initialization memset() is moved into slab_post_alloc_hook() that currently directly follows the initialization loop. A new argument is added to slab_post_alloc_hook() that indicates whether to initialize the memory or not. To avoid discrepancies with which memory gets initialized that can be caused by future changes, both KASAN hook and initialization memset() are put together and a warning comment is added. Combining setting allocation tags with memory initialization improves HW_TAGS KASAN performance when init_on_alloc is enabled. Link: https://lkml.kernel.org/r/c1292aeb5d519da221ec74a0684a949b027d7720.1615296150.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Reviewed-by: Marco Elver <elver@google.com> Cc: Alexander Potapenko <glider@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Branislav Rankov <Branislav.Rankov@arm.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christoph Lameter <cl@linux.com> Cc: David Rientjes <rientjes@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Evgenii Stepanov <eugenis@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kevin Brodsky <kevin.brodsky@arm.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Collingbourne <pcc@google.com> Cc: Vincenzo Frascino <vincenzo.frascino@arm.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-30 14:00:06 +08:00
slab_post_alloc_hook(s, objcg, gfpflags, 1, &object, init);
slub: support concurrent local and remote frees and allocs on a slab Avoid atomic overhead in slab_alloc and slab_free SLUB needs to use the slab_lock for the per cpu slabs to synchronize with potential kfree operations. This patch avoids that need by moving all free objects onto a lockless_freelist. The regular freelist continues to exist and will be used to free objects. So while we consume the lockless_freelist the regular freelist may build up objects. If we are out of objects on the lockless_freelist then we may check the regular freelist. If it has objects then we move those over to the lockless_freelist and do this again. There is a significant savings in terms of atomic operations that have to be performed. We can even free directly to the lockless_freelist if we know that we are running on the same processor. So this speeds up short lived objects. They may be allocated and freed without taking the slab_lock. This is particular good for netperf. In order to maximize the effect of the new faster hotpath we extract the hottest performance pieces into inlined functions. These are then inlined into kmem_cache_alloc and kmem_cache_free. So hotpath allocation and freeing no longer requires a subroutine call within SLUB. [I am not sure that it is worth doing this because it changes the easy to read structure of slub just to reduce atomic ops. However, there is someone out there with a benchmark on 4 way and 8 way processor systems that seems to show a 5% regression vs. Slab. Seems that the regression is due to increased atomic operations use vs. SLAB in SLUB). I wonder if this is applicable or discernable at all in a real workload?] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-10 18:15:16 +08:00
return object;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
static __always_inline void *slab_alloc(struct kmem_cache *s,
mm, kfence: insert KFENCE hooks for SLUB Inserts KFENCE hooks into the SLUB allocator. To pass the originally requested size to KFENCE, add an argument 'orig_size' to slab_alloc*(). The additional argument is required to preserve the requested original size for kmalloc() allocations, which uses size classes (e.g. an allocation of 272 bytes will return an object of size 512). Therefore, kmem_cache::size does not represent the kmalloc-caller's requested size, and we must introduce the argument 'orig_size' to propagate the originally requested size to KFENCE. Without the originally requested size, we would not be able to detect out-of-bounds accesses for objects placed at the end of a KFENCE object page if that object is not equal to the kmalloc-size class it was bucketed into. When KFENCE is disabled, there is no additional overhead, since slab_alloc*() functions are __always_inline. Link: https://lkml.kernel.org/r/20201103175841.3495947-6-elver@google.com Signed-off-by: Marco Elver <elver@google.com> Signed-off-by: Alexander Potapenko <glider@google.com> Reviewed-by: Dmitry Vyukov <dvyukov@google.com> Reviewed-by: Jann Horn <jannh@google.com> Co-developed-by: Marco Elver <elver@google.com> Cc: Andrey Konovalov <andreyknvl@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christopher Lameter <cl@linux.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: David Rientjes <rientjes@google.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Hillf Danton <hdanton@sina.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Joern Engel <joern@purestorage.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kees Cook <keescook@chromium.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Paul E. McKenney <paulmck@kernel.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: SeongJae Park <sjpark@amazon.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-26 09:19:16 +08:00
gfp_t gfpflags, unsigned long addr, size_t orig_size)
{
mm, kfence: insert KFENCE hooks for SLUB Inserts KFENCE hooks into the SLUB allocator. To pass the originally requested size to KFENCE, add an argument 'orig_size' to slab_alloc*(). The additional argument is required to preserve the requested original size for kmalloc() allocations, which uses size classes (e.g. an allocation of 272 bytes will return an object of size 512). Therefore, kmem_cache::size does not represent the kmalloc-caller's requested size, and we must introduce the argument 'orig_size' to propagate the originally requested size to KFENCE. Without the originally requested size, we would not be able to detect out-of-bounds accesses for objects placed at the end of a KFENCE object page if that object is not equal to the kmalloc-size class it was bucketed into. When KFENCE is disabled, there is no additional overhead, since slab_alloc*() functions are __always_inline. Link: https://lkml.kernel.org/r/20201103175841.3495947-6-elver@google.com Signed-off-by: Marco Elver <elver@google.com> Signed-off-by: Alexander Potapenko <glider@google.com> Reviewed-by: Dmitry Vyukov <dvyukov@google.com> Reviewed-by: Jann Horn <jannh@google.com> Co-developed-by: Marco Elver <elver@google.com> Cc: Andrey Konovalov <andreyknvl@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christopher Lameter <cl@linux.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: David Rientjes <rientjes@google.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Hillf Danton <hdanton@sina.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Joern Engel <joern@purestorage.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kees Cook <keescook@chromium.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Paul E. McKenney <paulmck@kernel.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: SeongJae Park <sjpark@amazon.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-26 09:19:16 +08:00
return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr, orig_size);
}
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
{
mm, kfence: insert KFENCE hooks for SLUB Inserts KFENCE hooks into the SLUB allocator. To pass the originally requested size to KFENCE, add an argument 'orig_size' to slab_alloc*(). The additional argument is required to preserve the requested original size for kmalloc() allocations, which uses size classes (e.g. an allocation of 272 bytes will return an object of size 512). Therefore, kmem_cache::size does not represent the kmalloc-caller's requested size, and we must introduce the argument 'orig_size' to propagate the originally requested size to KFENCE. Without the originally requested size, we would not be able to detect out-of-bounds accesses for objects placed at the end of a KFENCE object page if that object is not equal to the kmalloc-size class it was bucketed into. When KFENCE is disabled, there is no additional overhead, since slab_alloc*() functions are __always_inline. Link: https://lkml.kernel.org/r/20201103175841.3495947-6-elver@google.com Signed-off-by: Marco Elver <elver@google.com> Signed-off-by: Alexander Potapenko <glider@google.com> Reviewed-by: Dmitry Vyukov <dvyukov@google.com> Reviewed-by: Jann Horn <jannh@google.com> Co-developed-by: Marco Elver <elver@google.com> Cc: Andrey Konovalov <andreyknvl@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christopher Lameter <cl@linux.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: David Rientjes <rientjes@google.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Hillf Danton <hdanton@sina.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Joern Engel <joern@purestorage.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kees Cook <keescook@chromium.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Paul E. McKenney <paulmck@kernel.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: SeongJae Park <sjpark@amazon.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-26 09:19:16 +08:00
void *ret = slab_alloc(s, gfpflags, _RET_IP_, s->object_size);
trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
s->size, gfpflags);
return ret;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
EXPORT_SYMBOL(kmem_cache_alloc);
#ifdef CONFIG_TRACING
slub tracing: move trace calls out of always inlined functions to reduce kernel code size Having the trace calls defined in the always inlined kmalloc functions in include/linux/slub_def.h causes a lot of code duplication as the trace functions get instantiated for each kamalloc call site. This can simply be removed by pushing the trace calls down into the functions in slub.c. On my x86_64 built this patch shrinks the code size of the kernel by approx 36K and also shrinks the code size of many modules -- too many to list here ;) size vmlinux (2.6.36) reports text data bss dec hex filename 5410611 743172 828928 6982711 6a8c37 vmlinux 5373738 744244 828928 6946910 6a005e vmlinux + patch The resulting kernel has had some testing & kmalloc trace still seems to work. This patch - moves trace_kmalloc out of the inlined kmalloc() and pushes it down into kmem_cache_alloc_trace() so this it only get instantiated once. - rename kmem_cache_alloc_notrace() to kmem_cache_alloc_trace() to indicate that now is does have tracing. (maybe this would better being called something like kmalloc_kmem_cache ?) - adds a new function kmalloc_order() to handle allocation and tracing of large allocations of page order. - removes tracing from the inlined kmalloc_large() replacing them with a call to kmalloc_order(); - move tracing out of inlined kmalloc_node() and pushing it down into kmem_cache_alloc_node_trace - rename kmem_cache_alloc_node_notrace() to kmem_cache_alloc_node_trace() - removes the include of trace/events/kmem.h from slub_def.h. v2 - keep kmalloc_order_trace inline when !CONFIG_TRACE Signed-off-by: Richard Kennedy <richard@rsk.demon.co.uk> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2010-10-21 17:29:19 +08:00
void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
{
mm, kfence: insert KFENCE hooks for SLUB Inserts KFENCE hooks into the SLUB allocator. To pass the originally requested size to KFENCE, add an argument 'orig_size' to slab_alloc*(). The additional argument is required to preserve the requested original size for kmalloc() allocations, which uses size classes (e.g. an allocation of 272 bytes will return an object of size 512). Therefore, kmem_cache::size does not represent the kmalloc-caller's requested size, and we must introduce the argument 'orig_size' to propagate the originally requested size to KFENCE. Without the originally requested size, we would not be able to detect out-of-bounds accesses for objects placed at the end of a KFENCE object page if that object is not equal to the kmalloc-size class it was bucketed into. When KFENCE is disabled, there is no additional overhead, since slab_alloc*() functions are __always_inline. Link: https://lkml.kernel.org/r/20201103175841.3495947-6-elver@google.com Signed-off-by: Marco Elver <elver@google.com> Signed-off-by: Alexander Potapenko <glider@google.com> Reviewed-by: Dmitry Vyukov <dvyukov@google.com> Reviewed-by: Jann Horn <jannh@google.com> Co-developed-by: Marco Elver <elver@google.com> Cc: Andrey Konovalov <andreyknvl@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christopher Lameter <cl@linux.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: David Rientjes <rientjes@google.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Hillf Danton <hdanton@sina.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Joern Engel <joern@purestorage.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kees Cook <keescook@chromium.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Paul E. McKenney <paulmck@kernel.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: SeongJae Park <sjpark@amazon.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-26 09:19:16 +08:00
void *ret = slab_alloc(s, gfpflags, _RET_IP_, size);
slub tracing: move trace calls out of always inlined functions to reduce kernel code size Having the trace calls defined in the always inlined kmalloc functions in include/linux/slub_def.h causes a lot of code duplication as the trace functions get instantiated for each kamalloc call site. This can simply be removed by pushing the trace calls down into the functions in slub.c. On my x86_64 built this patch shrinks the code size of the kernel by approx 36K and also shrinks the code size of many modules -- too many to list here ;) size vmlinux (2.6.36) reports text data bss dec hex filename 5410611 743172 828928 6982711 6a8c37 vmlinux 5373738 744244 828928 6946910 6a005e vmlinux + patch The resulting kernel has had some testing & kmalloc trace still seems to work. This patch - moves trace_kmalloc out of the inlined kmalloc() and pushes it down into kmem_cache_alloc_trace() so this it only get instantiated once. - rename kmem_cache_alloc_notrace() to kmem_cache_alloc_trace() to indicate that now is does have tracing. (maybe this would better being called something like kmalloc_kmem_cache ?) - adds a new function kmalloc_order() to handle allocation and tracing of large allocations of page order. - removes tracing from the inlined kmalloc_large() replacing them with a call to kmalloc_order(); - move tracing out of inlined kmalloc_node() and pushing it down into kmem_cache_alloc_node_trace - rename kmem_cache_alloc_node_notrace() to kmem_cache_alloc_node_trace() - removes the include of trace/events/kmem.h from slub_def.h. v2 - keep kmalloc_order_trace inline when !CONFIG_TRACE Signed-off-by: Richard Kennedy <richard@rsk.demon.co.uk> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2010-10-21 17:29:19 +08:00
trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
kasan, mm: change hooks signatures Patch series "kasan: add software tag-based mode for arm64", v13. This patchset adds a new software tag-based mode to KASAN [1]. (Initially this mode was called KHWASAN, but it got renamed, see the naming rationale at the end of this section). The plan is to implement HWASan [2] for the kernel with the incentive, that it's going to have comparable to KASAN performance, but in the same time consume much less memory, trading that off for somewhat imprecise bug detection and being supported only for arm64. The underlying ideas of the approach used by software tag-based KASAN are: 1. By using the Top Byte Ignore (TBI) arm64 CPU feature, we can store pointer tags in the top byte of each kernel pointer. 2. Using shadow memory, we can store memory tags for each chunk of kernel memory. 3. On each memory allocation, we can generate a random tag, embed it into the returned pointer and set the memory tags that correspond to this chunk of memory to the same value. 4. By using compiler instrumentation, before each memory access we can add a check that the pointer tag matches the tag of the memory that is being accessed. 5. On a tag mismatch we report an error. With this patchset the existing KASAN mode gets renamed to generic KASAN, with the word "generic" meaning that the implementation can be supported by any architecture as it is purely software. The new mode this patchset adds is called software tag-based KASAN. The word "tag-based" refers to the fact that this mode uses tags embedded into the top byte of kernel pointers and the TBI arm64 CPU feature that allows to dereference such pointers. The word "software" here means that shadow memory manipulation and tag checking on pointer dereference is done in software. As it is the only tag-based implementation right now, "software tag-based" KASAN is sometimes referred to as simply "tag-based" in this patchset. A potential expansion of this mode is a hardware tag-based mode, which would use hardware memory tagging support (announced by Arm [3]) instead of compiler instrumentation and manual shadow memory manipulation. Same as generic KASAN, software tag-based KASAN is strictly a debugging feature. [1] https://www.kernel.org/doc/html/latest/dev-tools/kasan.html [2] http://clang.llvm.org/docs/HardwareAssistedAddressSanitizerDesign.html [3] https://community.arm.com/processors/b/blog/posts/arm-a-profile-architecture-2018-developments-armv85a ====== Rationale On mobile devices generic KASAN's memory usage is significant problem. One of the main reasons to have tag-based KASAN is to be able to perform a similar set of checks as the generic one does, but with lower memory requirements. Comment from Vishwath Mohan <vishwath@google.com>: I don't have data on-hand, but anecdotally both ASAN and KASAN have proven problematic to enable for environments that don't tolerate the increased memory pressure well. This includes (a) Low-memory form factors - Wear, TV, Things, lower-tier phones like Go, (c) Connected components like Pixel's visual core [1]. These are both places I'd love to have a low(er) memory footprint option at my disposal. Comment from Evgenii Stepanov <eugenis@google.com>: Looking at a live Android device under load, slab (according to /proc/meminfo) + kernel stack take 8-10% available RAM (~350MB). KASAN's overhead of 2x - 3x on top of it is not insignificant. Not having this overhead enables near-production use - ex. running KASAN/KHWASAN kernel on a personal, daily-use device to catch bugs that do not reproduce in test configuration. These are the ones that often cost the most engineering time to track down. CPU overhead is bad, but generally tolerable. RAM is critical, in our experience. Once it gets low enough, OOM-killer makes your life miserable. [1] https://www.blog.google/products/pixel/pixel-visual-core-image-processing-and-machine-learning-pixel-2/ ====== Technical details Software tag-based KASAN mode is implemented in a very similar way to the generic one. This patchset essentially does the following: 1. TCR_TBI1 is set to enable Top Byte Ignore. 2. Shadow memory is used (with a different scale, 1:16, so each shadow byte corresponds to 16 bytes of kernel memory) to store memory tags. 3. All slab objects are aligned to shadow scale, which is 16 bytes. 4. All pointers returned from the slab allocator are tagged with a random tag and the corresponding shadow memory is poisoned with the same value. 5. Compiler instrumentation is used to insert tag checks. Either by calling callbacks or by inlining them (CONFIG_KASAN_OUTLINE and CONFIG_KASAN_INLINE flags are reused). 6. When a tag mismatch is detected in callback instrumentation mode KASAN simply prints a bug report. In case of inline instrumentation, clang inserts a brk instruction, and KASAN has it's own brk handler, which reports the bug. 7. The memory in between slab objects is marked with a reserved tag, and acts as a redzone. 8. When a slab object is freed it's marked with a reserved tag. Bug detection is imprecise for two reasons: 1. We won't catch some small out-of-bounds accesses, that fall into the same shadow cell, as the last byte of a slab object. 2. We only have 1 byte to store tags, which means we have a 1/256 probability of a tag match for an incorrect access (actually even slightly less due to reserved tag values). Despite that there's a particular type of bugs that tag-based KASAN can detect compared to generic KASAN: use-after-free after the object has been allocated by someone else. ====== Testing Some kernel developers voiced a concern that changing the top byte of kernel pointers may lead to subtle bugs that are difficult to discover. To address this concern deliberate testing has been performed. It doesn't seem feasible to do some kind of static checking to find potential issues with pointer tagging, so a dynamic approach was taken. All pointer comparisons/subtractions have been instrumented in an LLVM compiler pass and a kernel module that would print a bug report whenever two pointers with different tags are being compared/subtracted (ignoring comparisons with NULL pointers and with pointers obtained by casting an error code to a pointer type) has been used. Then the kernel has been booted in QEMU and on an Odroid C2 board and syzkaller has been run. This yielded the following results. The two places that look interesting are: is_vmalloc_addr in include/linux/mm.h is_kernel_rodata in mm/util.c Here we compare a pointer with some fixed untagged values to make sure that the pointer lies in a particular part of the kernel address space. Since tag-based KASAN doesn't add tags to pointers that belong to rodata or vmalloc regions, this should work as is. To make sure debug checks to those two functions that check that the result doesn't change whether we operate on pointers with or without untagging has been added. A few other cases that don't look that interesting: Comparing pointers to achieve unique sorting order of pointee objects (e.g. sorting locks addresses before performing a double lock): tty_ldisc_lock_pair_timeout in drivers/tty/tty_ldisc.c pipe_double_lock in fs/pipe.c unix_state_double_lock in net/unix/af_unix.c lock_two_nondirectories in fs/inode.c mutex_lock_double in kernel/events/core.c ep_cmp_ffd in fs/eventpoll.c fsnotify_compare_groups fs/notify/mark.c Nothing needs to be done here, since the tags embedded into pointers don't change, so the sorting order would still be unique. Checks that a pointer belongs to some particular allocation: is_sibling_entry in lib/radix-tree.c object_is_on_stack in include/linux/sched/task_stack.h Nothing needs to be done here either, since two pointers can only belong to the same allocation if they have the same tag. Overall, since the kernel boots and works, there are no critical bugs. As for the rest, the traditional kernel testing way (use until fails) is the only one that looks feasible. Another point here is that tag-based KASAN is available under a separate config option that needs to be deliberately enabled. Even though it might be used in a "near-production" environment to find bugs that are not found during fuzzing or running tests, it is still a debug tool. ====== Benchmarks The following numbers were collected on Odroid C2 board. Both generic and tag-based KASAN were used in inline instrumentation mode. Boot time [1]: * ~1.7 sec for clean kernel * ~5.0 sec for generic KASAN * ~5.0 sec for tag-based KASAN Network performance [2]: * 8.33 Gbits/sec for clean kernel * 3.17 Gbits/sec for generic KASAN * 2.85 Gbits/sec for tag-based KASAN Slab memory usage after boot [3]: * ~40 kb for clean kernel * ~105 kb (~260% overhead) for generic KASAN * ~47 kb (~20% overhead) for tag-based KASAN KASAN memory overhead consists of three main parts: 1. Increased slab memory usage due to redzones. 2. Shadow memory (the whole reserved once during boot). 3. Quaratine (grows gradually until some preset limit; the more the limit, the more the chance to detect a use-after-free). Comparing tag-based vs generic KASAN for each of these points: 1. 20% vs 260% overhead. 2. 1/16th vs 1/8th of physical memory. 3. Tag-based KASAN doesn't require quarantine. [1] Time before the ext4 driver is initialized. [2] Measured as `iperf -s & iperf -c 127.0.0.1 -t 30`. [3] Measured as `cat /proc/meminfo | grep Slab`. ====== Some notes A few notes: 1. The patchset can be found here: https://github.com/xairy/kasan-prototype/tree/khwasan 2. Building requires a recent Clang version (7.0.0 or later). 3. Stack instrumentation is not supported yet and will be added later. This patch (of 25): Tag-based KASAN changes the value of the top byte of pointers returned from the kernel allocation functions (such as kmalloc). This patch updates KASAN hooks signatures and their usage in SLAB and SLUB code to reflect that. Link: http://lkml.kernel.org/r/aec2b5e3973781ff8a6bb6760f8543643202c451.1544099024.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Reviewed-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Reviewed-by: Dmitry Vyukov <dvyukov@google.com> Cc: Christoph Lameter <cl@linux.com> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 16:29:37 +08:00
ret = kasan_kmalloc(s, ret, size, gfpflags);
slub tracing: move trace calls out of always inlined functions to reduce kernel code size Having the trace calls defined in the always inlined kmalloc functions in include/linux/slub_def.h causes a lot of code duplication as the trace functions get instantiated for each kamalloc call site. This can simply be removed by pushing the trace calls down into the functions in slub.c. On my x86_64 built this patch shrinks the code size of the kernel by approx 36K and also shrinks the code size of many modules -- too many to list here ;) size vmlinux (2.6.36) reports text data bss dec hex filename 5410611 743172 828928 6982711 6a8c37 vmlinux 5373738 744244 828928 6946910 6a005e vmlinux + patch The resulting kernel has had some testing & kmalloc trace still seems to work. This patch - moves trace_kmalloc out of the inlined kmalloc() and pushes it down into kmem_cache_alloc_trace() so this it only get instantiated once. - rename kmem_cache_alloc_notrace() to kmem_cache_alloc_trace() to indicate that now is does have tracing. (maybe this would better being called something like kmalloc_kmem_cache ?) - adds a new function kmalloc_order() to handle allocation and tracing of large allocations of page order. - removes tracing from the inlined kmalloc_large() replacing them with a call to kmalloc_order(); - move tracing out of inlined kmalloc_node() and pushing it down into kmem_cache_alloc_node_trace - rename kmem_cache_alloc_node_notrace() to kmem_cache_alloc_node_trace() - removes the include of trace/events/kmem.h from slub_def.h. v2 - keep kmalloc_order_trace inline when !CONFIG_TRACE Signed-off-by: Richard Kennedy <richard@rsk.demon.co.uk> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2010-10-21 17:29:19 +08:00
return ret;
}
EXPORT_SYMBOL(kmem_cache_alloc_trace);
#endif
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
#ifdef CONFIG_NUMA
void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
{
mm, kfence: insert KFENCE hooks for SLUB Inserts KFENCE hooks into the SLUB allocator. To pass the originally requested size to KFENCE, add an argument 'orig_size' to slab_alloc*(). The additional argument is required to preserve the requested original size for kmalloc() allocations, which uses size classes (e.g. an allocation of 272 bytes will return an object of size 512). Therefore, kmem_cache::size does not represent the kmalloc-caller's requested size, and we must introduce the argument 'orig_size' to propagate the originally requested size to KFENCE. Without the originally requested size, we would not be able to detect out-of-bounds accesses for objects placed at the end of a KFENCE object page if that object is not equal to the kmalloc-size class it was bucketed into. When KFENCE is disabled, there is no additional overhead, since slab_alloc*() functions are __always_inline. Link: https://lkml.kernel.org/r/20201103175841.3495947-6-elver@google.com Signed-off-by: Marco Elver <elver@google.com> Signed-off-by: Alexander Potapenko <glider@google.com> Reviewed-by: Dmitry Vyukov <dvyukov@google.com> Reviewed-by: Jann Horn <jannh@google.com> Co-developed-by: Marco Elver <elver@google.com> Cc: Andrey Konovalov <andreyknvl@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christopher Lameter <cl@linux.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: David Rientjes <rientjes@google.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Hillf Danton <hdanton@sina.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Joern Engel <joern@purestorage.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kees Cook <keescook@chromium.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Paul E. McKenney <paulmck@kernel.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: SeongJae Park <sjpark@amazon.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-26 09:19:16 +08:00
void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_, s->object_size);
trace_kmem_cache_alloc_node(_RET_IP_, ret,
s->object_size, s->size, gfpflags, node);
return ret;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
EXPORT_SYMBOL(kmem_cache_alloc_node);
#ifdef CONFIG_TRACING
slub tracing: move trace calls out of always inlined functions to reduce kernel code size Having the trace calls defined in the always inlined kmalloc functions in include/linux/slub_def.h causes a lot of code duplication as the trace functions get instantiated for each kamalloc call site. This can simply be removed by pushing the trace calls down into the functions in slub.c. On my x86_64 built this patch shrinks the code size of the kernel by approx 36K and also shrinks the code size of many modules -- too many to list here ;) size vmlinux (2.6.36) reports text data bss dec hex filename 5410611 743172 828928 6982711 6a8c37 vmlinux 5373738 744244 828928 6946910 6a005e vmlinux + patch The resulting kernel has had some testing & kmalloc trace still seems to work. This patch - moves trace_kmalloc out of the inlined kmalloc() and pushes it down into kmem_cache_alloc_trace() so this it only get instantiated once. - rename kmem_cache_alloc_notrace() to kmem_cache_alloc_trace() to indicate that now is does have tracing. (maybe this would better being called something like kmalloc_kmem_cache ?) - adds a new function kmalloc_order() to handle allocation and tracing of large allocations of page order. - removes tracing from the inlined kmalloc_large() replacing them with a call to kmalloc_order(); - move tracing out of inlined kmalloc_node() and pushing it down into kmem_cache_alloc_node_trace - rename kmem_cache_alloc_node_notrace() to kmem_cache_alloc_node_trace() - removes the include of trace/events/kmem.h from slub_def.h. v2 - keep kmalloc_order_trace inline when !CONFIG_TRACE Signed-off-by: Richard Kennedy <richard@rsk.demon.co.uk> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2010-10-21 17:29:19 +08:00
void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
gfp_t gfpflags,
slub tracing: move trace calls out of always inlined functions to reduce kernel code size Having the trace calls defined in the always inlined kmalloc functions in include/linux/slub_def.h causes a lot of code duplication as the trace functions get instantiated for each kamalloc call site. This can simply be removed by pushing the trace calls down into the functions in slub.c. On my x86_64 built this patch shrinks the code size of the kernel by approx 36K and also shrinks the code size of many modules -- too many to list here ;) size vmlinux (2.6.36) reports text data bss dec hex filename 5410611 743172 828928 6982711 6a8c37 vmlinux 5373738 744244 828928 6946910 6a005e vmlinux + patch The resulting kernel has had some testing & kmalloc trace still seems to work. This patch - moves trace_kmalloc out of the inlined kmalloc() and pushes it down into kmem_cache_alloc_trace() so this it only get instantiated once. - rename kmem_cache_alloc_notrace() to kmem_cache_alloc_trace() to indicate that now is does have tracing. (maybe this would better being called something like kmalloc_kmem_cache ?) - adds a new function kmalloc_order() to handle allocation and tracing of large allocations of page order. - removes tracing from the inlined kmalloc_large() replacing them with a call to kmalloc_order(); - move tracing out of inlined kmalloc_node() and pushing it down into kmem_cache_alloc_node_trace - rename kmem_cache_alloc_node_notrace() to kmem_cache_alloc_node_trace() - removes the include of trace/events/kmem.h from slub_def.h. v2 - keep kmalloc_order_trace inline when !CONFIG_TRACE Signed-off-by: Richard Kennedy <richard@rsk.demon.co.uk> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2010-10-21 17:29:19 +08:00
int node, size_t size)
{
mm, kfence: insert KFENCE hooks for SLUB Inserts KFENCE hooks into the SLUB allocator. To pass the originally requested size to KFENCE, add an argument 'orig_size' to slab_alloc*(). The additional argument is required to preserve the requested original size for kmalloc() allocations, which uses size classes (e.g. an allocation of 272 bytes will return an object of size 512). Therefore, kmem_cache::size does not represent the kmalloc-caller's requested size, and we must introduce the argument 'orig_size' to propagate the originally requested size to KFENCE. Without the originally requested size, we would not be able to detect out-of-bounds accesses for objects placed at the end of a KFENCE object page if that object is not equal to the kmalloc-size class it was bucketed into. When KFENCE is disabled, there is no additional overhead, since slab_alloc*() functions are __always_inline. Link: https://lkml.kernel.org/r/20201103175841.3495947-6-elver@google.com Signed-off-by: Marco Elver <elver@google.com> Signed-off-by: Alexander Potapenko <glider@google.com> Reviewed-by: Dmitry Vyukov <dvyukov@google.com> Reviewed-by: Jann Horn <jannh@google.com> Co-developed-by: Marco Elver <elver@google.com> Cc: Andrey Konovalov <andreyknvl@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christopher Lameter <cl@linux.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: David Rientjes <rientjes@google.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Hillf Danton <hdanton@sina.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Joern Engel <joern@purestorage.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kees Cook <keescook@chromium.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Paul E. McKenney <paulmck@kernel.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: SeongJae Park <sjpark@amazon.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-26 09:19:16 +08:00
void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_, size);
slub tracing: move trace calls out of always inlined functions to reduce kernel code size Having the trace calls defined in the always inlined kmalloc functions in include/linux/slub_def.h causes a lot of code duplication as the trace functions get instantiated for each kamalloc call site. This can simply be removed by pushing the trace calls down into the functions in slub.c. On my x86_64 built this patch shrinks the code size of the kernel by approx 36K and also shrinks the code size of many modules -- too many to list here ;) size vmlinux (2.6.36) reports text data bss dec hex filename 5410611 743172 828928 6982711 6a8c37 vmlinux 5373738 744244 828928 6946910 6a005e vmlinux + patch The resulting kernel has had some testing & kmalloc trace still seems to work. This patch - moves trace_kmalloc out of the inlined kmalloc() and pushes it down into kmem_cache_alloc_trace() so this it only get instantiated once. - rename kmem_cache_alloc_notrace() to kmem_cache_alloc_trace() to indicate that now is does have tracing. (maybe this would better being called something like kmalloc_kmem_cache ?) - adds a new function kmalloc_order() to handle allocation and tracing of large allocations of page order. - removes tracing from the inlined kmalloc_large() replacing them with a call to kmalloc_order(); - move tracing out of inlined kmalloc_node() and pushing it down into kmem_cache_alloc_node_trace - rename kmem_cache_alloc_node_notrace() to kmem_cache_alloc_node_trace() - removes the include of trace/events/kmem.h from slub_def.h. v2 - keep kmalloc_order_trace inline when !CONFIG_TRACE Signed-off-by: Richard Kennedy <richard@rsk.demon.co.uk> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2010-10-21 17:29:19 +08:00
trace_kmalloc_node(_RET_IP_, ret,
size, s->size, gfpflags, node);
mm: slub: add kernel address sanitizer support for slub allocator With this patch kasan will be able to catch bugs in memory allocated by slub. Initially all objects in newly allocated slab page, marked as redzone. Later, when allocation of slub object happens, requested by caller number of bytes marked as accessible, and the rest of the object (including slub's metadata) marked as redzone (inaccessible). We also mark object as accessible if ksize was called for this object. There is some places in kernel where ksize function is called to inquire size of really allocated area. Such callers could validly access whole allocated memory, so it should be marked as accessible. Code in slub.c and slab_common.c files could validly access to object's metadata, so instrumentation for this files are disabled. Signed-off-by: Andrey Ryabinin <a.ryabinin@samsung.com> Signed-off-by: Dmitry Chernenkov <dmitryc@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Konstantin Serebryany <kcc@google.com> Signed-off-by: Andrey Konovalov <adech.fo@gmail.com> Cc: Yuri Gribov <tetra2005@gmail.com> Cc: Konstantin Khlebnikov <koct9i@gmail.com> Cc: Sasha Levin <sasha.levin@oracle.com> Cc: Christoph Lameter <cl@linux.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Ingo Molnar <mingo@elte.hu> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-14 06:39:42 +08:00
kasan, mm: change hooks signatures Patch series "kasan: add software tag-based mode for arm64", v13. This patchset adds a new software tag-based mode to KASAN [1]. (Initially this mode was called KHWASAN, but it got renamed, see the naming rationale at the end of this section). The plan is to implement HWASan [2] for the kernel with the incentive, that it's going to have comparable to KASAN performance, but in the same time consume much less memory, trading that off for somewhat imprecise bug detection and being supported only for arm64. The underlying ideas of the approach used by software tag-based KASAN are: 1. By using the Top Byte Ignore (TBI) arm64 CPU feature, we can store pointer tags in the top byte of each kernel pointer. 2. Using shadow memory, we can store memory tags for each chunk of kernel memory. 3. On each memory allocation, we can generate a random tag, embed it into the returned pointer and set the memory tags that correspond to this chunk of memory to the same value. 4. By using compiler instrumentation, before each memory access we can add a check that the pointer tag matches the tag of the memory that is being accessed. 5. On a tag mismatch we report an error. With this patchset the existing KASAN mode gets renamed to generic KASAN, with the word "generic" meaning that the implementation can be supported by any architecture as it is purely software. The new mode this patchset adds is called software tag-based KASAN. The word "tag-based" refers to the fact that this mode uses tags embedded into the top byte of kernel pointers and the TBI arm64 CPU feature that allows to dereference such pointers. The word "software" here means that shadow memory manipulation and tag checking on pointer dereference is done in software. As it is the only tag-based implementation right now, "software tag-based" KASAN is sometimes referred to as simply "tag-based" in this patchset. A potential expansion of this mode is a hardware tag-based mode, which would use hardware memory tagging support (announced by Arm [3]) instead of compiler instrumentation and manual shadow memory manipulation. Same as generic KASAN, software tag-based KASAN is strictly a debugging feature. [1] https://www.kernel.org/doc/html/latest/dev-tools/kasan.html [2] http://clang.llvm.org/docs/HardwareAssistedAddressSanitizerDesign.html [3] https://community.arm.com/processors/b/blog/posts/arm-a-profile-architecture-2018-developments-armv85a ====== Rationale On mobile devices generic KASAN's memory usage is significant problem. One of the main reasons to have tag-based KASAN is to be able to perform a similar set of checks as the generic one does, but with lower memory requirements. Comment from Vishwath Mohan <vishwath@google.com>: I don't have data on-hand, but anecdotally both ASAN and KASAN have proven problematic to enable for environments that don't tolerate the increased memory pressure well. This includes (a) Low-memory form factors - Wear, TV, Things, lower-tier phones like Go, (c) Connected components like Pixel's visual core [1]. These are both places I'd love to have a low(er) memory footprint option at my disposal. Comment from Evgenii Stepanov <eugenis@google.com>: Looking at a live Android device under load, slab (according to /proc/meminfo) + kernel stack take 8-10% available RAM (~350MB). KASAN's overhead of 2x - 3x on top of it is not insignificant. Not having this overhead enables near-production use - ex. running KASAN/KHWASAN kernel on a personal, daily-use device to catch bugs that do not reproduce in test configuration. These are the ones that often cost the most engineering time to track down. CPU overhead is bad, but generally tolerable. RAM is critical, in our experience. Once it gets low enough, OOM-killer makes your life miserable. [1] https://www.blog.google/products/pixel/pixel-visual-core-image-processing-and-machine-learning-pixel-2/ ====== Technical details Software tag-based KASAN mode is implemented in a very similar way to the generic one. This patchset essentially does the following: 1. TCR_TBI1 is set to enable Top Byte Ignore. 2. Shadow memory is used (with a different scale, 1:16, so each shadow byte corresponds to 16 bytes of kernel memory) to store memory tags. 3. All slab objects are aligned to shadow scale, which is 16 bytes. 4. All pointers returned from the slab allocator are tagged with a random tag and the corresponding shadow memory is poisoned with the same value. 5. Compiler instrumentation is used to insert tag checks. Either by calling callbacks or by inlining them (CONFIG_KASAN_OUTLINE and CONFIG_KASAN_INLINE flags are reused). 6. When a tag mismatch is detected in callback instrumentation mode KASAN simply prints a bug report. In case of inline instrumentation, clang inserts a brk instruction, and KASAN has it's own brk handler, which reports the bug. 7. The memory in between slab objects is marked with a reserved tag, and acts as a redzone. 8. When a slab object is freed it's marked with a reserved tag. Bug detection is imprecise for two reasons: 1. We won't catch some small out-of-bounds accesses, that fall into the same shadow cell, as the last byte of a slab object. 2. We only have 1 byte to store tags, which means we have a 1/256 probability of a tag match for an incorrect access (actually even slightly less due to reserved tag values). Despite that there's a particular type of bugs that tag-based KASAN can detect compared to generic KASAN: use-after-free after the object has been allocated by someone else. ====== Testing Some kernel developers voiced a concern that changing the top byte of kernel pointers may lead to subtle bugs that are difficult to discover. To address this concern deliberate testing has been performed. It doesn't seem feasible to do some kind of static checking to find potential issues with pointer tagging, so a dynamic approach was taken. All pointer comparisons/subtractions have been instrumented in an LLVM compiler pass and a kernel module that would print a bug report whenever two pointers with different tags are being compared/subtracted (ignoring comparisons with NULL pointers and with pointers obtained by casting an error code to a pointer type) has been used. Then the kernel has been booted in QEMU and on an Odroid C2 board and syzkaller has been run. This yielded the following results. The two places that look interesting are: is_vmalloc_addr in include/linux/mm.h is_kernel_rodata in mm/util.c Here we compare a pointer with some fixed untagged values to make sure that the pointer lies in a particular part of the kernel address space. Since tag-based KASAN doesn't add tags to pointers that belong to rodata or vmalloc regions, this should work as is. To make sure debug checks to those two functions that check that the result doesn't change whether we operate on pointers with or without untagging has been added. A few other cases that don't look that interesting: Comparing pointers to achieve unique sorting order of pointee objects (e.g. sorting locks addresses before performing a double lock): tty_ldisc_lock_pair_timeout in drivers/tty/tty_ldisc.c pipe_double_lock in fs/pipe.c unix_state_double_lock in net/unix/af_unix.c lock_two_nondirectories in fs/inode.c mutex_lock_double in kernel/events/core.c ep_cmp_ffd in fs/eventpoll.c fsnotify_compare_groups fs/notify/mark.c Nothing needs to be done here, since the tags embedded into pointers don't change, so the sorting order would still be unique. Checks that a pointer belongs to some particular allocation: is_sibling_entry in lib/radix-tree.c object_is_on_stack in include/linux/sched/task_stack.h Nothing needs to be done here either, since two pointers can only belong to the same allocation if they have the same tag. Overall, since the kernel boots and works, there are no critical bugs. As for the rest, the traditional kernel testing way (use until fails) is the only one that looks feasible. Another point here is that tag-based KASAN is available under a separate config option that needs to be deliberately enabled. Even though it might be used in a "near-production" environment to find bugs that are not found during fuzzing or running tests, it is still a debug tool. ====== Benchmarks The following numbers were collected on Odroid C2 board. Both generic and tag-based KASAN were used in inline instrumentation mode. Boot time [1]: * ~1.7 sec for clean kernel * ~5.0 sec for generic KASAN * ~5.0 sec for tag-based KASAN Network performance [2]: * 8.33 Gbits/sec for clean kernel * 3.17 Gbits/sec for generic KASAN * 2.85 Gbits/sec for tag-based KASAN Slab memory usage after boot [3]: * ~40 kb for clean kernel * ~105 kb (~260% overhead) for generic KASAN * ~47 kb (~20% overhead) for tag-based KASAN KASAN memory overhead consists of three main parts: 1. Increased slab memory usage due to redzones. 2. Shadow memory (the whole reserved once during boot). 3. Quaratine (grows gradually until some preset limit; the more the limit, the more the chance to detect a use-after-free). Comparing tag-based vs generic KASAN for each of these points: 1. 20% vs 260% overhead. 2. 1/16th vs 1/8th of physical memory. 3. Tag-based KASAN doesn't require quarantine. [1] Time before the ext4 driver is initialized. [2] Measured as `iperf -s & iperf -c 127.0.0.1 -t 30`. [3] Measured as `cat /proc/meminfo | grep Slab`. ====== Some notes A few notes: 1. The patchset can be found here: https://github.com/xairy/kasan-prototype/tree/khwasan 2. Building requires a recent Clang version (7.0.0 or later). 3. Stack instrumentation is not supported yet and will be added later. This patch (of 25): Tag-based KASAN changes the value of the top byte of pointers returned from the kernel allocation functions (such as kmalloc). This patch updates KASAN hooks signatures and their usage in SLAB and SLUB code to reflect that. Link: http://lkml.kernel.org/r/aec2b5e3973781ff8a6bb6760f8543643202c451.1544099024.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Reviewed-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Reviewed-by: Dmitry Vyukov <dvyukov@google.com> Cc: Christoph Lameter <cl@linux.com> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 16:29:37 +08:00
ret = kasan_kmalloc(s, ret, size, gfpflags);
slub tracing: move trace calls out of always inlined functions to reduce kernel code size Having the trace calls defined in the always inlined kmalloc functions in include/linux/slub_def.h causes a lot of code duplication as the trace functions get instantiated for each kamalloc call site. This can simply be removed by pushing the trace calls down into the functions in slub.c. On my x86_64 built this patch shrinks the code size of the kernel by approx 36K and also shrinks the code size of many modules -- too many to list here ;) size vmlinux (2.6.36) reports text data bss dec hex filename 5410611 743172 828928 6982711 6a8c37 vmlinux 5373738 744244 828928 6946910 6a005e vmlinux + patch The resulting kernel has had some testing & kmalloc trace still seems to work. This patch - moves trace_kmalloc out of the inlined kmalloc() and pushes it down into kmem_cache_alloc_trace() so this it only get instantiated once. - rename kmem_cache_alloc_notrace() to kmem_cache_alloc_trace() to indicate that now is does have tracing. (maybe this would better being called something like kmalloc_kmem_cache ?) - adds a new function kmalloc_order() to handle allocation and tracing of large allocations of page order. - removes tracing from the inlined kmalloc_large() replacing them with a call to kmalloc_order(); - move tracing out of inlined kmalloc_node() and pushing it down into kmem_cache_alloc_node_trace - rename kmem_cache_alloc_node_notrace() to kmem_cache_alloc_node_trace() - removes the include of trace/events/kmem.h from slub_def.h. v2 - keep kmalloc_order_trace inline when !CONFIG_TRACE Signed-off-by: Richard Kennedy <richard@rsk.demon.co.uk> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2010-10-21 17:29:19 +08:00
return ret;
}
slub tracing: move trace calls out of always inlined functions to reduce kernel code size Having the trace calls defined in the always inlined kmalloc functions in include/linux/slub_def.h causes a lot of code duplication as the trace functions get instantiated for each kamalloc call site. This can simply be removed by pushing the trace calls down into the functions in slub.c. On my x86_64 built this patch shrinks the code size of the kernel by approx 36K and also shrinks the code size of many modules -- too many to list here ;) size vmlinux (2.6.36) reports text data bss dec hex filename 5410611 743172 828928 6982711 6a8c37 vmlinux 5373738 744244 828928 6946910 6a005e vmlinux + patch The resulting kernel has had some testing & kmalloc trace still seems to work. This patch - moves trace_kmalloc out of the inlined kmalloc() and pushes it down into kmem_cache_alloc_trace() so this it only get instantiated once. - rename kmem_cache_alloc_notrace() to kmem_cache_alloc_trace() to indicate that now is does have tracing. (maybe this would better being called something like kmalloc_kmem_cache ?) - adds a new function kmalloc_order() to handle allocation and tracing of large allocations of page order. - removes tracing from the inlined kmalloc_large() replacing them with a call to kmalloc_order(); - move tracing out of inlined kmalloc_node() and pushing it down into kmem_cache_alloc_node_trace - rename kmem_cache_alloc_node_notrace() to kmem_cache_alloc_node_trace() - removes the include of trace/events/kmem.h from slub_def.h. v2 - keep kmalloc_order_trace inline when !CONFIG_TRACE Signed-off-by: Richard Kennedy <richard@rsk.demon.co.uk> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2010-10-21 17:29:19 +08:00
EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
#endif
#endif /* CONFIG_NUMA */
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/*
* Slow path handling. This may still be called frequently since objects
slub: support concurrent local and remote frees and allocs on a slab Avoid atomic overhead in slab_alloc and slab_free SLUB needs to use the slab_lock for the per cpu slabs to synchronize with potential kfree operations. This patch avoids that need by moving all free objects onto a lockless_freelist. The regular freelist continues to exist and will be used to free objects. So while we consume the lockless_freelist the regular freelist may build up objects. If we are out of objects on the lockless_freelist then we may check the regular freelist. If it has objects then we move those over to the lockless_freelist and do this again. There is a significant savings in terms of atomic operations that have to be performed. We can even free directly to the lockless_freelist if we know that we are running on the same processor. So this speeds up short lived objects. They may be allocated and freed without taking the slab_lock. This is particular good for netperf. In order to maximize the effect of the new faster hotpath we extract the hottest performance pieces into inlined functions. These are then inlined into kmem_cache_alloc and kmem_cache_free. So hotpath allocation and freeing no longer requires a subroutine call within SLUB. [I am not sure that it is worth doing this because it changes the easy to read structure of slub just to reduce atomic ops. However, there is someone out there with a benchmark on 4 way and 8 way processor systems that seems to show a 5% regression vs. Slab. Seems that the regression is due to increased atomic operations use vs. SLAB in SLUB). I wonder if this is applicable or discernable at all in a real workload?] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-10 18:15:16 +08:00
* have a longer lifetime than the cpu slabs in most processing loads.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*
slub: support concurrent local and remote frees and allocs on a slab Avoid atomic overhead in slab_alloc and slab_free SLUB needs to use the slab_lock for the per cpu slabs to synchronize with potential kfree operations. This patch avoids that need by moving all free objects onto a lockless_freelist. The regular freelist continues to exist and will be used to free objects. So while we consume the lockless_freelist the regular freelist may build up objects. If we are out of objects on the lockless_freelist then we may check the regular freelist. If it has objects then we move those over to the lockless_freelist and do this again. There is a significant savings in terms of atomic operations that have to be performed. We can even free directly to the lockless_freelist if we know that we are running on the same processor. So this speeds up short lived objects. They may be allocated and freed without taking the slab_lock. This is particular good for netperf. In order to maximize the effect of the new faster hotpath we extract the hottest performance pieces into inlined functions. These are then inlined into kmem_cache_alloc and kmem_cache_free. So hotpath allocation and freeing no longer requires a subroutine call within SLUB. [I am not sure that it is worth doing this because it changes the easy to read structure of slub just to reduce atomic ops. However, there is someone out there with a benchmark on 4 way and 8 way processor systems that seems to show a 5% regression vs. Slab. Seems that the regression is due to increased atomic operations use vs. SLAB in SLUB). I wonder if this is applicable or discernable at all in a real workload?] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-10 18:15:16 +08:00
* So we still attempt to reduce cache line usage. Just take the slab
* lock and free the item. If there is no additional partial page
* handling required then we can return immediately.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*/
slub: support concurrent local and remote frees and allocs on a slab Avoid atomic overhead in slab_alloc and slab_free SLUB needs to use the slab_lock for the per cpu slabs to synchronize with potential kfree operations. This patch avoids that need by moving all free objects onto a lockless_freelist. The regular freelist continues to exist and will be used to free objects. So while we consume the lockless_freelist the regular freelist may build up objects. If we are out of objects on the lockless_freelist then we may check the regular freelist. If it has objects then we move those over to the lockless_freelist and do this again. There is a significant savings in terms of atomic operations that have to be performed. We can even free directly to the lockless_freelist if we know that we are running on the same processor. So this speeds up short lived objects. They may be allocated and freed without taking the slab_lock. This is particular good for netperf. In order to maximize the effect of the new faster hotpath we extract the hottest performance pieces into inlined functions. These are then inlined into kmem_cache_alloc and kmem_cache_free. So hotpath allocation and freeing no longer requires a subroutine call within SLUB. [I am not sure that it is worth doing this because it changes the easy to read structure of slub just to reduce atomic ops. However, there is someone out there with a benchmark on 4 way and 8 way processor systems that seems to show a 5% regression vs. Slab. Seems that the regression is due to increased atomic operations use vs. SLAB in SLUB). I wonder if this is applicable or discernable at all in a real workload?] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-10 18:15:16 +08:00
static void __slab_free(struct kmem_cache *s, struct page *page,
void *head, void *tail, int cnt,
unsigned long addr)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
{
void *prior;
int was_frozen;
struct page new;
unsigned long counters;
struct kmem_cache_node *n = NULL;
treewide: Remove uninitialized_var() usage Using uninitialized_var() is dangerous as it papers over real bugs[1] (or can in the future), and suppresses unrelated compiler warnings (e.g. "unused variable"). If the compiler thinks it is uninitialized, either simply initialize the variable or make compiler changes. In preparation for removing[2] the[3] macro[4], remove all remaining needless uses with the following script: git grep '\buninitialized_var\b' | cut -d: -f1 | sort -u | \ xargs perl -pi -e \ 's/\buninitialized_var\(([^\)]+)\)/\1/g; s:\s*/\* (GCC be quiet|to make compiler happy) \*/$::g;' drivers/video/fbdev/riva/riva_hw.c was manually tweaked to avoid pathological white-space. No outstanding warnings were found building allmodconfig with GCC 9.3.0 for x86_64, i386, arm64, arm, powerpc, powerpc64le, s390x, mips, sparc64, alpha, and m68k. [1] https://lore.kernel.org/lkml/20200603174714.192027-1-glider@google.com/ [2] https://lore.kernel.org/lkml/CA+55aFw+Vbj0i=1TGqCR5vQkCzWJ0QxK6CernOU6eedsudAixw@mail.gmail.com/ [3] https://lore.kernel.org/lkml/CA+55aFwgbgqhbp1fkxvRKEpzyR5J8n1vKT1VZdz9knmPuXhOeg@mail.gmail.com/ [4] https://lore.kernel.org/lkml/CA+55aFz2500WfbKXAx8s67wrm9=yVJu65TpLgN_ybYNv0VEOKA@mail.gmail.com/ Reviewed-by: Leon Romanovsky <leonro@mellanox.com> # drivers/infiniband and mlx4/mlx5 Acked-by: Jason Gunthorpe <jgg@mellanox.com> # IB Acked-by: Kalle Valo <kvalo@codeaurora.org> # wireless drivers Reviewed-by: Chao Yu <yuchao0@huawei.com> # erofs Signed-off-by: Kees Cook <keescook@chromium.org>
2020-06-04 04:09:38 +08:00
unsigned long flags;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
Lockless (and preemptless) fastpaths for slub Use the this_cpu_cmpxchg_double functionality to implement a lockless allocation algorithm on arches that support fast this_cpu_ops. Each of the per cpu pointers is paired with a transaction id that ensures that updates of the per cpu information can only occur in sequence on a certain cpu. A transaction id is a "long" integer that is comprised of an event number and the cpu number. The event number is incremented for every change to the per cpu state. This means that the cmpxchg instruction can verify for an update that nothing interfered and that we are updating the percpu structure for the processor where we picked up the information and that we are also currently on that processor when we update the information. This results in a significant decrease of the overhead in the fastpaths. It also makes it easy to adopt the fast path for realtime kernels since this is lockless and does not require the use of the current per cpu area over the critical section. It is only important that the per cpu area is current at the beginning of the critical section and at the end. So there is no need even to disable preemption. Test results show that the fastpath cycle count is reduced by up to ~ 40% (alloc/free test goes from ~140 cycles down to ~80). The slowpath for kfree adds a few cycles. Sadly this does nothing for the slowpath which is where the main issues with performance in slub are but the best case performance rises significantly. (For that see the more complex slub patches that require cmpxchg_double) Kmalloc: alloc/free test Before: 10000 times kmalloc(8)/kfree -> 134 cycles 10000 times kmalloc(16)/kfree -> 152 cycles 10000 times kmalloc(32)/kfree -> 144 cycles 10000 times kmalloc(64)/kfree -> 142 cycles 10000 times kmalloc(128)/kfree -> 142 cycles 10000 times kmalloc(256)/kfree -> 132 cycles 10000 times kmalloc(512)/kfree -> 132 cycles 10000 times kmalloc(1024)/kfree -> 135 cycles 10000 times kmalloc(2048)/kfree -> 135 cycles 10000 times kmalloc(4096)/kfree -> 135 cycles 10000 times kmalloc(8192)/kfree -> 144 cycles 10000 times kmalloc(16384)/kfree -> 754 cycles After: 10000 times kmalloc(8)/kfree -> 78 cycles 10000 times kmalloc(16)/kfree -> 78 cycles 10000 times kmalloc(32)/kfree -> 82 cycles 10000 times kmalloc(64)/kfree -> 88 cycles 10000 times kmalloc(128)/kfree -> 79 cycles 10000 times kmalloc(256)/kfree -> 79 cycles 10000 times kmalloc(512)/kfree -> 85 cycles 10000 times kmalloc(1024)/kfree -> 82 cycles 10000 times kmalloc(2048)/kfree -> 82 cycles 10000 times kmalloc(4096)/kfree -> 85 cycles 10000 times kmalloc(8192)/kfree -> 82 cycles 10000 times kmalloc(16384)/kfree -> 706 cycles Kmalloc: Repeatedly allocate then free test Before: 10000 times kmalloc(8) -> 211 cycles kfree -> 113 cycles 10000 times kmalloc(16) -> 174 cycles kfree -> 115 cycles 10000 times kmalloc(32) -> 235 cycles kfree -> 129 cycles 10000 times kmalloc(64) -> 222 cycles kfree -> 120 cycles 10000 times kmalloc(128) -> 343 cycles kfree -> 139 cycles 10000 times kmalloc(256) -> 827 cycles kfree -> 147 cycles 10000 times kmalloc(512) -> 1048 cycles kfree -> 272 cycles 10000 times kmalloc(1024) -> 2043 cycles kfree -> 528 cycles 10000 times kmalloc(2048) -> 4002 cycles kfree -> 571 cycles 10000 times kmalloc(4096) -> 7740 cycles kfree -> 628 cycles 10000 times kmalloc(8192) -> 8062 cycles kfree -> 850 cycles 10000 times kmalloc(16384) -> 8895 cycles kfree -> 1249 cycles After: 10000 times kmalloc(8) -> 190 cycles kfree -> 129 cycles 10000 times kmalloc(16) -> 76 cycles kfree -> 123 cycles 10000 times kmalloc(32) -> 126 cycles kfree -> 124 cycles 10000 times kmalloc(64) -> 181 cycles kfree -> 128 cycles 10000 times kmalloc(128) -> 310 cycles kfree -> 140 cycles 10000 times kmalloc(256) -> 809 cycles kfree -> 165 cycles 10000 times kmalloc(512) -> 1005 cycles kfree -> 269 cycles 10000 times kmalloc(1024) -> 1999 cycles kfree -> 527 cycles 10000 times kmalloc(2048) -> 3967 cycles kfree -> 570 cycles 10000 times kmalloc(4096) -> 7658 cycles kfree -> 637 cycles 10000 times kmalloc(8192) -> 8111 cycles kfree -> 859 cycles 10000 times kmalloc(16384) -> 8791 cycles kfree -> 1173 cycles Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-02-26 01:38:54 +08:00
stat(s, FREE_SLOWPATH);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
mm, kfence: insert KFENCE hooks for SLUB Inserts KFENCE hooks into the SLUB allocator. To pass the originally requested size to KFENCE, add an argument 'orig_size' to slab_alloc*(). The additional argument is required to preserve the requested original size for kmalloc() allocations, which uses size classes (e.g. an allocation of 272 bytes will return an object of size 512). Therefore, kmem_cache::size does not represent the kmalloc-caller's requested size, and we must introduce the argument 'orig_size' to propagate the originally requested size to KFENCE. Without the originally requested size, we would not be able to detect out-of-bounds accesses for objects placed at the end of a KFENCE object page if that object is not equal to the kmalloc-size class it was bucketed into. When KFENCE is disabled, there is no additional overhead, since slab_alloc*() functions are __always_inline. Link: https://lkml.kernel.org/r/20201103175841.3495947-6-elver@google.com Signed-off-by: Marco Elver <elver@google.com> Signed-off-by: Alexander Potapenko <glider@google.com> Reviewed-by: Dmitry Vyukov <dvyukov@google.com> Reviewed-by: Jann Horn <jannh@google.com> Co-developed-by: Marco Elver <elver@google.com> Cc: Andrey Konovalov <andreyknvl@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christopher Lameter <cl@linux.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: David Rientjes <rientjes@google.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Hillf Danton <hdanton@sina.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Joern Engel <joern@purestorage.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kees Cook <keescook@chromium.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Paul E. McKenney <paulmck@kernel.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: SeongJae Park <sjpark@amazon.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-26 09:19:16 +08:00
if (kfence_free(head))
return;
if (kmem_cache_debug(s) &&
slub: drop lock at the end of free_debug_processing This series takes the suggestion of Christoph Lameter and only focuses on optimizing the slow path where the debug processing runs. The two main optimizations in this series are letting the consistency checks be skipped and relaxing the cmpxchg restrictions when we are not doing consistency checks. With hackbench -g 20 -l 1000 averaged over 100 runs: Before slub_debug=P mean 15.607 variance .086 stdev .294 After slub_debug=P mean 10.836 variance .155 stdev .394 This still isn't as fast as what is in grsecurity unfortunately so there's still work to be done. Profiling ___slab_alloc shows that 25-50% of time is spent in deactivate_slab. I haven't looked too closely to see if this is something that can be optimized. My plan for now is to focus on getting all of this merged (if appropriate) before digging in to another task. This patch (of 4): Currently, free_debug_processing has a comment "Keep node_lock to preserve integrity until the object is actually freed". In actuallity, the lock is dropped immediately in __slab_free. Rather than wait until __slab_free and potentially throw off the unlikely marking, just drop the lock in __slab_free. This also lets free_debug_processing take its own copy of the spinlock flags rather than trying to share the ones from __slab_free. Since there is no use for the node afterwards, change the return type of free_debug_processing to return an int like alloc_debug_processing. Credit to Mathias Krause for the original work which inspired this series [akpm@linux-foundation.org: fix build] Signed-off-by: Laura Abbott <labbott@fedoraproject.org> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <js1304@gmail.com> Cc: Kees Cook <keescook@chromium.org> Cc: Mathias Krause <minipli@googlemail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-03-16 05:54:59 +08:00
!free_debug_processing(s, page, head, tail, cnt, addr))
return;
do {
if (unlikely(n)) {
spin_unlock_irqrestore(&n->list_lock, flags);
n = NULL;
}
prior = page->freelist;
counters = page->counters;
set_freepointer(s, tail, prior);
new.counters = counters;
was_frozen = new.frozen;
new.inuse -= cnt;
if ((!new.inuse || !prior) && !was_frozen) {
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
if (kmem_cache_has_cpu_partial(s) && !prior) {
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
/*
* Slab was on no list before and will be
* partially empty
* We can defer the list move and instead
* freeze it.
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
*/
new.frozen = 1;
} else { /* Needs to be taken off a list */
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
n = get_node(s, page_to_nid(page));
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
/*
* Speculatively acquire the list_lock.
* If the cmpxchg does not succeed then we may
* drop the list_lock without any processing.
*
* Otherwise the list_lock will synchronize with
* other processors updating the list of slabs.
*/
spin_lock_irqsave(&n->list_lock, flags);
}
}
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
} while (!cmpxchg_double_slab(s, page,
prior, counters,
head, new.counters,
"__slab_free"));
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (likely(!n)) {
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
if (likely(was_frozen)) {
/*
* The list lock was not taken therefore no list
* activity can be necessary.
*/
stat(s, FREE_FROZEN);
} else if (new.frozen) {
/*
* If we just froze the page then put it onto the
* per cpu partial list.
*/
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
put_cpu_partial(s, page, 1);
stat(s, CPU_PARTIAL_FREE);
}
return;
}
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
goto slab_empty;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/*
* Objects left in the slab. If it was not on the partial list before
* then add it.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*/
if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
remove_full(s, n, page);
add_partial(n, page, DEACTIVATE_TO_TAIL);
stat(s, FREE_ADD_PARTIAL);
SLUB: Support for performance statistics The statistics provided here allow the monitoring of allocator behavior but at the cost of some (minimal) loss of performance. Counters are placed in SLUB's per cpu data structure. The per cpu structure may be extended by the statistics to grow larger than one cacheline which will increase the cache footprint of SLUB. There is a compile option to enable/disable the inclusion of the runtime statistics and its off by default. The slabinfo tool is enhanced to support these statistics via two options: -D Switches the line of information displayed for a slab from size mode to activity mode. -A Sorts the slabs displayed by activity. This allows the display of the slabs most important to the performance of a certain load. -r Report option will report detailed statistics on Example (tbench load): slabinfo -AD ->Shows the most active slabs Name Objects Alloc Free %Fast skbuff_fclone_cache 33 111953835 111953835 99 99 :0000192 2666 5283688 5281047 99 99 :0001024 849 5247230 5246389 83 83 vm_area_struct 1349 119642 118355 91 22 :0004096 15 66753 66751 98 98 :0000064 2067 25297 23383 98 78 dentry 10259 28635 18464 91 45 :0000080 11004 18950 8089 98 98 :0000096 1703 12358 10784 99 98 :0000128 762 10582 9875 94 18 :0000512 184 9807 9647 95 81 :0002048 479 9669 9195 83 65 anon_vma 777 9461 9002 99 71 kmalloc-8 6492 9981 5624 99 97 :0000768 258 7174 6931 58 15 So the skbuff_fclone_cache is of highest importance for the tbench load. Pretty high load on the 192 sized slab. Look for the aliases slabinfo -a | grep 000192 :0000192 <- xfs_btree_cur filp kmalloc-192 uid_cache tw_sock_TCP request_sock_TCPv6 tw_sock_TCPv6 skbuff_head_cache xfs_ili Likely skbuff_head_cache. Looking into the statistics of the skbuff_fclone_cache is possible through slabinfo skbuff_fclone_cache ->-r option implied if cache name is mentioned .... Usual output ... Slab Perf Counter Alloc Free %Al %Fr -------------------------------------------------- Fastpath 111953360 111946981 99 99 Slowpath 1044 7423 0 0 Page Alloc 272 264 0 0 Add partial 25 325 0 0 Remove partial 86 264 0 0 RemoteObj/SlabFrozen 350 4832 0 0 Total 111954404 111954404 Flushes 49 Refill 0 Deactivate Full=325(92%) Empty=0(0%) ToHead=24(6%) ToTail=1(0%) Looks good because the fastpath is overwhelmingly taken. skbuff_head_cache: Slab Perf Counter Alloc Free %Al %Fr -------------------------------------------------- Fastpath 5297262 5259882 99 99 Slowpath 4477 39586 0 0 Page Alloc 937 824 0 0 Add partial 0 2515 0 0 Remove partial 1691 824 0 0 RemoteObj/SlabFrozen 2621 9684 0 0 Total 5301739 5299468 Deactivate Full=2620(100%) Empty=0(0%) ToHead=0(0%) ToTail=0(0%) Descriptions of the output: Total: The total number of allocation and frees that occurred for a slab Fastpath: The number of allocations/frees that used the fastpath. Slowpath: Other allocations Page Alloc: Number of calls to the page allocator as a result of slowpath processing Add Partial: Number of slabs added to the partial list through free or alloc (occurs during cpuslab flushes) Remove Partial: Number of slabs removed from the partial list as a result of allocations retrieving a partial slab or by a free freeing the last object of a slab. RemoteObj/Froz: How many times were remotely freed object encountered when a slab was about to be deactivated. Frozen: How many times was free able to skip list processing because the slab was in use as the cpuslab of another processor. Flushes: Number of times the cpuslab was flushed on request (kmem_cache_shrink, may result from races in __slab_alloc) Refill: Number of times we were able to refill the cpuslab from remotely freed objects for the same slab. Deactivate: Statistics how slabs were deactivated. Shows how they were put onto the partial list. In general fastpath is very good. Slowpath without partial list processing is also desirable. Any touching of partial list uses node specific locks which may potentially cause list lock contention. Signed-off-by: Christoph Lameter <clameter@sgi.com>
2008-02-08 09:47:41 +08:00
}
spin_unlock_irqrestore(&n->list_lock, flags);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
return;
slab_empty:
if (prior) {
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/*
slub: Fix full list corruption if debugging is on When a slab is freed by __slab_free() and the slab can only contain a single object ever then it was full (and therefore not on the partial lists but on the full list in the debug case) before we reached slab_empty. This caused the following full list corruption when SLUB debugging was enabled: [ 5913.233035] ------------[ cut here ]------------ [ 5913.233097] WARNING: at lib/list_debug.c:53 __list_del_entry+0x8d/0x98() [ 5913.233101] Hardware name: Adamo 13 [ 5913.233105] list_del corruption. prev->next should be ffffea000434fd20, but was ffffea0004199520 [ 5913.233108] Modules linked in: nfs fscache fuse ebtable_nat ebtables ppdev parport_pc lp parport ipt_MASQUERADE iptable_nat nf_nat nfsd lockd nfs_acl auth_rpcgss xt_CHECKSUM sunrpc iptable_mangle bridge stp llc cpufreq_ondemand acpi_cpufreq freq_table mperf ip6t_REJECT nf_conntrack_ipv6 nf_defrag_ipv6 ip6table_filter ip6_tables rfcomm bnep arc4 iwlagn snd_hda_codec_hdmi snd_hda_codec_idt snd_hda_intel btusb mac80211 snd_hda_codec bluetooth snd_hwdep snd_seq snd_seq_device snd_pcm usb_debug dell_wmi sparse_keymap cdc_ether usbnet cdc_acm uvcvideo cdc_wdm mii cfg80211 snd_timer dell_laptop videodev dcdbas snd microcode v4l2_compat_ioctl32 soundcore joydev tg3 pcspkr snd_page_alloc iTCO_wdt i2c_i801 rfkill iTCO_vendor_support wmi virtio_net kvm_intel kvm ipv6 xts gf128mul dm_crypt i915 drm_kms_helper drm i2c_algo_bit i2c_core video [last unloaded: scsi_wait_scan] [ 5913.233213] Pid: 0, comm: swapper Not tainted 3.0.0+ #127 [ 5913.233213] Call Trace: [ 5913.233213] <IRQ> [<ffffffff8105df18>] warn_slowpath_common+0x83/0x9b [ 5913.233213] [<ffffffff8105dfd3>] warn_slowpath_fmt+0x46/0x48 [ 5913.233213] [<ffffffff8127e7c1>] __list_del_entry+0x8d/0x98 [ 5913.233213] [<ffffffff8127e7da>] list_del+0xe/0x2d [ 5913.233213] [<ffffffff814e0430>] __slab_free+0x1db/0x235 [ 5913.233213] [<ffffffff811706ab>] ? bvec_free_bs+0x35/0x37 [ 5913.233213] [<ffffffff811706ab>] ? bvec_free_bs+0x35/0x37 [ 5913.233213] [<ffffffff811706ab>] ? bvec_free_bs+0x35/0x37 [ 5913.233213] [<ffffffff81133085>] kmem_cache_free+0x88/0x102 [ 5913.233213] [<ffffffff811706ab>] bvec_free_bs+0x35/0x37 [ 5913.233213] [<ffffffff811706e1>] bio_free+0x34/0x64 [ 5913.233213] [<ffffffff813dc390>] dm_bio_destructor+0x12/0x14 [ 5913.233213] [<ffffffff8116fef6>] bio_put+0x2b/0x2d [ 5913.233213] [<ffffffff813dccab>] clone_endio+0x9e/0xb4 [ 5913.233213] [<ffffffff8116f7dd>] bio_endio+0x2d/0x2f [ 5913.233213] [<ffffffffa00148da>] crypt_dec_pending+0x5c/0x8b [dm_crypt] [ 5913.233213] [<ffffffffa00150a9>] crypt_endio+0x78/0x81 [dm_crypt] [ Full discussion here: https://lkml.org/lkml/2011/8/4/375 ] Make sure that we remove such a slab also from the full lists. Reported-and-tested-by: Dave Jones <davej@redhat.com> Reported-and-tested-by: Xiaotian Feng <xtfeng@gmail.com> Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-09 00:16:56 +08:00
* Slab on the partial list.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*/
remove_partial(n, page);
stat(s, FREE_REMOVE_PARTIAL);
} else {
slub: Fix full list corruption if debugging is on When a slab is freed by __slab_free() and the slab can only contain a single object ever then it was full (and therefore not on the partial lists but on the full list in the debug case) before we reached slab_empty. This caused the following full list corruption when SLUB debugging was enabled: [ 5913.233035] ------------[ cut here ]------------ [ 5913.233097] WARNING: at lib/list_debug.c:53 __list_del_entry+0x8d/0x98() [ 5913.233101] Hardware name: Adamo 13 [ 5913.233105] list_del corruption. prev->next should be ffffea000434fd20, but was ffffea0004199520 [ 5913.233108] Modules linked in: nfs fscache fuse ebtable_nat ebtables ppdev parport_pc lp parport ipt_MASQUERADE iptable_nat nf_nat nfsd lockd nfs_acl auth_rpcgss xt_CHECKSUM sunrpc iptable_mangle bridge stp llc cpufreq_ondemand acpi_cpufreq freq_table mperf ip6t_REJECT nf_conntrack_ipv6 nf_defrag_ipv6 ip6table_filter ip6_tables rfcomm bnep arc4 iwlagn snd_hda_codec_hdmi snd_hda_codec_idt snd_hda_intel btusb mac80211 snd_hda_codec bluetooth snd_hwdep snd_seq snd_seq_device snd_pcm usb_debug dell_wmi sparse_keymap cdc_ether usbnet cdc_acm uvcvideo cdc_wdm mii cfg80211 snd_timer dell_laptop videodev dcdbas snd microcode v4l2_compat_ioctl32 soundcore joydev tg3 pcspkr snd_page_alloc iTCO_wdt i2c_i801 rfkill iTCO_vendor_support wmi virtio_net kvm_intel kvm ipv6 xts gf128mul dm_crypt i915 drm_kms_helper drm i2c_algo_bit i2c_core video [last unloaded: scsi_wait_scan] [ 5913.233213] Pid: 0, comm: swapper Not tainted 3.0.0+ #127 [ 5913.233213] Call Trace: [ 5913.233213] <IRQ> [<ffffffff8105df18>] warn_slowpath_common+0x83/0x9b [ 5913.233213] [<ffffffff8105dfd3>] warn_slowpath_fmt+0x46/0x48 [ 5913.233213] [<ffffffff8127e7c1>] __list_del_entry+0x8d/0x98 [ 5913.233213] [<ffffffff8127e7da>] list_del+0xe/0x2d [ 5913.233213] [<ffffffff814e0430>] __slab_free+0x1db/0x235 [ 5913.233213] [<ffffffff811706ab>] ? bvec_free_bs+0x35/0x37 [ 5913.233213] [<ffffffff811706ab>] ? bvec_free_bs+0x35/0x37 [ 5913.233213] [<ffffffff811706ab>] ? bvec_free_bs+0x35/0x37 [ 5913.233213] [<ffffffff81133085>] kmem_cache_free+0x88/0x102 [ 5913.233213] [<ffffffff811706ab>] bvec_free_bs+0x35/0x37 [ 5913.233213] [<ffffffff811706e1>] bio_free+0x34/0x64 [ 5913.233213] [<ffffffff813dc390>] dm_bio_destructor+0x12/0x14 [ 5913.233213] [<ffffffff8116fef6>] bio_put+0x2b/0x2d [ 5913.233213] [<ffffffff813dccab>] clone_endio+0x9e/0xb4 [ 5913.233213] [<ffffffff8116f7dd>] bio_endio+0x2d/0x2f [ 5913.233213] [<ffffffffa00148da>] crypt_dec_pending+0x5c/0x8b [dm_crypt] [ 5913.233213] [<ffffffffa00150a9>] crypt_endio+0x78/0x81 [dm_crypt] [ Full discussion here: https://lkml.org/lkml/2011/8/4/375 ] Make sure that we remove such a slab also from the full lists. Reported-and-tested-by: Dave Jones <davej@redhat.com> Reported-and-tested-by: Xiaotian Feng <xtfeng@gmail.com> Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-09 00:16:56 +08:00
/* Slab must be on the full list */
remove_full(s, n, page);
}
spin_unlock_irqrestore(&n->list_lock, flags);
stat(s, FREE_SLAB);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
discard_slab(s, page);
}
slub: support concurrent local and remote frees and allocs on a slab Avoid atomic overhead in slab_alloc and slab_free SLUB needs to use the slab_lock for the per cpu slabs to synchronize with potential kfree operations. This patch avoids that need by moving all free objects onto a lockless_freelist. The regular freelist continues to exist and will be used to free objects. So while we consume the lockless_freelist the regular freelist may build up objects. If we are out of objects on the lockless_freelist then we may check the regular freelist. If it has objects then we move those over to the lockless_freelist and do this again. There is a significant savings in terms of atomic operations that have to be performed. We can even free directly to the lockless_freelist if we know that we are running on the same processor. So this speeds up short lived objects. They may be allocated and freed without taking the slab_lock. This is particular good for netperf. In order to maximize the effect of the new faster hotpath we extract the hottest performance pieces into inlined functions. These are then inlined into kmem_cache_alloc and kmem_cache_free. So hotpath allocation and freeing no longer requires a subroutine call within SLUB. [I am not sure that it is worth doing this because it changes the easy to read structure of slub just to reduce atomic ops. However, there is someone out there with a benchmark on 4 way and 8 way processor systems that seems to show a 5% regression vs. Slab. Seems that the regression is due to increased atomic operations use vs. SLAB in SLUB). I wonder if this is applicable or discernable at all in a real workload?] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-10 18:15:16 +08:00
/*
* Fastpath with forced inlining to produce a kfree and kmem_cache_free that
* can perform fastpath freeing without additional function calls.
*
* The fastpath is only possible if we are freeing to the current cpu slab
* of this processor. This typically the case if we have just allocated
* the item before.
*
* If fastpath is not possible then fall back to __slab_free where we deal
* with all sorts of special processing.
*
* Bulk free of a freelist with several objects (all pointing to the
* same page) possible by specifying head and tail ptr, plus objects
* count (cnt). Bulk free indicated by tail pointer being set.
slub: support concurrent local and remote frees and allocs on a slab Avoid atomic overhead in slab_alloc and slab_free SLUB needs to use the slab_lock for the per cpu slabs to synchronize with potential kfree operations. This patch avoids that need by moving all free objects onto a lockless_freelist. The regular freelist continues to exist and will be used to free objects. So while we consume the lockless_freelist the regular freelist may build up objects. If we are out of objects on the lockless_freelist then we may check the regular freelist. If it has objects then we move those over to the lockless_freelist and do this again. There is a significant savings in terms of atomic operations that have to be performed. We can even free directly to the lockless_freelist if we know that we are running on the same processor. So this speeds up short lived objects. They may be allocated and freed without taking the slab_lock. This is particular good for netperf. In order to maximize the effect of the new faster hotpath we extract the hottest performance pieces into inlined functions. These are then inlined into kmem_cache_alloc and kmem_cache_free. So hotpath allocation and freeing no longer requires a subroutine call within SLUB. [I am not sure that it is worth doing this because it changes the easy to read structure of slub just to reduce atomic ops. However, there is someone out there with a benchmark on 4 way and 8 way processor systems that seems to show a 5% regression vs. Slab. Seems that the regression is due to increased atomic operations use vs. SLAB in SLUB). I wonder if this is applicable or discernable at all in a real workload?] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-10 18:15:16 +08:00
*/
static __always_inline void do_slab_free(struct kmem_cache *s,
struct page *page, void *head, void *tail,
int cnt, unsigned long addr)
slub: support concurrent local and remote frees and allocs on a slab Avoid atomic overhead in slab_alloc and slab_free SLUB needs to use the slab_lock for the per cpu slabs to synchronize with potential kfree operations. This patch avoids that need by moving all free objects onto a lockless_freelist. The regular freelist continues to exist and will be used to free objects. So while we consume the lockless_freelist the regular freelist may build up objects. If we are out of objects on the lockless_freelist then we may check the regular freelist. If it has objects then we move those over to the lockless_freelist and do this again. There is a significant savings in terms of atomic operations that have to be performed. We can even free directly to the lockless_freelist if we know that we are running on the same processor. So this speeds up short lived objects. They may be allocated and freed without taking the slab_lock. This is particular good for netperf. In order to maximize the effect of the new faster hotpath we extract the hottest performance pieces into inlined functions. These are then inlined into kmem_cache_alloc and kmem_cache_free. So hotpath allocation and freeing no longer requires a subroutine call within SLUB. [I am not sure that it is worth doing this because it changes the easy to read structure of slub just to reduce atomic ops. However, there is someone out there with a benchmark on 4 way and 8 way processor systems that seems to show a 5% regression vs. Slab. Seems that the regression is due to increased atomic operations use vs. SLAB in SLUB). I wonder if this is applicable or discernable at all in a real workload?] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-10 18:15:16 +08:00
{
void *tail_obj = tail ? : head;
SLUB: Avoid page struct cacheline bouncing due to remote frees to cpu slab A remote free may access the same page struct that also contains the lockless freelist for the cpu slab. If objects have a short lifetime and are freed by a different processor then remote frees back to the slab from which we are currently allocating are frequent. The cacheline with the page struct needs to be repeately acquired in exclusive mode by both the allocating thread and the freeing thread. If this is frequent enough then performance will suffer because of cacheline bouncing. This patchset puts the lockless_freelist pointer in its own cacheline. In order to make that happen we introduce a per cpu structure called kmem_cache_cpu. Instead of keeping an array of pointers to page structs we now keep an array to a per cpu structure that--among other things--contains the pointer to the lockless freelist. The freeing thread can then keep possession of exclusive access to the page struct cacheline while the allocating thread keeps its exclusive access to the cacheline containing the per cpu structure. This works as long as the allocating cpu is able to service its request from the lockless freelist. If the lockless freelist runs empty then the allocating thread needs to acquire exclusive access to the cacheline with the page struct lock the slab. The allocating thread will then check if new objects were freed to the per cpu slab. If so it will keep the slab as the cpu slab and continue with the recently remote freed objects. So the allocating thread can take a series of just freed remote pages and dish them out again. Ideally allocations could be just recycling objects in the same slab this way which will lead to an ideal allocation / remote free pattern. The number of objects that can be handled in this way is limited by the capacity of one slab. Increasing slab size via slub_min_objects/ slub_max_order may increase the number of objects and therefore performance. If the allocating thread runs out of objects and finds that no objects were put back by the remote processor then it will retrieve a new slab (from the partial lists or from the page allocator) and start with a whole new set of objects while the remote thread may still be freeing objects to the old cpu slab. This may then repeat until the new slab is also exhausted. If remote freeing has freed objects in the earlier slab then that earlier slab will now be on the partial freelist and the allocating thread will pick that slab next for allocation. So the loop is extended. However, both threads need to take the list_lock to make the swizzling via the partial list happen. It is likely that this kind of scheme will keep the objects being passed around to a small set that can be kept in the cpu caches leading to increased performance. More code cleanups become possible: - Instead of passing a cpu we can now pass a kmem_cache_cpu structure around. Allows reducing the number of parameters to various functions. - Can define a new node_match() function for NUMA to encapsulate locality checks. Effect on allocations: Cachelines touched before this patch: Write: page cache struct and first cacheline of object Cachelines touched after this patch: Write: kmem_cache_cpu cacheline and first cacheline of object Read: page cache struct (but see later patch that avoids touching that cacheline) The handling when the lockless alloc list runs empty gets to be a bit more complicated since another cacheline has now to be written to. But that is halfway out of the hot path. Effect on freeing: Cachelines touched before this patch: Write: page_struct and first cacheline of object Cachelines touched after this patch depending on how we free: Write(to cpu_slab): kmem_cache_cpu struct and first cacheline of object Write(to other): page struct and first cacheline of object Read(to cpu_slab): page struct to id slab etc. (but see later patch that avoids touching the page struct on free) Read(to other): cpu local kmem_cache_cpu struct to verify its not the cpu slab. Summary: Pro: - Distinct cachelines so that concurrent remote frees and local allocs on a cpuslab can occur without cacheline bouncing. - Avoids potential bouncing cachelines because of neighboring per cpu pointer updates in kmem_cache's cpu_slab structure since it now grows to a cacheline (Therefore remove the comment that talks about that concern). Cons: - Freeing objects now requires the reading of one additional cacheline. That can be mitigated for some cases by the following patches but its not possible to completely eliminate these references. - Memory usage grows slightly. The size of each per cpu object is blown up from one word (pointing to the page_struct) to one cacheline with various data. So this is NR_CPUS*NR_SLABS*L1_BYTES more memory use. Lets say NR_SLABS is 100 and a cache line size of 128 then we have just increased SLAB metadata requirements by 12.8k per cpu. (Another later patch reduces these requirements) Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 16:26:05 +08:00
struct kmem_cache_cpu *c;
Lockless (and preemptless) fastpaths for slub Use the this_cpu_cmpxchg_double functionality to implement a lockless allocation algorithm on arches that support fast this_cpu_ops. Each of the per cpu pointers is paired with a transaction id that ensures that updates of the per cpu information can only occur in sequence on a certain cpu. A transaction id is a "long" integer that is comprised of an event number and the cpu number. The event number is incremented for every change to the per cpu state. This means that the cmpxchg instruction can verify for an update that nothing interfered and that we are updating the percpu structure for the processor where we picked up the information and that we are also currently on that processor when we update the information. This results in a significant decrease of the overhead in the fastpaths. It also makes it easy to adopt the fast path for realtime kernels since this is lockless and does not require the use of the current per cpu area over the critical section. It is only important that the per cpu area is current at the beginning of the critical section and at the end. So there is no need even to disable preemption. Test results show that the fastpath cycle count is reduced by up to ~ 40% (alloc/free test goes from ~140 cycles down to ~80). The slowpath for kfree adds a few cycles. Sadly this does nothing for the slowpath which is where the main issues with performance in slub are but the best case performance rises significantly. (For that see the more complex slub patches that require cmpxchg_double) Kmalloc: alloc/free test Before: 10000 times kmalloc(8)/kfree -> 134 cycles 10000 times kmalloc(16)/kfree -> 152 cycles 10000 times kmalloc(32)/kfree -> 144 cycles 10000 times kmalloc(64)/kfree -> 142 cycles 10000 times kmalloc(128)/kfree -> 142 cycles 10000 times kmalloc(256)/kfree -> 132 cycles 10000 times kmalloc(512)/kfree -> 132 cycles 10000 times kmalloc(1024)/kfree -> 135 cycles 10000 times kmalloc(2048)/kfree -> 135 cycles 10000 times kmalloc(4096)/kfree -> 135 cycles 10000 times kmalloc(8192)/kfree -> 144 cycles 10000 times kmalloc(16384)/kfree -> 754 cycles After: 10000 times kmalloc(8)/kfree -> 78 cycles 10000 times kmalloc(16)/kfree -> 78 cycles 10000 times kmalloc(32)/kfree -> 82 cycles 10000 times kmalloc(64)/kfree -> 88 cycles 10000 times kmalloc(128)/kfree -> 79 cycles 10000 times kmalloc(256)/kfree -> 79 cycles 10000 times kmalloc(512)/kfree -> 85 cycles 10000 times kmalloc(1024)/kfree -> 82 cycles 10000 times kmalloc(2048)/kfree -> 82 cycles 10000 times kmalloc(4096)/kfree -> 85 cycles 10000 times kmalloc(8192)/kfree -> 82 cycles 10000 times kmalloc(16384)/kfree -> 706 cycles Kmalloc: Repeatedly allocate then free test Before: 10000 times kmalloc(8) -> 211 cycles kfree -> 113 cycles 10000 times kmalloc(16) -> 174 cycles kfree -> 115 cycles 10000 times kmalloc(32) -> 235 cycles kfree -> 129 cycles 10000 times kmalloc(64) -> 222 cycles kfree -> 120 cycles 10000 times kmalloc(128) -> 343 cycles kfree -> 139 cycles 10000 times kmalloc(256) -> 827 cycles kfree -> 147 cycles 10000 times kmalloc(512) -> 1048 cycles kfree -> 272 cycles 10000 times kmalloc(1024) -> 2043 cycles kfree -> 528 cycles 10000 times kmalloc(2048) -> 4002 cycles kfree -> 571 cycles 10000 times kmalloc(4096) -> 7740 cycles kfree -> 628 cycles 10000 times kmalloc(8192) -> 8062 cycles kfree -> 850 cycles 10000 times kmalloc(16384) -> 8895 cycles kfree -> 1249 cycles After: 10000 times kmalloc(8) -> 190 cycles kfree -> 129 cycles 10000 times kmalloc(16) -> 76 cycles kfree -> 123 cycles 10000 times kmalloc(32) -> 126 cycles kfree -> 124 cycles 10000 times kmalloc(64) -> 181 cycles kfree -> 128 cycles 10000 times kmalloc(128) -> 310 cycles kfree -> 140 cycles 10000 times kmalloc(256) -> 809 cycles kfree -> 165 cycles 10000 times kmalloc(512) -> 1005 cycles kfree -> 269 cycles 10000 times kmalloc(1024) -> 1999 cycles kfree -> 527 cycles 10000 times kmalloc(2048) -> 3967 cycles kfree -> 570 cycles 10000 times kmalloc(4096) -> 7658 cycles kfree -> 637 cycles 10000 times kmalloc(8192) -> 8111 cycles kfree -> 859 cycles 10000 times kmalloc(16384) -> 8791 cycles kfree -> 1173 cycles Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-02-26 01:38:54 +08:00
unsigned long tid;
memcg_slab_free_hook(s, &head, 1);
Lockless (and preemptless) fastpaths for slub Use the this_cpu_cmpxchg_double functionality to implement a lockless allocation algorithm on arches that support fast this_cpu_ops. Each of the per cpu pointers is paired with a transaction id that ensures that updates of the per cpu information can only occur in sequence on a certain cpu. A transaction id is a "long" integer that is comprised of an event number and the cpu number. The event number is incremented for every change to the per cpu state. This means that the cmpxchg instruction can verify for an update that nothing interfered and that we are updating the percpu structure for the processor where we picked up the information and that we are also currently on that processor when we update the information. This results in a significant decrease of the overhead in the fastpaths. It also makes it easy to adopt the fast path for realtime kernels since this is lockless and does not require the use of the current per cpu area over the critical section. It is only important that the per cpu area is current at the beginning of the critical section and at the end. So there is no need even to disable preemption. Test results show that the fastpath cycle count is reduced by up to ~ 40% (alloc/free test goes from ~140 cycles down to ~80). The slowpath for kfree adds a few cycles. Sadly this does nothing for the slowpath which is where the main issues with performance in slub are but the best case performance rises significantly. (For that see the more complex slub patches that require cmpxchg_double) Kmalloc: alloc/free test Before: 10000 times kmalloc(8)/kfree -> 134 cycles 10000 times kmalloc(16)/kfree -> 152 cycles 10000 times kmalloc(32)/kfree -> 144 cycles 10000 times kmalloc(64)/kfree -> 142 cycles 10000 times kmalloc(128)/kfree -> 142 cycles 10000 times kmalloc(256)/kfree -> 132 cycles 10000 times kmalloc(512)/kfree -> 132 cycles 10000 times kmalloc(1024)/kfree -> 135 cycles 10000 times kmalloc(2048)/kfree -> 135 cycles 10000 times kmalloc(4096)/kfree -> 135 cycles 10000 times kmalloc(8192)/kfree -> 144 cycles 10000 times kmalloc(16384)/kfree -> 754 cycles After: 10000 times kmalloc(8)/kfree -> 78 cycles 10000 times kmalloc(16)/kfree -> 78 cycles 10000 times kmalloc(32)/kfree -> 82 cycles 10000 times kmalloc(64)/kfree -> 88 cycles 10000 times kmalloc(128)/kfree -> 79 cycles 10000 times kmalloc(256)/kfree -> 79 cycles 10000 times kmalloc(512)/kfree -> 85 cycles 10000 times kmalloc(1024)/kfree -> 82 cycles 10000 times kmalloc(2048)/kfree -> 82 cycles 10000 times kmalloc(4096)/kfree -> 85 cycles 10000 times kmalloc(8192)/kfree -> 82 cycles 10000 times kmalloc(16384)/kfree -> 706 cycles Kmalloc: Repeatedly allocate then free test Before: 10000 times kmalloc(8) -> 211 cycles kfree -> 113 cycles 10000 times kmalloc(16) -> 174 cycles kfree -> 115 cycles 10000 times kmalloc(32) -> 235 cycles kfree -> 129 cycles 10000 times kmalloc(64) -> 222 cycles kfree -> 120 cycles 10000 times kmalloc(128) -> 343 cycles kfree -> 139 cycles 10000 times kmalloc(256) -> 827 cycles kfree -> 147 cycles 10000 times kmalloc(512) -> 1048 cycles kfree -> 272 cycles 10000 times kmalloc(1024) -> 2043 cycles kfree -> 528 cycles 10000 times kmalloc(2048) -> 4002 cycles kfree -> 571 cycles 10000 times kmalloc(4096) -> 7740 cycles kfree -> 628 cycles 10000 times kmalloc(8192) -> 8062 cycles kfree -> 850 cycles 10000 times kmalloc(16384) -> 8895 cycles kfree -> 1249 cycles After: 10000 times kmalloc(8) -> 190 cycles kfree -> 129 cycles 10000 times kmalloc(16) -> 76 cycles kfree -> 123 cycles 10000 times kmalloc(32) -> 126 cycles kfree -> 124 cycles 10000 times kmalloc(64) -> 181 cycles kfree -> 128 cycles 10000 times kmalloc(128) -> 310 cycles kfree -> 140 cycles 10000 times kmalloc(256) -> 809 cycles kfree -> 165 cycles 10000 times kmalloc(512) -> 1005 cycles kfree -> 269 cycles 10000 times kmalloc(1024) -> 1999 cycles kfree -> 527 cycles 10000 times kmalloc(2048) -> 3967 cycles kfree -> 570 cycles 10000 times kmalloc(4096) -> 7658 cycles kfree -> 637 cycles 10000 times kmalloc(8192) -> 8111 cycles kfree -> 859 cycles 10000 times kmalloc(16384) -> 8791 cycles kfree -> 1173 cycles Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-02-26 01:38:54 +08:00
redo:
/*
* Determine the currently cpus per cpu slab.
* The cpu may change afterward. However that does not matter since
* data is retrieved via this pointer. If we are on the same cpu
slub: fix spelling succedd to succeed With this patchset the SLUB allocator now has both bulk alloc and free implemented. This patchset mostly optimizes the "fastpath" where objects are available on the per CPU fastpath page. This mostly amortize the less-heavy none-locked cmpxchg_double used on fastpath. The "fallback" bulking (e.g __kmem_cache_free_bulk) provides a good basis for comparison. Measurements[1] of the fallback functions __kmem_cache_{free,alloc}_bulk have been copied from slab_common.c and forced "noinline" to force a function call like slab_common.c. Measurements on CPU CPU i7-4790K @ 4.00GHz Baseline normal fastpath (alloc+free cost): 42 cycles(tsc) 10.601 ns Measurements last-patch with disabled debugging: Bulk- fallback - this-patch 1 - 57 cycles(tsc) 14.448 ns - 44 cycles(tsc) 11.236 ns improved 22.8% 2 - 51 cycles(tsc) 12.768 ns - 28 cycles(tsc) 7.019 ns improved 45.1% 3 - 48 cycles(tsc) 12.232 ns - 22 cycles(tsc) 5.526 ns improved 54.2% 4 - 48 cycles(tsc) 12.025 ns - 19 cycles(tsc) 4.786 ns improved 60.4% 8 - 46 cycles(tsc) 11.558 ns - 18 cycles(tsc) 4.572 ns improved 60.9% 16 - 45 cycles(tsc) 11.458 ns - 18 cycles(tsc) 4.658 ns improved 60.0% 30 - 45 cycles(tsc) 11.499 ns - 18 cycles(tsc) 4.568 ns improved 60.0% 32 - 79 cycles(tsc) 19.917 ns - 65 cycles(tsc) 16.454 ns improved 17.7% 34 - 78 cycles(tsc) 19.655 ns - 63 cycles(tsc) 15.932 ns improved 19.2% 48 - 68 cycles(tsc) 17.049 ns - 50 cycles(tsc) 12.506 ns improved 26.5% 64 - 80 cycles(tsc) 20.009 ns - 63 cycles(tsc) 15.929 ns improved 21.3% 128 - 94 cycles(tsc) 23.749 ns - 86 cycles(tsc) 21.583 ns improved 8.5% 158 - 97 cycles(tsc) 24.299 ns - 90 cycles(tsc) 22.552 ns improved 7.2% 250 - 102 cycles(tsc) 25.681 ns - 98 cycles(tsc) 24.589 ns improved 3.9% Benchmarking shows impressive improvements in the "fastpath" with a small number of objects in the working set. Once the working set increases, resulting in activating the "slowpath" (that contains the heavier locked cmpxchg_double) the improvement decreases. I'm currently working on also optimizing the "slowpath" (as network stack use-case hits this), but this patchset should provide a good foundation for further improvements. Rest of my patch queue in this area needs some more work, but preliminary results are good. I'm attending Netfilter Workshop[2] next week, and I'll hopefully return working on further improvements in this area. This patch (of 6): s/succedd/succeed/ Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-05 06:45:31 +08:00
* during the cmpxchg then the free will succeed.
Lockless (and preemptless) fastpaths for slub Use the this_cpu_cmpxchg_double functionality to implement a lockless allocation algorithm on arches that support fast this_cpu_ops. Each of the per cpu pointers is paired with a transaction id that ensures that updates of the per cpu information can only occur in sequence on a certain cpu. A transaction id is a "long" integer that is comprised of an event number and the cpu number. The event number is incremented for every change to the per cpu state. This means that the cmpxchg instruction can verify for an update that nothing interfered and that we are updating the percpu structure for the processor where we picked up the information and that we are also currently on that processor when we update the information. This results in a significant decrease of the overhead in the fastpaths. It also makes it easy to adopt the fast path for realtime kernels since this is lockless and does not require the use of the current per cpu area over the critical section. It is only important that the per cpu area is current at the beginning of the critical section and at the end. So there is no need even to disable preemption. Test results show that the fastpath cycle count is reduced by up to ~ 40% (alloc/free test goes from ~140 cycles down to ~80). The slowpath for kfree adds a few cycles. Sadly this does nothing for the slowpath which is where the main issues with performance in slub are but the best case performance rises significantly. (For that see the more complex slub patches that require cmpxchg_double) Kmalloc: alloc/free test Before: 10000 times kmalloc(8)/kfree -> 134 cycles 10000 times kmalloc(16)/kfree -> 152 cycles 10000 times kmalloc(32)/kfree -> 144 cycles 10000 times kmalloc(64)/kfree -> 142 cycles 10000 times kmalloc(128)/kfree -> 142 cycles 10000 times kmalloc(256)/kfree -> 132 cycles 10000 times kmalloc(512)/kfree -> 132 cycles 10000 times kmalloc(1024)/kfree -> 135 cycles 10000 times kmalloc(2048)/kfree -> 135 cycles 10000 times kmalloc(4096)/kfree -> 135 cycles 10000 times kmalloc(8192)/kfree -> 144 cycles 10000 times kmalloc(16384)/kfree -> 754 cycles After: 10000 times kmalloc(8)/kfree -> 78 cycles 10000 times kmalloc(16)/kfree -> 78 cycles 10000 times kmalloc(32)/kfree -> 82 cycles 10000 times kmalloc(64)/kfree -> 88 cycles 10000 times kmalloc(128)/kfree -> 79 cycles 10000 times kmalloc(256)/kfree -> 79 cycles 10000 times kmalloc(512)/kfree -> 85 cycles 10000 times kmalloc(1024)/kfree -> 82 cycles 10000 times kmalloc(2048)/kfree -> 82 cycles 10000 times kmalloc(4096)/kfree -> 85 cycles 10000 times kmalloc(8192)/kfree -> 82 cycles 10000 times kmalloc(16384)/kfree -> 706 cycles Kmalloc: Repeatedly allocate then free test Before: 10000 times kmalloc(8) -> 211 cycles kfree -> 113 cycles 10000 times kmalloc(16) -> 174 cycles kfree -> 115 cycles 10000 times kmalloc(32) -> 235 cycles kfree -> 129 cycles 10000 times kmalloc(64) -> 222 cycles kfree -> 120 cycles 10000 times kmalloc(128) -> 343 cycles kfree -> 139 cycles 10000 times kmalloc(256) -> 827 cycles kfree -> 147 cycles 10000 times kmalloc(512) -> 1048 cycles kfree -> 272 cycles 10000 times kmalloc(1024) -> 2043 cycles kfree -> 528 cycles 10000 times kmalloc(2048) -> 4002 cycles kfree -> 571 cycles 10000 times kmalloc(4096) -> 7740 cycles kfree -> 628 cycles 10000 times kmalloc(8192) -> 8062 cycles kfree -> 850 cycles 10000 times kmalloc(16384) -> 8895 cycles kfree -> 1249 cycles After: 10000 times kmalloc(8) -> 190 cycles kfree -> 129 cycles 10000 times kmalloc(16) -> 76 cycles kfree -> 123 cycles 10000 times kmalloc(32) -> 126 cycles kfree -> 124 cycles 10000 times kmalloc(64) -> 181 cycles kfree -> 128 cycles 10000 times kmalloc(128) -> 310 cycles kfree -> 140 cycles 10000 times kmalloc(256) -> 809 cycles kfree -> 165 cycles 10000 times kmalloc(512) -> 1005 cycles kfree -> 269 cycles 10000 times kmalloc(1024) -> 1999 cycles kfree -> 527 cycles 10000 times kmalloc(2048) -> 3967 cycles kfree -> 570 cycles 10000 times kmalloc(4096) -> 7658 cycles kfree -> 637 cycles 10000 times kmalloc(8192) -> 8111 cycles kfree -> 859 cycles 10000 times kmalloc(16384) -> 8791 cycles kfree -> 1173 cycles Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-02-26 01:38:54 +08:00
*/
mm/slub: optimize alloc/free fastpath by removing preemption on/off We had to insert a preempt enable/disable in the fastpath a while ago in order to guarantee that tid and kmem_cache_cpu are retrieved on the same cpu. It is the problem only for CONFIG_PREEMPT in which scheduler can move the process to other cpu during retrieving data. Now, I reach the solution to remove preempt enable/disable in the fastpath. If tid is matched with kmem_cache_cpu's tid after tid and kmem_cache_cpu are retrieved by separate this_cpu operation, it means that they are retrieved on the same cpu. If not matched, we just have to retry it. With this guarantee, preemption enable/disable isn't need at all even if CONFIG_PREEMPT, so this patch removes it. I saw roughly 5% win in a fast-path loop over kmem_cache_alloc/free in CONFIG_PREEMPT. (14.821 ns -> 14.049 ns) Below is the result of Christoph's slab_test reported by Jesper Dangaard Brouer. * Before Single thread testing ===================== 1. Kmalloc: Repeatedly allocate then free test 10000 times kmalloc(8) -> 49 cycles kfree -> 62 cycles 10000 times kmalloc(16) -> 48 cycles kfree -> 64 cycles 10000 times kmalloc(32) -> 53 cycles kfree -> 70 cycles 10000 times kmalloc(64) -> 64 cycles kfree -> 77 cycles 10000 times kmalloc(128) -> 74 cycles kfree -> 84 cycles 10000 times kmalloc(256) -> 84 cycles kfree -> 114 cycles 10000 times kmalloc(512) -> 83 cycles kfree -> 116 cycles 10000 times kmalloc(1024) -> 81 cycles kfree -> 120 cycles 10000 times kmalloc(2048) -> 104 cycles kfree -> 136 cycles 10000 times kmalloc(4096) -> 142 cycles kfree -> 165 cycles 10000 times kmalloc(8192) -> 238 cycles kfree -> 226 cycles 10000 times kmalloc(16384) -> 403 cycles kfree -> 264 cycles 2. Kmalloc: alloc/free test 10000 times kmalloc(8)/kfree -> 68 cycles 10000 times kmalloc(16)/kfree -> 68 cycles 10000 times kmalloc(32)/kfree -> 69 cycles 10000 times kmalloc(64)/kfree -> 68 cycles 10000 times kmalloc(128)/kfree -> 68 cycles 10000 times kmalloc(256)/kfree -> 68 cycles 10000 times kmalloc(512)/kfree -> 74 cycles 10000 times kmalloc(1024)/kfree -> 75 cycles 10000 times kmalloc(2048)/kfree -> 74 cycles 10000 times kmalloc(4096)/kfree -> 74 cycles 10000 times kmalloc(8192)/kfree -> 75 cycles 10000 times kmalloc(16384)/kfree -> 510 cycles * After Single thread testing ===================== 1. Kmalloc: Repeatedly allocate then free test 10000 times kmalloc(8) -> 46 cycles kfree -> 61 cycles 10000 times kmalloc(16) -> 46 cycles kfree -> 63 cycles 10000 times kmalloc(32) -> 49 cycles kfree -> 69 cycles 10000 times kmalloc(64) -> 57 cycles kfree -> 76 cycles 10000 times kmalloc(128) -> 66 cycles kfree -> 83 cycles 10000 times kmalloc(256) -> 84 cycles kfree -> 110 cycles 10000 times kmalloc(512) -> 77 cycles kfree -> 114 cycles 10000 times kmalloc(1024) -> 80 cycles kfree -> 116 cycles 10000 times kmalloc(2048) -> 102 cycles kfree -> 131 cycles 10000 times kmalloc(4096) -> 135 cycles kfree -> 163 cycles 10000 times kmalloc(8192) -> 238 cycles kfree -> 218 cycles 10000 times kmalloc(16384) -> 399 cycles kfree -> 262 cycles 2. Kmalloc: alloc/free test 10000 times kmalloc(8)/kfree -> 65 cycles 10000 times kmalloc(16)/kfree -> 66 cycles 10000 times kmalloc(32)/kfree -> 65 cycles 10000 times kmalloc(64)/kfree -> 66 cycles 10000 times kmalloc(128)/kfree -> 66 cycles 10000 times kmalloc(256)/kfree -> 71 cycles 10000 times kmalloc(512)/kfree -> 72 cycles 10000 times kmalloc(1024)/kfree -> 71 cycles 10000 times kmalloc(2048)/kfree -> 71 cycles 10000 times kmalloc(4096)/kfree -> 71 cycles 10000 times kmalloc(8192)/kfree -> 65 cycles 10000 times kmalloc(16384)/kfree -> 511 cycles Most of the results are better than before. Note that this change slightly worses performance in !CONFIG_PREEMPT, roughly 0.3%. Implementing each case separately would help performance, but, since it's so marginal, I didn't do that. This would help maintanance since we have same code for all cases. Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com> Acked-by: Christoph Lameter <cl@linux.com> Tested-by: Jesper Dangaard Brouer <brouer@redhat.com> Acked-by: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 06:09:32 +08:00
do {
tid = this_cpu_read(s->cpu_slab->tid);
c = raw_cpu_ptr(s->cpu_slab);
} while (IS_ENABLED(CONFIG_PREEMPTION) &&
mm/slub: fix lockups on PREEMPT && !SMP kernels Commit 9aabf810a67c ("mm/slub: optimize alloc/free fastpath by removing preemption on/off") introduced an occasional hang for kernels built with CONFIG_PREEMPT && !CONFIG_SMP. The problem is the following loop the patch introduced to slab_alloc_node and slab_free: do { tid = this_cpu_read(s->cpu_slab->tid); c = raw_cpu_ptr(s->cpu_slab); } while (IS_ENABLED(CONFIG_PREEMPT) && unlikely(tid != c->tid)); GCC 4.9 has been observed to hoist the load of c and c->tid above the loop for !SMP kernels (as in this case raw_cpu_ptr(x) is compile-time constant and does not force a reload). On arm64 the generated assembly looks like: ldr x4, [x0,#8] loop: ldr x1, [x0,#8] cmp x1, x4 b.ne loop If the thread is preempted between the load of c->tid (into x1) and tid (into x4), and an allocation or free occurs in another thread (bumping the cpu_slab's tid), the thread will be stuck in the loop until s->cpu_slab->tid wraps, which may be forever in the absence of allocations/frees on the same CPU. This patch changes the loop condition to access c->tid with READ_ONCE. This ensures that the value is reloaded even when the compiler would otherwise assume it could cache the value, and also ensures that the load will not be torn. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: David Rientjes <rientjes@google.com> Cc: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: Steve Capper <steve.capper@linaro.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-03-26 06:55:23 +08:00
unlikely(tid != READ_ONCE(c->tid)));
mm/slub: optimize alloc/free fastpath by removing preemption on/off We had to insert a preempt enable/disable in the fastpath a while ago in order to guarantee that tid and kmem_cache_cpu are retrieved on the same cpu. It is the problem only for CONFIG_PREEMPT in which scheduler can move the process to other cpu during retrieving data. Now, I reach the solution to remove preempt enable/disable in the fastpath. If tid is matched with kmem_cache_cpu's tid after tid and kmem_cache_cpu are retrieved by separate this_cpu operation, it means that they are retrieved on the same cpu. If not matched, we just have to retry it. With this guarantee, preemption enable/disable isn't need at all even if CONFIG_PREEMPT, so this patch removes it. I saw roughly 5% win in a fast-path loop over kmem_cache_alloc/free in CONFIG_PREEMPT. (14.821 ns -> 14.049 ns) Below is the result of Christoph's slab_test reported by Jesper Dangaard Brouer. * Before Single thread testing ===================== 1. Kmalloc: Repeatedly allocate then free test 10000 times kmalloc(8) -> 49 cycles kfree -> 62 cycles 10000 times kmalloc(16) -> 48 cycles kfree -> 64 cycles 10000 times kmalloc(32) -> 53 cycles kfree -> 70 cycles 10000 times kmalloc(64) -> 64 cycles kfree -> 77 cycles 10000 times kmalloc(128) -> 74 cycles kfree -> 84 cycles 10000 times kmalloc(256) -> 84 cycles kfree -> 114 cycles 10000 times kmalloc(512) -> 83 cycles kfree -> 116 cycles 10000 times kmalloc(1024) -> 81 cycles kfree -> 120 cycles 10000 times kmalloc(2048) -> 104 cycles kfree -> 136 cycles 10000 times kmalloc(4096) -> 142 cycles kfree -> 165 cycles 10000 times kmalloc(8192) -> 238 cycles kfree -> 226 cycles 10000 times kmalloc(16384) -> 403 cycles kfree -> 264 cycles 2. Kmalloc: alloc/free test 10000 times kmalloc(8)/kfree -> 68 cycles 10000 times kmalloc(16)/kfree -> 68 cycles 10000 times kmalloc(32)/kfree -> 69 cycles 10000 times kmalloc(64)/kfree -> 68 cycles 10000 times kmalloc(128)/kfree -> 68 cycles 10000 times kmalloc(256)/kfree -> 68 cycles 10000 times kmalloc(512)/kfree -> 74 cycles 10000 times kmalloc(1024)/kfree -> 75 cycles 10000 times kmalloc(2048)/kfree -> 74 cycles 10000 times kmalloc(4096)/kfree -> 74 cycles 10000 times kmalloc(8192)/kfree -> 75 cycles 10000 times kmalloc(16384)/kfree -> 510 cycles * After Single thread testing ===================== 1. Kmalloc: Repeatedly allocate then free test 10000 times kmalloc(8) -> 46 cycles kfree -> 61 cycles 10000 times kmalloc(16) -> 46 cycles kfree -> 63 cycles 10000 times kmalloc(32) -> 49 cycles kfree -> 69 cycles 10000 times kmalloc(64) -> 57 cycles kfree -> 76 cycles 10000 times kmalloc(128) -> 66 cycles kfree -> 83 cycles 10000 times kmalloc(256) -> 84 cycles kfree -> 110 cycles 10000 times kmalloc(512) -> 77 cycles kfree -> 114 cycles 10000 times kmalloc(1024) -> 80 cycles kfree -> 116 cycles 10000 times kmalloc(2048) -> 102 cycles kfree -> 131 cycles 10000 times kmalloc(4096) -> 135 cycles kfree -> 163 cycles 10000 times kmalloc(8192) -> 238 cycles kfree -> 218 cycles 10000 times kmalloc(16384) -> 399 cycles kfree -> 262 cycles 2. Kmalloc: alloc/free test 10000 times kmalloc(8)/kfree -> 65 cycles 10000 times kmalloc(16)/kfree -> 66 cycles 10000 times kmalloc(32)/kfree -> 65 cycles 10000 times kmalloc(64)/kfree -> 66 cycles 10000 times kmalloc(128)/kfree -> 66 cycles 10000 times kmalloc(256)/kfree -> 71 cycles 10000 times kmalloc(512)/kfree -> 72 cycles 10000 times kmalloc(1024)/kfree -> 71 cycles 10000 times kmalloc(2048)/kfree -> 71 cycles 10000 times kmalloc(4096)/kfree -> 71 cycles 10000 times kmalloc(8192)/kfree -> 65 cycles 10000 times kmalloc(16384)/kfree -> 511 cycles Most of the results are better than before. Note that this change slightly worses performance in !CONFIG_PREEMPT, roughly 0.3%. Implementing each case separately would help performance, but, since it's so marginal, I didn't do that. This would help maintanance since we have same code for all cases. Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com> Acked-by: Christoph Lameter <cl@linux.com> Tested-by: Jesper Dangaard Brouer <brouer@redhat.com> Acked-by: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-11 06:09:32 +08:00
/* Same with comment on barrier() in slab_alloc_node() */
barrier();
if (likely(page == c->page)) {
void **freelist = READ_ONCE(c->freelist);
set_freepointer(s, tail_obj, freelist);
Lockless (and preemptless) fastpaths for slub Use the this_cpu_cmpxchg_double functionality to implement a lockless allocation algorithm on arches that support fast this_cpu_ops. Each of the per cpu pointers is paired with a transaction id that ensures that updates of the per cpu information can only occur in sequence on a certain cpu. A transaction id is a "long" integer that is comprised of an event number and the cpu number. The event number is incremented for every change to the per cpu state. This means that the cmpxchg instruction can verify for an update that nothing interfered and that we are updating the percpu structure for the processor where we picked up the information and that we are also currently on that processor when we update the information. This results in a significant decrease of the overhead in the fastpaths. It also makes it easy to adopt the fast path for realtime kernels since this is lockless and does not require the use of the current per cpu area over the critical section. It is only important that the per cpu area is current at the beginning of the critical section and at the end. So there is no need even to disable preemption. Test results show that the fastpath cycle count is reduced by up to ~ 40% (alloc/free test goes from ~140 cycles down to ~80). The slowpath for kfree adds a few cycles. Sadly this does nothing for the slowpath which is where the main issues with performance in slub are but the best case performance rises significantly. (For that see the more complex slub patches that require cmpxchg_double) Kmalloc: alloc/free test Before: 10000 times kmalloc(8)/kfree -> 134 cycles 10000 times kmalloc(16)/kfree -> 152 cycles 10000 times kmalloc(32)/kfree -> 144 cycles 10000 times kmalloc(64)/kfree -> 142 cycles 10000 times kmalloc(128)/kfree -> 142 cycles 10000 times kmalloc(256)/kfree -> 132 cycles 10000 times kmalloc(512)/kfree -> 132 cycles 10000 times kmalloc(1024)/kfree -> 135 cycles 10000 times kmalloc(2048)/kfree -> 135 cycles 10000 times kmalloc(4096)/kfree -> 135 cycles 10000 times kmalloc(8192)/kfree -> 144 cycles 10000 times kmalloc(16384)/kfree -> 754 cycles After: 10000 times kmalloc(8)/kfree -> 78 cycles 10000 times kmalloc(16)/kfree -> 78 cycles 10000 times kmalloc(32)/kfree -> 82 cycles 10000 times kmalloc(64)/kfree -> 88 cycles 10000 times kmalloc(128)/kfree -> 79 cycles 10000 times kmalloc(256)/kfree -> 79 cycles 10000 times kmalloc(512)/kfree -> 85 cycles 10000 times kmalloc(1024)/kfree -> 82 cycles 10000 times kmalloc(2048)/kfree -> 82 cycles 10000 times kmalloc(4096)/kfree -> 85 cycles 10000 times kmalloc(8192)/kfree -> 82 cycles 10000 times kmalloc(16384)/kfree -> 706 cycles Kmalloc: Repeatedly allocate then free test Before: 10000 times kmalloc(8) -> 211 cycles kfree -> 113 cycles 10000 times kmalloc(16) -> 174 cycles kfree -> 115 cycles 10000 times kmalloc(32) -> 235 cycles kfree -> 129 cycles 10000 times kmalloc(64) -> 222 cycles kfree -> 120 cycles 10000 times kmalloc(128) -> 343 cycles kfree -> 139 cycles 10000 times kmalloc(256) -> 827 cycles kfree -> 147 cycles 10000 times kmalloc(512) -> 1048 cycles kfree -> 272 cycles 10000 times kmalloc(1024) -> 2043 cycles kfree -> 528 cycles 10000 times kmalloc(2048) -> 4002 cycles kfree -> 571 cycles 10000 times kmalloc(4096) -> 7740 cycles kfree -> 628 cycles 10000 times kmalloc(8192) -> 8062 cycles kfree -> 850 cycles 10000 times kmalloc(16384) -> 8895 cycles kfree -> 1249 cycles After: 10000 times kmalloc(8) -> 190 cycles kfree -> 129 cycles 10000 times kmalloc(16) -> 76 cycles kfree -> 123 cycles 10000 times kmalloc(32) -> 126 cycles kfree -> 124 cycles 10000 times kmalloc(64) -> 181 cycles kfree -> 128 cycles 10000 times kmalloc(128) -> 310 cycles kfree -> 140 cycles 10000 times kmalloc(256) -> 809 cycles kfree -> 165 cycles 10000 times kmalloc(512) -> 1005 cycles kfree -> 269 cycles 10000 times kmalloc(1024) -> 1999 cycles kfree -> 527 cycles 10000 times kmalloc(2048) -> 3967 cycles kfree -> 570 cycles 10000 times kmalloc(4096) -> 7658 cycles kfree -> 637 cycles 10000 times kmalloc(8192) -> 8111 cycles kfree -> 859 cycles 10000 times kmalloc(16384) -> 8791 cycles kfree -> 1173 cycles Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-02-26 01:38:54 +08:00
if (unlikely(!this_cpu_cmpxchg_double(
Lockless (and preemptless) fastpaths for slub Use the this_cpu_cmpxchg_double functionality to implement a lockless allocation algorithm on arches that support fast this_cpu_ops. Each of the per cpu pointers is paired with a transaction id that ensures that updates of the per cpu information can only occur in sequence on a certain cpu. A transaction id is a "long" integer that is comprised of an event number and the cpu number. The event number is incremented for every change to the per cpu state. This means that the cmpxchg instruction can verify for an update that nothing interfered and that we are updating the percpu structure for the processor where we picked up the information and that we are also currently on that processor when we update the information. This results in a significant decrease of the overhead in the fastpaths. It also makes it easy to adopt the fast path for realtime kernels since this is lockless and does not require the use of the current per cpu area over the critical section. It is only important that the per cpu area is current at the beginning of the critical section and at the end. So there is no need even to disable preemption. Test results show that the fastpath cycle count is reduced by up to ~ 40% (alloc/free test goes from ~140 cycles down to ~80). The slowpath for kfree adds a few cycles. Sadly this does nothing for the slowpath which is where the main issues with performance in slub are but the best case performance rises significantly. (For that see the more complex slub patches that require cmpxchg_double) Kmalloc: alloc/free test Before: 10000 times kmalloc(8)/kfree -> 134 cycles 10000 times kmalloc(16)/kfree -> 152 cycles 10000 times kmalloc(32)/kfree -> 144 cycles 10000 times kmalloc(64)/kfree -> 142 cycles 10000 times kmalloc(128)/kfree -> 142 cycles 10000 times kmalloc(256)/kfree -> 132 cycles 10000 times kmalloc(512)/kfree -> 132 cycles 10000 times kmalloc(1024)/kfree -> 135 cycles 10000 times kmalloc(2048)/kfree -> 135 cycles 10000 times kmalloc(4096)/kfree -> 135 cycles 10000 times kmalloc(8192)/kfree -> 144 cycles 10000 times kmalloc(16384)/kfree -> 754 cycles After: 10000 times kmalloc(8)/kfree -> 78 cycles 10000 times kmalloc(16)/kfree -> 78 cycles 10000 times kmalloc(32)/kfree -> 82 cycles 10000 times kmalloc(64)/kfree -> 88 cycles 10000 times kmalloc(128)/kfree -> 79 cycles 10000 times kmalloc(256)/kfree -> 79 cycles 10000 times kmalloc(512)/kfree -> 85 cycles 10000 times kmalloc(1024)/kfree -> 82 cycles 10000 times kmalloc(2048)/kfree -> 82 cycles 10000 times kmalloc(4096)/kfree -> 85 cycles 10000 times kmalloc(8192)/kfree -> 82 cycles 10000 times kmalloc(16384)/kfree -> 706 cycles Kmalloc: Repeatedly allocate then free test Before: 10000 times kmalloc(8) -> 211 cycles kfree -> 113 cycles 10000 times kmalloc(16) -> 174 cycles kfree -> 115 cycles 10000 times kmalloc(32) -> 235 cycles kfree -> 129 cycles 10000 times kmalloc(64) -> 222 cycles kfree -> 120 cycles 10000 times kmalloc(128) -> 343 cycles kfree -> 139 cycles 10000 times kmalloc(256) -> 827 cycles kfree -> 147 cycles 10000 times kmalloc(512) -> 1048 cycles kfree -> 272 cycles 10000 times kmalloc(1024) -> 2043 cycles kfree -> 528 cycles 10000 times kmalloc(2048) -> 4002 cycles kfree -> 571 cycles 10000 times kmalloc(4096) -> 7740 cycles kfree -> 628 cycles 10000 times kmalloc(8192) -> 8062 cycles kfree -> 850 cycles 10000 times kmalloc(16384) -> 8895 cycles kfree -> 1249 cycles After: 10000 times kmalloc(8) -> 190 cycles kfree -> 129 cycles 10000 times kmalloc(16) -> 76 cycles kfree -> 123 cycles 10000 times kmalloc(32) -> 126 cycles kfree -> 124 cycles 10000 times kmalloc(64) -> 181 cycles kfree -> 128 cycles 10000 times kmalloc(128) -> 310 cycles kfree -> 140 cycles 10000 times kmalloc(256) -> 809 cycles kfree -> 165 cycles 10000 times kmalloc(512) -> 1005 cycles kfree -> 269 cycles 10000 times kmalloc(1024) -> 1999 cycles kfree -> 527 cycles 10000 times kmalloc(2048) -> 3967 cycles kfree -> 570 cycles 10000 times kmalloc(4096) -> 7658 cycles kfree -> 637 cycles 10000 times kmalloc(8192) -> 8111 cycles kfree -> 859 cycles 10000 times kmalloc(16384) -> 8791 cycles kfree -> 1173 cycles Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-02-26 01:38:54 +08:00
s->cpu_slab->freelist, s->cpu_slab->tid,
freelist, tid,
head, next_tid(tid)))) {
Lockless (and preemptless) fastpaths for slub Use the this_cpu_cmpxchg_double functionality to implement a lockless allocation algorithm on arches that support fast this_cpu_ops. Each of the per cpu pointers is paired with a transaction id that ensures that updates of the per cpu information can only occur in sequence on a certain cpu. A transaction id is a "long" integer that is comprised of an event number and the cpu number. The event number is incremented for every change to the per cpu state. This means that the cmpxchg instruction can verify for an update that nothing interfered and that we are updating the percpu structure for the processor where we picked up the information and that we are also currently on that processor when we update the information. This results in a significant decrease of the overhead in the fastpaths. It also makes it easy to adopt the fast path for realtime kernels since this is lockless and does not require the use of the current per cpu area over the critical section. It is only important that the per cpu area is current at the beginning of the critical section and at the end. So there is no need even to disable preemption. Test results show that the fastpath cycle count is reduced by up to ~ 40% (alloc/free test goes from ~140 cycles down to ~80). The slowpath for kfree adds a few cycles. Sadly this does nothing for the slowpath which is where the main issues with performance in slub are but the best case performance rises significantly. (For that see the more complex slub patches that require cmpxchg_double) Kmalloc: alloc/free test Before: 10000 times kmalloc(8)/kfree -> 134 cycles 10000 times kmalloc(16)/kfree -> 152 cycles 10000 times kmalloc(32)/kfree -> 144 cycles 10000 times kmalloc(64)/kfree -> 142 cycles 10000 times kmalloc(128)/kfree -> 142 cycles 10000 times kmalloc(256)/kfree -> 132 cycles 10000 times kmalloc(512)/kfree -> 132 cycles 10000 times kmalloc(1024)/kfree -> 135 cycles 10000 times kmalloc(2048)/kfree -> 135 cycles 10000 times kmalloc(4096)/kfree -> 135 cycles 10000 times kmalloc(8192)/kfree -> 144 cycles 10000 times kmalloc(16384)/kfree -> 754 cycles After: 10000 times kmalloc(8)/kfree -> 78 cycles 10000 times kmalloc(16)/kfree -> 78 cycles 10000 times kmalloc(32)/kfree -> 82 cycles 10000 times kmalloc(64)/kfree -> 88 cycles 10000 times kmalloc(128)/kfree -> 79 cycles 10000 times kmalloc(256)/kfree -> 79 cycles 10000 times kmalloc(512)/kfree -> 85 cycles 10000 times kmalloc(1024)/kfree -> 82 cycles 10000 times kmalloc(2048)/kfree -> 82 cycles 10000 times kmalloc(4096)/kfree -> 85 cycles 10000 times kmalloc(8192)/kfree -> 82 cycles 10000 times kmalloc(16384)/kfree -> 706 cycles Kmalloc: Repeatedly allocate then free test Before: 10000 times kmalloc(8) -> 211 cycles kfree -> 113 cycles 10000 times kmalloc(16) -> 174 cycles kfree -> 115 cycles 10000 times kmalloc(32) -> 235 cycles kfree -> 129 cycles 10000 times kmalloc(64) -> 222 cycles kfree -> 120 cycles 10000 times kmalloc(128) -> 343 cycles kfree -> 139 cycles 10000 times kmalloc(256) -> 827 cycles kfree -> 147 cycles 10000 times kmalloc(512) -> 1048 cycles kfree -> 272 cycles 10000 times kmalloc(1024) -> 2043 cycles kfree -> 528 cycles 10000 times kmalloc(2048) -> 4002 cycles kfree -> 571 cycles 10000 times kmalloc(4096) -> 7740 cycles kfree -> 628 cycles 10000 times kmalloc(8192) -> 8062 cycles kfree -> 850 cycles 10000 times kmalloc(16384) -> 8895 cycles kfree -> 1249 cycles After: 10000 times kmalloc(8) -> 190 cycles kfree -> 129 cycles 10000 times kmalloc(16) -> 76 cycles kfree -> 123 cycles 10000 times kmalloc(32) -> 126 cycles kfree -> 124 cycles 10000 times kmalloc(64) -> 181 cycles kfree -> 128 cycles 10000 times kmalloc(128) -> 310 cycles kfree -> 140 cycles 10000 times kmalloc(256) -> 809 cycles kfree -> 165 cycles 10000 times kmalloc(512) -> 1005 cycles kfree -> 269 cycles 10000 times kmalloc(1024) -> 1999 cycles kfree -> 527 cycles 10000 times kmalloc(2048) -> 3967 cycles kfree -> 570 cycles 10000 times kmalloc(4096) -> 7658 cycles kfree -> 637 cycles 10000 times kmalloc(8192) -> 8111 cycles kfree -> 859 cycles 10000 times kmalloc(16384) -> 8791 cycles kfree -> 1173 cycles Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-02-26 01:38:54 +08:00
note_cmpxchg_failure("slab_free", s, tid);
goto redo;
}
stat(s, FREE_FASTPATH);
slub: support concurrent local and remote frees and allocs on a slab Avoid atomic overhead in slab_alloc and slab_free SLUB needs to use the slab_lock for the per cpu slabs to synchronize with potential kfree operations. This patch avoids that need by moving all free objects onto a lockless_freelist. The regular freelist continues to exist and will be used to free objects. So while we consume the lockless_freelist the regular freelist may build up objects. If we are out of objects on the lockless_freelist then we may check the regular freelist. If it has objects then we move those over to the lockless_freelist and do this again. There is a significant savings in terms of atomic operations that have to be performed. We can even free directly to the lockless_freelist if we know that we are running on the same processor. So this speeds up short lived objects. They may be allocated and freed without taking the slab_lock. This is particular good for netperf. In order to maximize the effect of the new faster hotpath we extract the hottest performance pieces into inlined functions. These are then inlined into kmem_cache_alloc and kmem_cache_free. So hotpath allocation and freeing no longer requires a subroutine call within SLUB. [I am not sure that it is worth doing this because it changes the easy to read structure of slub just to reduce atomic ops. However, there is someone out there with a benchmark on 4 way and 8 way processor systems that seems to show a 5% regression vs. Slab. Seems that the regression is due to increased atomic operations use vs. SLAB in SLUB). I wonder if this is applicable or discernable at all in a real workload?] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-10 18:15:16 +08:00
} else
__slab_free(s, page, head, tail_obj, cnt, addr);
slub: support concurrent local and remote frees and allocs on a slab Avoid atomic overhead in slab_alloc and slab_free SLUB needs to use the slab_lock for the per cpu slabs to synchronize with potential kfree operations. This patch avoids that need by moving all free objects onto a lockless_freelist. The regular freelist continues to exist and will be used to free objects. So while we consume the lockless_freelist the regular freelist may build up objects. If we are out of objects on the lockless_freelist then we may check the regular freelist. If it has objects then we move those over to the lockless_freelist and do this again. There is a significant savings in terms of atomic operations that have to be performed. We can even free directly to the lockless_freelist if we know that we are running on the same processor. So this speeds up short lived objects. They may be allocated and freed without taking the slab_lock. This is particular good for netperf. In order to maximize the effect of the new faster hotpath we extract the hottest performance pieces into inlined functions. These are then inlined into kmem_cache_alloc and kmem_cache_free. So hotpath allocation and freeing no longer requires a subroutine call within SLUB. [I am not sure that it is worth doing this because it changes the easy to read structure of slub just to reduce atomic ops. However, there is someone out there with a benchmark on 4 way and 8 way processor systems that seems to show a 5% regression vs. Slab. Seems that the regression is due to increased atomic operations use vs. SLAB in SLUB). I wonder if this is applicable or discernable at all in a real workload?] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-10 18:15:16 +08:00
}
static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
void *head, void *tail, int cnt,
unsigned long addr)
{
/*
kasan, slub: fix handling of kasan_slab_free hook The kasan_slab_free hook's return value denotes whether the reuse of a slab object must be delayed (e.g. when the object is put into memory qurantine). The current way SLUB handles this hook is by ignoring its return value and hardcoding checks similar (but not exactly the same) to the ones performed in kasan_slab_free, which is prone to making mistakes. The main difference between the hardcoded checks and the ones in kasan_slab_free is whether we want to perform a free in case when an invalid-free or a double-free was detected (we don't). This patch changes the way SLUB handles this by: 1. taking into account the return value of kasan_slab_free for each of the objects, that are being freed; 2. reconstructing the freelist of objects to exclude the ones, whose reuse must be delayed. [andreyknvl@google.com: eliminate unnecessary branch in slab_free] Link: http://lkml.kernel.org/r/a62759a2545fddf69b0c034547212ca1eb1b3ce2.1520359686.git.andreyknvl@google.com Link: http://lkml.kernel.org/r/083f58501e54731203801d899632d76175868e97.1519400992.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Acked-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Alexander Potapenko <glider@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Kostya Serebryany <kcc@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-04-11 07:30:31 +08:00
* With KASAN enabled slab_free_freelist_hook modifies the freelist
* to remove objects, whose reuse must be delayed.
*/
kasan, slub: fix handling of kasan_slab_free hook The kasan_slab_free hook's return value denotes whether the reuse of a slab object must be delayed (e.g. when the object is put into memory qurantine). The current way SLUB handles this hook is by ignoring its return value and hardcoding checks similar (but not exactly the same) to the ones performed in kasan_slab_free, which is prone to making mistakes. The main difference between the hardcoded checks and the ones in kasan_slab_free is whether we want to perform a free in case when an invalid-free or a double-free was detected (we don't). This patch changes the way SLUB handles this by: 1. taking into account the return value of kasan_slab_free for each of the objects, that are being freed; 2. reconstructing the freelist of objects to exclude the ones, whose reuse must be delayed. [andreyknvl@google.com: eliminate unnecessary branch in slab_free] Link: http://lkml.kernel.org/r/a62759a2545fddf69b0c034547212ca1eb1b3ce2.1520359686.git.andreyknvl@google.com Link: http://lkml.kernel.org/r/083f58501e54731203801d899632d76175868e97.1519400992.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Acked-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Alexander Potapenko <glider@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Kostya Serebryany <kcc@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-04-11 07:30:31 +08:00
if (slab_free_freelist_hook(s, &head, &tail))
do_slab_free(s, page, head, tail, cnt, addr);
}
kasan: add CONFIG_KASAN_GENERIC and CONFIG_KASAN_SW_TAGS This commit splits the current CONFIG_KASAN config option into two: 1. CONFIG_KASAN_GENERIC, that enables the generic KASAN mode (the one that exists now); 2. CONFIG_KASAN_SW_TAGS, that enables the software tag-based KASAN mode. The name CONFIG_KASAN_SW_TAGS is chosen as in the future we will have another hardware tag-based KASAN mode, that will rely on hardware memory tagging support in arm64. With CONFIG_KASAN_SW_TAGS enabled, compiler options are changed to instrument kernel files with -fsantize=kernel-hwaddress (except the ones for which KASAN_SANITIZE := n is set). Both CONFIG_KASAN_GENERIC and CONFIG_KASAN_SW_TAGS support both CONFIG_KASAN_INLINE and CONFIG_KASAN_OUTLINE instrumentation modes. This commit also adds empty placeholder (for now) implementation of tag-based KASAN specific hooks inserted by the compiler and adjusts common hooks implementation. While this commit adds the CONFIG_KASAN_SW_TAGS config option, this option is not selectable, as it depends on HAVE_ARCH_KASAN_SW_TAGS, which we will enable once all the infrastracture code has been added. Link: http://lkml.kernel.org/r/b2550106eb8a68b10fefbabce820910b115aa853.1544099024.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Reviewed-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Reviewed-by: Dmitry Vyukov <dvyukov@google.com> Cc: Christoph Lameter <cl@linux.com> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 16:29:53 +08:00
#ifdef CONFIG_KASAN_GENERIC
void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
{
do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr);
}
#endif
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
void kmem_cache_free(struct kmem_cache *s, void *x)
{
s = cache_from_obj(s, x);
if (!s)
return;
slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
trace_kmem_cache_free(_RET_IP_, x, s->name);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
EXPORT_SYMBOL(kmem_cache_free);
slub: optimize bulk slowpath free by detached freelist This change focus on improving the speed of object freeing in the "slowpath" of kmem_cache_free_bulk. The calls slab_free (fastpath) and __slab_free (slowpath) have been extended with support for bulk free, which amortize the overhead of the (locked) cmpxchg_double. To use the new bulking feature, we build what I call a detached freelist. The detached freelist takes advantage of three properties: 1) the free function call owns the object that is about to be freed, thus writing into this memory is synchronization-free. 2) many freelist's can co-exist side-by-side in the same slab-page each with a separate head pointer. 3) it is the visibility of the head pointer that needs synchronization. Given these properties, the brilliant part is that the detached freelist can be constructed without any need for synchronization. The freelist is constructed directly in the page objects, without any synchronization needed. The detached freelist is allocated on the stack of the function call kmem_cache_free_bulk. Thus, the freelist head pointer is not visible to other CPUs. All objects in a SLUB freelist must belong to the same slab-page. Thus, constructing the detached freelist is about matching objects that belong to the same slab-page. The bulk free array is scanned is a progressive manor with a limited look-ahead facility. Kmem debug support is handled in call of slab_free(). Notice kmem_cache_free_bulk no longer need to disable IRQs. This only slowed down single free bulk with approx 3 cycles. Performance data: Benchmarked[1] obj size 256 bytes on CPU i7-4790K @ 4.00GHz SLUB fastpath single object quick reuse: 47 cycles(tsc) 11.931 ns To get stable and comparable numbers, the kernel have been booted with "slab_merge" (this also improve performance for larger bulk sizes). Performance data, compared against fallback bulking: bulk - fallback bulk - improvement with this patch 1 - 62 cycles(tsc) 15.662 ns - 49 cycles(tsc) 12.407 ns- improved 21.0% 2 - 55 cycles(tsc) 13.935 ns - 30 cycles(tsc) 7.506 ns - improved 45.5% 3 - 53 cycles(tsc) 13.341 ns - 23 cycles(tsc) 5.865 ns - improved 56.6% 4 - 52 cycles(tsc) 13.081 ns - 20 cycles(tsc) 5.048 ns - improved 61.5% 8 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.659 ns - improved 64.0% 16 - 49 cycles(tsc) 12.412 ns - 17 cycles(tsc) 4.495 ns - improved 65.3% 30 - 49 cycles(tsc) 12.484 ns - 18 cycles(tsc) 4.533 ns - improved 63.3% 32 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.707 ns - improved 64.0% 34 - 96 cycles(tsc) 24.243 ns - 23 cycles(tsc) 5.976 ns - improved 76.0% 48 - 83 cycles(tsc) 20.818 ns - 21 cycles(tsc) 5.329 ns - improved 74.7% 64 - 74 cycles(tsc) 18.700 ns - 20 cycles(tsc) 5.127 ns - improved 73.0% 128 - 90 cycles(tsc) 22.734 ns - 27 cycles(tsc) 6.833 ns - improved 70.0% 158 - 99 cycles(tsc) 24.776 ns - 30 cycles(tsc) 7.583 ns - improved 69.7% 250 - 104 cycles(tsc) 26.089 ns - 37 cycles(tsc) 9.280 ns - improved 64.4% Performance data, compared current in-kernel bulking: bulk - curr in-kernel - improvement with this patch 1 - 46 cycles(tsc) - 49 cycles(tsc) - improved (cycles:-3) -6.5% 2 - 27 cycles(tsc) - 30 cycles(tsc) - improved (cycles:-3) -11.1% 3 - 21 cycles(tsc) - 23 cycles(tsc) - improved (cycles:-2) -9.5% 4 - 18 cycles(tsc) - 20 cycles(tsc) - improved (cycles:-2) -11.1% 8 - 17 cycles(tsc) - 18 cycles(tsc) - improved (cycles:-1) -5.9% 16 - 18 cycles(tsc) - 17 cycles(tsc) - improved (cycles: 1) 5.6% 30 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 32 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 34 - 78 cycles(tsc) - 23 cycles(tsc) - improved (cycles:55) 70.5% 48 - 60 cycles(tsc) - 21 cycles(tsc) - improved (cycles:39) 65.0% 64 - 49 cycles(tsc) - 20 cycles(tsc) - improved (cycles:29) 59.2% 128 - 69 cycles(tsc) - 27 cycles(tsc) - improved (cycles:42) 60.9% 158 - 79 cycles(tsc) - 30 cycles(tsc) - improved (cycles:49) 62.0% 250 - 86 cycles(tsc) - 37 cycles(tsc) - improved (cycles:49) 57.0% Performance with normal SLUB merging is significantly slower for larger bulking. This is believed to (primarily) be an effect of not having to share the per-CPU data-structures, as tuning per-CPU size can achieve similar performance. bulk - slab_nomerge - normal SLUB merge 1 - 49 cycles(tsc) - 49 cycles(tsc) - merge slower with cycles:0 2 - 30 cycles(tsc) - 30 cycles(tsc) - merge slower with cycles:0 3 - 23 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:0 4 - 20 cycles(tsc) - 20 cycles(tsc) - merge slower with cycles:0 8 - 18 cycles(tsc) - 18 cycles(tsc) - merge slower with cycles:0 16 - 17 cycles(tsc) - 17 cycles(tsc) - merge slower with cycles:0 30 - 18 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:5 32 - 18 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:4 34 - 23 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:-1 48 - 21 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:1 64 - 20 cycles(tsc) - 48 cycles(tsc) - merge slower with cycles:28 128 - 27 cycles(tsc) - 57 cycles(tsc) - merge slower with cycles:30 158 - 30 cycles(tsc) - 59 cycles(tsc) - merge slower with cycles:29 250 - 37 cycles(tsc) - 56 cycles(tsc) - merge slower with cycles:19 Joint work with Alexander Duyck. [1] https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/mm/slab_bulk_test01.c [akpm@linux-foundation.org: BUG_ON -> WARN_ON;return] Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Alexander Duyck <alexander.h.duyck@redhat.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-21 07:57:49 +08:00
struct detached_freelist {
slub: initial bulk free implementation This implements SLUB specific kmem_cache_free_bulk(). SLUB allocator now both have bulk alloc and free implemented. Choose to reenable local IRQs while calling slowpath __slab_free(). In worst case, where all objects hit slowpath call, the performance should still be faster than fallback function __kmem_cache_free_bulk(), because local_irq_{disable+enable} is very fast (7-cycles), while the fallback invokes this_cpu_cmpxchg() which is slightly slower (9-cycles). Nitpicking, this should be faster for N>=4, due to the entry cost of local_irq_{disable+enable}. Do notice that the save+restore variant is very expensive, this is key to why this optimization works. CPU: i7-4790K CPU @ 4.00GHz * local_irq_{disable,enable}: 7 cycles(tsc) - 1.821 ns * local_irq_{save,restore} : 37 cycles(tsc) - 9.443 ns Measurements on CPU CPU i7-4790K @ 4.00GHz Baseline normal fastpath (alloc+free cost): 43 cycles(tsc) 10.834 ns Bulk- fallback - this-patch 1 - 58 cycles(tsc) 14.542 ns - 43 cycles(tsc) 10.811 ns improved 25.9% 2 - 50 cycles(tsc) 12.659 ns - 27 cycles(tsc) 6.867 ns improved 46.0% 3 - 48 cycles(tsc) 12.168 ns - 21 cycles(tsc) 5.496 ns improved 56.2% 4 - 47 cycles(tsc) 11.987 ns - 24 cycles(tsc) 6.038 ns improved 48.9% 8 - 46 cycles(tsc) 11.518 ns - 17 cycles(tsc) 4.280 ns improved 63.0% 16 - 45 cycles(tsc) 11.366 ns - 17 cycles(tsc) 4.483 ns improved 62.2% 30 - 45 cycles(tsc) 11.433 ns - 18 cycles(tsc) 4.531 ns improved 60.0% 32 - 75 cycles(tsc) 18.983 ns - 58 cycles(tsc) 14.586 ns improved 22.7% 34 - 71 cycles(tsc) 17.940 ns - 53 cycles(tsc) 13.391 ns improved 25.4% 48 - 80 cycles(tsc) 20.077 ns - 65 cycles(tsc) 16.268 ns improved 18.8% 64 - 71 cycles(tsc) 17.799 ns - 53 cycles(tsc) 13.440 ns improved 25.4% 128 - 91 cycles(tsc) 22.980 ns - 79 cycles(tsc) 19.899 ns improved 13.2% 158 - 100 cycles(tsc) 25.241 ns - 90 cycles(tsc) 22.732 ns improved 10.0% 250 - 102 cycles(tsc) 25.583 ns - 95 cycles(tsc) 23.916 ns improved 6.9% Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-05 06:45:43 +08:00
struct page *page;
slub: optimize bulk slowpath free by detached freelist This change focus on improving the speed of object freeing in the "slowpath" of kmem_cache_free_bulk. The calls slab_free (fastpath) and __slab_free (slowpath) have been extended with support for bulk free, which amortize the overhead of the (locked) cmpxchg_double. To use the new bulking feature, we build what I call a detached freelist. The detached freelist takes advantage of three properties: 1) the free function call owns the object that is about to be freed, thus writing into this memory is synchronization-free. 2) many freelist's can co-exist side-by-side in the same slab-page each with a separate head pointer. 3) it is the visibility of the head pointer that needs synchronization. Given these properties, the brilliant part is that the detached freelist can be constructed without any need for synchronization. The freelist is constructed directly in the page objects, without any synchronization needed. The detached freelist is allocated on the stack of the function call kmem_cache_free_bulk. Thus, the freelist head pointer is not visible to other CPUs. All objects in a SLUB freelist must belong to the same slab-page. Thus, constructing the detached freelist is about matching objects that belong to the same slab-page. The bulk free array is scanned is a progressive manor with a limited look-ahead facility. Kmem debug support is handled in call of slab_free(). Notice kmem_cache_free_bulk no longer need to disable IRQs. This only slowed down single free bulk with approx 3 cycles. Performance data: Benchmarked[1] obj size 256 bytes on CPU i7-4790K @ 4.00GHz SLUB fastpath single object quick reuse: 47 cycles(tsc) 11.931 ns To get stable and comparable numbers, the kernel have been booted with "slab_merge" (this also improve performance for larger bulk sizes). Performance data, compared against fallback bulking: bulk - fallback bulk - improvement with this patch 1 - 62 cycles(tsc) 15.662 ns - 49 cycles(tsc) 12.407 ns- improved 21.0% 2 - 55 cycles(tsc) 13.935 ns - 30 cycles(tsc) 7.506 ns - improved 45.5% 3 - 53 cycles(tsc) 13.341 ns - 23 cycles(tsc) 5.865 ns - improved 56.6% 4 - 52 cycles(tsc) 13.081 ns - 20 cycles(tsc) 5.048 ns - improved 61.5% 8 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.659 ns - improved 64.0% 16 - 49 cycles(tsc) 12.412 ns - 17 cycles(tsc) 4.495 ns - improved 65.3% 30 - 49 cycles(tsc) 12.484 ns - 18 cycles(tsc) 4.533 ns - improved 63.3% 32 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.707 ns - improved 64.0% 34 - 96 cycles(tsc) 24.243 ns - 23 cycles(tsc) 5.976 ns - improved 76.0% 48 - 83 cycles(tsc) 20.818 ns - 21 cycles(tsc) 5.329 ns - improved 74.7% 64 - 74 cycles(tsc) 18.700 ns - 20 cycles(tsc) 5.127 ns - improved 73.0% 128 - 90 cycles(tsc) 22.734 ns - 27 cycles(tsc) 6.833 ns - improved 70.0% 158 - 99 cycles(tsc) 24.776 ns - 30 cycles(tsc) 7.583 ns - improved 69.7% 250 - 104 cycles(tsc) 26.089 ns - 37 cycles(tsc) 9.280 ns - improved 64.4% Performance data, compared current in-kernel bulking: bulk - curr in-kernel - improvement with this patch 1 - 46 cycles(tsc) - 49 cycles(tsc) - improved (cycles:-3) -6.5% 2 - 27 cycles(tsc) - 30 cycles(tsc) - improved (cycles:-3) -11.1% 3 - 21 cycles(tsc) - 23 cycles(tsc) - improved (cycles:-2) -9.5% 4 - 18 cycles(tsc) - 20 cycles(tsc) - improved (cycles:-2) -11.1% 8 - 17 cycles(tsc) - 18 cycles(tsc) - improved (cycles:-1) -5.9% 16 - 18 cycles(tsc) - 17 cycles(tsc) - improved (cycles: 1) 5.6% 30 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 32 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 34 - 78 cycles(tsc) - 23 cycles(tsc) - improved (cycles:55) 70.5% 48 - 60 cycles(tsc) - 21 cycles(tsc) - improved (cycles:39) 65.0% 64 - 49 cycles(tsc) - 20 cycles(tsc) - improved (cycles:29) 59.2% 128 - 69 cycles(tsc) - 27 cycles(tsc) - improved (cycles:42) 60.9% 158 - 79 cycles(tsc) - 30 cycles(tsc) - improved (cycles:49) 62.0% 250 - 86 cycles(tsc) - 37 cycles(tsc) - improved (cycles:49) 57.0% Performance with normal SLUB merging is significantly slower for larger bulking. This is believed to (primarily) be an effect of not having to share the per-CPU data-structures, as tuning per-CPU size can achieve similar performance. bulk - slab_nomerge - normal SLUB merge 1 - 49 cycles(tsc) - 49 cycles(tsc) - merge slower with cycles:0 2 - 30 cycles(tsc) - 30 cycles(tsc) - merge slower with cycles:0 3 - 23 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:0 4 - 20 cycles(tsc) - 20 cycles(tsc) - merge slower with cycles:0 8 - 18 cycles(tsc) - 18 cycles(tsc) - merge slower with cycles:0 16 - 17 cycles(tsc) - 17 cycles(tsc) - merge slower with cycles:0 30 - 18 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:5 32 - 18 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:4 34 - 23 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:-1 48 - 21 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:1 64 - 20 cycles(tsc) - 48 cycles(tsc) - merge slower with cycles:28 128 - 27 cycles(tsc) - 57 cycles(tsc) - merge slower with cycles:30 158 - 30 cycles(tsc) - 59 cycles(tsc) - merge slower with cycles:29 250 - 37 cycles(tsc) - 56 cycles(tsc) - merge slower with cycles:19 Joint work with Alexander Duyck. [1] https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/mm/slab_bulk_test01.c [akpm@linux-foundation.org: BUG_ON -> WARN_ON;return] Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Alexander Duyck <alexander.h.duyck@redhat.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-21 07:57:49 +08:00
void *tail;
void *freelist;
int cnt;
struct kmem_cache *s;
slub: optimize bulk slowpath free by detached freelist This change focus on improving the speed of object freeing in the "slowpath" of kmem_cache_free_bulk. The calls slab_free (fastpath) and __slab_free (slowpath) have been extended with support for bulk free, which amortize the overhead of the (locked) cmpxchg_double. To use the new bulking feature, we build what I call a detached freelist. The detached freelist takes advantage of three properties: 1) the free function call owns the object that is about to be freed, thus writing into this memory is synchronization-free. 2) many freelist's can co-exist side-by-side in the same slab-page each with a separate head pointer. 3) it is the visibility of the head pointer that needs synchronization. Given these properties, the brilliant part is that the detached freelist can be constructed without any need for synchronization. The freelist is constructed directly in the page objects, without any synchronization needed. The detached freelist is allocated on the stack of the function call kmem_cache_free_bulk. Thus, the freelist head pointer is not visible to other CPUs. All objects in a SLUB freelist must belong to the same slab-page. Thus, constructing the detached freelist is about matching objects that belong to the same slab-page. The bulk free array is scanned is a progressive manor with a limited look-ahead facility. Kmem debug support is handled in call of slab_free(). Notice kmem_cache_free_bulk no longer need to disable IRQs. This only slowed down single free bulk with approx 3 cycles. Performance data: Benchmarked[1] obj size 256 bytes on CPU i7-4790K @ 4.00GHz SLUB fastpath single object quick reuse: 47 cycles(tsc) 11.931 ns To get stable and comparable numbers, the kernel have been booted with "slab_merge" (this also improve performance for larger bulk sizes). Performance data, compared against fallback bulking: bulk - fallback bulk - improvement with this patch 1 - 62 cycles(tsc) 15.662 ns - 49 cycles(tsc) 12.407 ns- improved 21.0% 2 - 55 cycles(tsc) 13.935 ns - 30 cycles(tsc) 7.506 ns - improved 45.5% 3 - 53 cycles(tsc) 13.341 ns - 23 cycles(tsc) 5.865 ns - improved 56.6% 4 - 52 cycles(tsc) 13.081 ns - 20 cycles(tsc) 5.048 ns - improved 61.5% 8 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.659 ns - improved 64.0% 16 - 49 cycles(tsc) 12.412 ns - 17 cycles(tsc) 4.495 ns - improved 65.3% 30 - 49 cycles(tsc) 12.484 ns - 18 cycles(tsc) 4.533 ns - improved 63.3% 32 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.707 ns - improved 64.0% 34 - 96 cycles(tsc) 24.243 ns - 23 cycles(tsc) 5.976 ns - improved 76.0% 48 - 83 cycles(tsc) 20.818 ns - 21 cycles(tsc) 5.329 ns - improved 74.7% 64 - 74 cycles(tsc) 18.700 ns - 20 cycles(tsc) 5.127 ns - improved 73.0% 128 - 90 cycles(tsc) 22.734 ns - 27 cycles(tsc) 6.833 ns - improved 70.0% 158 - 99 cycles(tsc) 24.776 ns - 30 cycles(tsc) 7.583 ns - improved 69.7% 250 - 104 cycles(tsc) 26.089 ns - 37 cycles(tsc) 9.280 ns - improved 64.4% Performance data, compared current in-kernel bulking: bulk - curr in-kernel - improvement with this patch 1 - 46 cycles(tsc) - 49 cycles(tsc) - improved (cycles:-3) -6.5% 2 - 27 cycles(tsc) - 30 cycles(tsc) - improved (cycles:-3) -11.1% 3 - 21 cycles(tsc) - 23 cycles(tsc) - improved (cycles:-2) -9.5% 4 - 18 cycles(tsc) - 20 cycles(tsc) - improved (cycles:-2) -11.1% 8 - 17 cycles(tsc) - 18 cycles(tsc) - improved (cycles:-1) -5.9% 16 - 18 cycles(tsc) - 17 cycles(tsc) - improved (cycles: 1) 5.6% 30 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 32 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 34 - 78 cycles(tsc) - 23 cycles(tsc) - improved (cycles:55) 70.5% 48 - 60 cycles(tsc) - 21 cycles(tsc) - improved (cycles:39) 65.0% 64 - 49 cycles(tsc) - 20 cycles(tsc) - improved (cycles:29) 59.2% 128 - 69 cycles(tsc) - 27 cycles(tsc) - improved (cycles:42) 60.9% 158 - 79 cycles(tsc) - 30 cycles(tsc) - improved (cycles:49) 62.0% 250 - 86 cycles(tsc) - 37 cycles(tsc) - improved (cycles:49) 57.0% Performance with normal SLUB merging is significantly slower for larger bulking. This is believed to (primarily) be an effect of not having to share the per-CPU data-structures, as tuning per-CPU size can achieve similar performance. bulk - slab_nomerge - normal SLUB merge 1 - 49 cycles(tsc) - 49 cycles(tsc) - merge slower with cycles:0 2 - 30 cycles(tsc) - 30 cycles(tsc) - merge slower with cycles:0 3 - 23 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:0 4 - 20 cycles(tsc) - 20 cycles(tsc) - merge slower with cycles:0 8 - 18 cycles(tsc) - 18 cycles(tsc) - merge slower with cycles:0 16 - 17 cycles(tsc) - 17 cycles(tsc) - merge slower with cycles:0 30 - 18 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:5 32 - 18 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:4 34 - 23 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:-1 48 - 21 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:1 64 - 20 cycles(tsc) - 48 cycles(tsc) - merge slower with cycles:28 128 - 27 cycles(tsc) - 57 cycles(tsc) - merge slower with cycles:30 158 - 30 cycles(tsc) - 59 cycles(tsc) - merge slower with cycles:29 250 - 37 cycles(tsc) - 56 cycles(tsc) - merge slower with cycles:19 Joint work with Alexander Duyck. [1] https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/mm/slab_bulk_test01.c [akpm@linux-foundation.org: BUG_ON -> WARN_ON;return] Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Alexander Duyck <alexander.h.duyck@redhat.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-21 07:57:49 +08:00
};
slub: initial bulk free implementation This implements SLUB specific kmem_cache_free_bulk(). SLUB allocator now both have bulk alloc and free implemented. Choose to reenable local IRQs while calling slowpath __slab_free(). In worst case, where all objects hit slowpath call, the performance should still be faster than fallback function __kmem_cache_free_bulk(), because local_irq_{disable+enable} is very fast (7-cycles), while the fallback invokes this_cpu_cmpxchg() which is slightly slower (9-cycles). Nitpicking, this should be faster for N>=4, due to the entry cost of local_irq_{disable+enable}. Do notice that the save+restore variant is very expensive, this is key to why this optimization works. CPU: i7-4790K CPU @ 4.00GHz * local_irq_{disable,enable}: 7 cycles(tsc) - 1.821 ns * local_irq_{save,restore} : 37 cycles(tsc) - 9.443 ns Measurements on CPU CPU i7-4790K @ 4.00GHz Baseline normal fastpath (alloc+free cost): 43 cycles(tsc) 10.834 ns Bulk- fallback - this-patch 1 - 58 cycles(tsc) 14.542 ns - 43 cycles(tsc) 10.811 ns improved 25.9% 2 - 50 cycles(tsc) 12.659 ns - 27 cycles(tsc) 6.867 ns improved 46.0% 3 - 48 cycles(tsc) 12.168 ns - 21 cycles(tsc) 5.496 ns improved 56.2% 4 - 47 cycles(tsc) 11.987 ns - 24 cycles(tsc) 6.038 ns improved 48.9% 8 - 46 cycles(tsc) 11.518 ns - 17 cycles(tsc) 4.280 ns improved 63.0% 16 - 45 cycles(tsc) 11.366 ns - 17 cycles(tsc) 4.483 ns improved 62.2% 30 - 45 cycles(tsc) 11.433 ns - 18 cycles(tsc) 4.531 ns improved 60.0% 32 - 75 cycles(tsc) 18.983 ns - 58 cycles(tsc) 14.586 ns improved 22.7% 34 - 71 cycles(tsc) 17.940 ns - 53 cycles(tsc) 13.391 ns improved 25.4% 48 - 80 cycles(tsc) 20.077 ns - 65 cycles(tsc) 16.268 ns improved 18.8% 64 - 71 cycles(tsc) 17.799 ns - 53 cycles(tsc) 13.440 ns improved 25.4% 128 - 91 cycles(tsc) 22.980 ns - 79 cycles(tsc) 19.899 ns improved 13.2% 158 - 100 cycles(tsc) 25.241 ns - 90 cycles(tsc) 22.732 ns improved 10.0% 250 - 102 cycles(tsc) 25.583 ns - 95 cycles(tsc) 23.916 ns improved 6.9% Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-05 06:45:43 +08:00
slub: optimize bulk slowpath free by detached freelist This change focus on improving the speed of object freeing in the "slowpath" of kmem_cache_free_bulk. The calls slab_free (fastpath) and __slab_free (slowpath) have been extended with support for bulk free, which amortize the overhead of the (locked) cmpxchg_double. To use the new bulking feature, we build what I call a detached freelist. The detached freelist takes advantage of three properties: 1) the free function call owns the object that is about to be freed, thus writing into this memory is synchronization-free. 2) many freelist's can co-exist side-by-side in the same slab-page each with a separate head pointer. 3) it is the visibility of the head pointer that needs synchronization. Given these properties, the brilliant part is that the detached freelist can be constructed without any need for synchronization. The freelist is constructed directly in the page objects, without any synchronization needed. The detached freelist is allocated on the stack of the function call kmem_cache_free_bulk. Thus, the freelist head pointer is not visible to other CPUs. All objects in a SLUB freelist must belong to the same slab-page. Thus, constructing the detached freelist is about matching objects that belong to the same slab-page. The bulk free array is scanned is a progressive manor with a limited look-ahead facility. Kmem debug support is handled in call of slab_free(). Notice kmem_cache_free_bulk no longer need to disable IRQs. This only slowed down single free bulk with approx 3 cycles. Performance data: Benchmarked[1] obj size 256 bytes on CPU i7-4790K @ 4.00GHz SLUB fastpath single object quick reuse: 47 cycles(tsc) 11.931 ns To get stable and comparable numbers, the kernel have been booted with "slab_merge" (this also improve performance for larger bulk sizes). Performance data, compared against fallback bulking: bulk - fallback bulk - improvement with this patch 1 - 62 cycles(tsc) 15.662 ns - 49 cycles(tsc) 12.407 ns- improved 21.0% 2 - 55 cycles(tsc) 13.935 ns - 30 cycles(tsc) 7.506 ns - improved 45.5% 3 - 53 cycles(tsc) 13.341 ns - 23 cycles(tsc) 5.865 ns - improved 56.6% 4 - 52 cycles(tsc) 13.081 ns - 20 cycles(tsc) 5.048 ns - improved 61.5% 8 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.659 ns - improved 64.0% 16 - 49 cycles(tsc) 12.412 ns - 17 cycles(tsc) 4.495 ns - improved 65.3% 30 - 49 cycles(tsc) 12.484 ns - 18 cycles(tsc) 4.533 ns - improved 63.3% 32 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.707 ns - improved 64.0% 34 - 96 cycles(tsc) 24.243 ns - 23 cycles(tsc) 5.976 ns - improved 76.0% 48 - 83 cycles(tsc) 20.818 ns - 21 cycles(tsc) 5.329 ns - improved 74.7% 64 - 74 cycles(tsc) 18.700 ns - 20 cycles(tsc) 5.127 ns - improved 73.0% 128 - 90 cycles(tsc) 22.734 ns - 27 cycles(tsc) 6.833 ns - improved 70.0% 158 - 99 cycles(tsc) 24.776 ns - 30 cycles(tsc) 7.583 ns - improved 69.7% 250 - 104 cycles(tsc) 26.089 ns - 37 cycles(tsc) 9.280 ns - improved 64.4% Performance data, compared current in-kernel bulking: bulk - curr in-kernel - improvement with this patch 1 - 46 cycles(tsc) - 49 cycles(tsc) - improved (cycles:-3) -6.5% 2 - 27 cycles(tsc) - 30 cycles(tsc) - improved (cycles:-3) -11.1% 3 - 21 cycles(tsc) - 23 cycles(tsc) - improved (cycles:-2) -9.5% 4 - 18 cycles(tsc) - 20 cycles(tsc) - improved (cycles:-2) -11.1% 8 - 17 cycles(tsc) - 18 cycles(tsc) - improved (cycles:-1) -5.9% 16 - 18 cycles(tsc) - 17 cycles(tsc) - improved (cycles: 1) 5.6% 30 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 32 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 34 - 78 cycles(tsc) - 23 cycles(tsc) - improved (cycles:55) 70.5% 48 - 60 cycles(tsc) - 21 cycles(tsc) - improved (cycles:39) 65.0% 64 - 49 cycles(tsc) - 20 cycles(tsc) - improved (cycles:29) 59.2% 128 - 69 cycles(tsc) - 27 cycles(tsc) - improved (cycles:42) 60.9% 158 - 79 cycles(tsc) - 30 cycles(tsc) - improved (cycles:49) 62.0% 250 - 86 cycles(tsc) - 37 cycles(tsc) - improved (cycles:49) 57.0% Performance with normal SLUB merging is significantly slower for larger bulking. This is believed to (primarily) be an effect of not having to share the per-CPU data-structures, as tuning per-CPU size can achieve similar performance. bulk - slab_nomerge - normal SLUB merge 1 - 49 cycles(tsc) - 49 cycles(tsc) - merge slower with cycles:0 2 - 30 cycles(tsc) - 30 cycles(tsc) - merge slower with cycles:0 3 - 23 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:0 4 - 20 cycles(tsc) - 20 cycles(tsc) - merge slower with cycles:0 8 - 18 cycles(tsc) - 18 cycles(tsc) - merge slower with cycles:0 16 - 17 cycles(tsc) - 17 cycles(tsc) - merge slower with cycles:0 30 - 18 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:5 32 - 18 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:4 34 - 23 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:-1 48 - 21 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:1 64 - 20 cycles(tsc) - 48 cycles(tsc) - merge slower with cycles:28 128 - 27 cycles(tsc) - 57 cycles(tsc) - merge slower with cycles:30 158 - 30 cycles(tsc) - 59 cycles(tsc) - merge slower with cycles:29 250 - 37 cycles(tsc) - 56 cycles(tsc) - merge slower with cycles:19 Joint work with Alexander Duyck. [1] https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/mm/slab_bulk_test01.c [akpm@linux-foundation.org: BUG_ON -> WARN_ON;return] Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Alexander Duyck <alexander.h.duyck@redhat.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-21 07:57:49 +08:00
/*
* This function progressively scans the array with free objects (with
* a limited look ahead) and extract objects belonging to the same
* page. It builds a detached freelist directly within the given
* page/objects. This can happen without any need for
* synchronization, because the objects are owned by running process.
* The freelist is build up as a single linked list in the objects.
* The idea is, that this detached freelist can then be bulk
* transferred to the real freelist(s), but only requiring a single
* synchronization primitive. Look ahead in the array is limited due
* to performance reasons.
*/
static inline
int build_detached_freelist(struct kmem_cache *s, size_t size,
void **p, struct detached_freelist *df)
slub: optimize bulk slowpath free by detached freelist This change focus on improving the speed of object freeing in the "slowpath" of kmem_cache_free_bulk. The calls slab_free (fastpath) and __slab_free (slowpath) have been extended with support for bulk free, which amortize the overhead of the (locked) cmpxchg_double. To use the new bulking feature, we build what I call a detached freelist. The detached freelist takes advantage of three properties: 1) the free function call owns the object that is about to be freed, thus writing into this memory is synchronization-free. 2) many freelist's can co-exist side-by-side in the same slab-page each with a separate head pointer. 3) it is the visibility of the head pointer that needs synchronization. Given these properties, the brilliant part is that the detached freelist can be constructed without any need for synchronization. The freelist is constructed directly in the page objects, without any synchronization needed. The detached freelist is allocated on the stack of the function call kmem_cache_free_bulk. Thus, the freelist head pointer is not visible to other CPUs. All objects in a SLUB freelist must belong to the same slab-page. Thus, constructing the detached freelist is about matching objects that belong to the same slab-page. The bulk free array is scanned is a progressive manor with a limited look-ahead facility. Kmem debug support is handled in call of slab_free(). Notice kmem_cache_free_bulk no longer need to disable IRQs. This only slowed down single free bulk with approx 3 cycles. Performance data: Benchmarked[1] obj size 256 bytes on CPU i7-4790K @ 4.00GHz SLUB fastpath single object quick reuse: 47 cycles(tsc) 11.931 ns To get stable and comparable numbers, the kernel have been booted with "slab_merge" (this also improve performance for larger bulk sizes). Performance data, compared against fallback bulking: bulk - fallback bulk - improvement with this patch 1 - 62 cycles(tsc) 15.662 ns - 49 cycles(tsc) 12.407 ns- improved 21.0% 2 - 55 cycles(tsc) 13.935 ns - 30 cycles(tsc) 7.506 ns - improved 45.5% 3 - 53 cycles(tsc) 13.341 ns - 23 cycles(tsc) 5.865 ns - improved 56.6% 4 - 52 cycles(tsc) 13.081 ns - 20 cycles(tsc) 5.048 ns - improved 61.5% 8 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.659 ns - improved 64.0% 16 - 49 cycles(tsc) 12.412 ns - 17 cycles(tsc) 4.495 ns - improved 65.3% 30 - 49 cycles(tsc) 12.484 ns - 18 cycles(tsc) 4.533 ns - improved 63.3% 32 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.707 ns - improved 64.0% 34 - 96 cycles(tsc) 24.243 ns - 23 cycles(tsc) 5.976 ns - improved 76.0% 48 - 83 cycles(tsc) 20.818 ns - 21 cycles(tsc) 5.329 ns - improved 74.7% 64 - 74 cycles(tsc) 18.700 ns - 20 cycles(tsc) 5.127 ns - improved 73.0% 128 - 90 cycles(tsc) 22.734 ns - 27 cycles(tsc) 6.833 ns - improved 70.0% 158 - 99 cycles(tsc) 24.776 ns - 30 cycles(tsc) 7.583 ns - improved 69.7% 250 - 104 cycles(tsc) 26.089 ns - 37 cycles(tsc) 9.280 ns - improved 64.4% Performance data, compared current in-kernel bulking: bulk - curr in-kernel - improvement with this patch 1 - 46 cycles(tsc) - 49 cycles(tsc) - improved (cycles:-3) -6.5% 2 - 27 cycles(tsc) - 30 cycles(tsc) - improved (cycles:-3) -11.1% 3 - 21 cycles(tsc) - 23 cycles(tsc) - improved (cycles:-2) -9.5% 4 - 18 cycles(tsc) - 20 cycles(tsc) - improved (cycles:-2) -11.1% 8 - 17 cycles(tsc) - 18 cycles(tsc) - improved (cycles:-1) -5.9% 16 - 18 cycles(tsc) - 17 cycles(tsc) - improved (cycles: 1) 5.6% 30 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 32 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 34 - 78 cycles(tsc) - 23 cycles(tsc) - improved (cycles:55) 70.5% 48 - 60 cycles(tsc) - 21 cycles(tsc) - improved (cycles:39) 65.0% 64 - 49 cycles(tsc) - 20 cycles(tsc) - improved (cycles:29) 59.2% 128 - 69 cycles(tsc) - 27 cycles(tsc) - improved (cycles:42) 60.9% 158 - 79 cycles(tsc) - 30 cycles(tsc) - improved (cycles:49) 62.0% 250 - 86 cycles(tsc) - 37 cycles(tsc) - improved (cycles:49) 57.0% Performance with normal SLUB merging is significantly slower for larger bulking. This is believed to (primarily) be an effect of not having to share the per-CPU data-structures, as tuning per-CPU size can achieve similar performance. bulk - slab_nomerge - normal SLUB merge 1 - 49 cycles(tsc) - 49 cycles(tsc) - merge slower with cycles:0 2 - 30 cycles(tsc) - 30 cycles(tsc) - merge slower with cycles:0 3 - 23 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:0 4 - 20 cycles(tsc) - 20 cycles(tsc) - merge slower with cycles:0 8 - 18 cycles(tsc) - 18 cycles(tsc) - merge slower with cycles:0 16 - 17 cycles(tsc) - 17 cycles(tsc) - merge slower with cycles:0 30 - 18 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:5 32 - 18 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:4 34 - 23 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:-1 48 - 21 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:1 64 - 20 cycles(tsc) - 48 cycles(tsc) - merge slower with cycles:28 128 - 27 cycles(tsc) - 57 cycles(tsc) - merge slower with cycles:30 158 - 30 cycles(tsc) - 59 cycles(tsc) - merge slower with cycles:29 250 - 37 cycles(tsc) - 56 cycles(tsc) - merge slower with cycles:19 Joint work with Alexander Duyck. [1] https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/mm/slab_bulk_test01.c [akpm@linux-foundation.org: BUG_ON -> WARN_ON;return] Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Alexander Duyck <alexander.h.duyck@redhat.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-21 07:57:49 +08:00
{
size_t first_skipped_index = 0;
int lookahead = 3;
void *object;
mm: new API kfree_bulk() for SLAB+SLUB allocators This patch introduce a new API call kfree_bulk() for bulk freeing memory objects not bound to a single kmem_cache. Christoph pointed out that it is possible to implement freeing of objects, without knowing the kmem_cache pointer as that information is available from the object's page->slab_cache. Proposing to remove the kmem_cache argument from the bulk free API. Jesper demonstrated that these extra steps per object comes at a performance cost. It is only in the case CONFIG_MEMCG_KMEM is compiled in and activated runtime that these steps are done anyhow. The extra cost is most visible for SLAB allocator, because the SLUB allocator does the page lookup (virt_to_head_page()) anyhow. Thus, the conclusion was to keep the kmem_cache free bulk API with a kmem_cache pointer, but we can still implement a kfree_bulk() API fairly easily. Simply by handling if kmem_cache_free_bulk() gets called with a kmem_cache NULL pointer. This does increase the code size a bit, but implementing a separate kfree_bulk() call would likely increase code size even more. Below benchmarks cost of alloc+free (obj size 256 bytes) on CPU i7-4790K @ 4.00GHz, no PREEMPT and CONFIG_MEMCG_KMEM=y. Code size increase for SLAB: add/remove: 0/0 grow/shrink: 1/0 up/down: 74/0 (74) function old new delta kmem_cache_free_bulk 660 734 +74 SLAB fastpath: 87 cycles(tsc) 21.814 sz - fallback - kmem_cache_free_bulk - kfree_bulk 1 - 103 cycles 25.878 ns - 41 cycles 10.498 ns - 81 cycles 20.312 ns 2 - 94 cycles 23.673 ns - 26 cycles 6.682 ns - 42 cycles 10.649 ns 3 - 92 cycles 23.181 ns - 21 cycles 5.325 ns - 39 cycles 9.950 ns 4 - 90 cycles 22.727 ns - 18 cycles 4.673 ns - 26 cycles 6.693 ns 8 - 89 cycles 22.270 ns - 14 cycles 3.664 ns - 23 cycles 5.835 ns 16 - 88 cycles 22.038 ns - 14 cycles 3.503 ns - 22 cycles 5.543 ns 30 - 89 cycles 22.284 ns - 13 cycles 3.310 ns - 20 cycles 5.197 ns 32 - 88 cycles 22.249 ns - 13 cycles 3.420 ns - 20 cycles 5.166 ns 34 - 88 cycles 22.224 ns - 14 cycles 3.643 ns - 20 cycles 5.170 ns 48 - 88 cycles 22.088 ns - 14 cycles 3.507 ns - 20 cycles 5.203 ns 64 - 88 cycles 22.063 ns - 13 cycles 3.428 ns - 20 cycles 5.152 ns 128 - 89 cycles 22.483 ns - 15 cycles 3.891 ns - 23 cycles 5.885 ns 158 - 89 cycles 22.381 ns - 15 cycles 3.779 ns - 22 cycles 5.548 ns 250 - 91 cycles 22.798 ns - 16 cycles 4.152 ns - 23 cycles 5.967 ns SLAB when enabling MEMCG_KMEM runtime: - kmemcg fastpath: 130 cycles(tsc) 32.684 ns (step:0) 1 - 148 cycles 37.220 ns - 66 cycles 16.622 ns - 66 cycles 16.583 ns 2 - 141 cycles 35.510 ns - 51 cycles 12.820 ns - 58 cycles 14.625 ns 3 - 140 cycles 35.017 ns - 37 cycles 9.326 ns - 33 cycles 8.474 ns 4 - 137 cycles 34.507 ns - 31 cycles 7.888 ns - 33 cycles 8.300 ns 8 - 140 cycles 35.069 ns - 25 cycles 6.461 ns - 25 cycles 6.436 ns 16 - 138 cycles 34.542 ns - 23 cycles 5.945 ns - 22 cycles 5.670 ns 30 - 136 cycles 34.227 ns - 22 cycles 5.502 ns - 22 cycles 5.587 ns 32 - 136 cycles 34.253 ns - 21 cycles 5.475 ns - 21 cycles 5.324 ns 34 - 136 cycles 34.254 ns - 21 cycles 5.448 ns - 20 cycles 5.194 ns 48 - 136 cycles 34.075 ns - 21 cycles 5.458 ns - 21 cycles 5.367 ns 64 - 135 cycles 33.994 ns - 21 cycles 5.350 ns - 21 cycles 5.259 ns 128 - 137 cycles 34.446 ns - 23 cycles 5.816 ns - 22 cycles 5.688 ns 158 - 137 cycles 34.379 ns - 22 cycles 5.727 ns - 22 cycles 5.602 ns 250 - 138 cycles 34.755 ns - 24 cycles 6.093 ns - 23 cycles 5.986 ns Code size increase for SLUB: function old new delta kmem_cache_free_bulk 717 799 +82 SLUB benchmark: SLUB fastpath: 46 cycles(tsc) 11.691 ns (step:0) sz - fallback - kmem_cache_free_bulk - kfree_bulk 1 - 61 cycles 15.486 ns - 53 cycles 13.364 ns - 57 cycles 14.464 ns 2 - 54 cycles 13.703 ns - 32 cycles 8.110 ns - 33 cycles 8.482 ns 3 - 53 cycles 13.272 ns - 25 cycles 6.362 ns - 27 cycles 6.947 ns 4 - 51 cycles 12.994 ns - 24 cycles 6.087 ns - 24 cycles 6.078 ns 8 - 50 cycles 12.576 ns - 21 cycles 5.354 ns - 22 cycles 5.513 ns 16 - 49 cycles 12.368 ns - 20 cycles 5.054 ns - 20 cycles 5.042 ns 30 - 49 cycles 12.273 ns - 18 cycles 4.748 ns - 19 cycles 4.758 ns 32 - 49 cycles 12.401 ns - 19 cycles 4.821 ns - 19 cycles 4.810 ns 34 - 98 cycles 24.519 ns - 24 cycles 6.154 ns - 24 cycles 6.157 ns 48 - 83 cycles 20.833 ns - 21 cycles 5.446 ns - 21 cycles 5.429 ns 64 - 75 cycles 18.891 ns - 20 cycles 5.247 ns - 20 cycles 5.238 ns 128 - 93 cycles 23.271 ns - 27 cycles 6.856 ns - 27 cycles 6.823 ns 158 - 102 cycles 25.581 ns - 30 cycles 7.714 ns - 30 cycles 7.695 ns 250 - 107 cycles 26.917 ns - 38 cycles 9.514 ns - 38 cycles 9.506 ns SLUB when enabling MEMCG_KMEM runtime: - kmemcg fastpath: 71 cycles(tsc) 17.897 ns (step:0) 1 - 85 cycles 21.484 ns - 78 cycles 19.569 ns - 75 cycles 18.938 ns 2 - 81 cycles 20.363 ns - 45 cycles 11.258 ns - 44 cycles 11.076 ns 3 - 78 cycles 19.709 ns - 33 cycles 8.354 ns - 32 cycles 8.044 ns 4 - 77 cycles 19.430 ns - 28 cycles 7.216 ns - 28 cycles 7.003 ns 8 - 101 cycles 25.288 ns - 23 cycles 5.849 ns - 23 cycles 5.787 ns 16 - 76 cycles 19.148 ns - 20 cycles 5.162 ns - 20 cycles 5.081 ns 30 - 76 cycles 19.067 ns - 19 cycles 4.868 ns - 19 cycles 4.821 ns 32 - 76 cycles 19.052 ns - 19 cycles 4.857 ns - 19 cycles 4.815 ns 34 - 121 cycles 30.291 ns - 25 cycles 6.333 ns - 25 cycles 6.268 ns 48 - 108 cycles 27.111 ns - 21 cycles 5.498 ns - 21 cycles 5.458 ns 64 - 100 cycles 25.164 ns - 20 cycles 5.242 ns - 20 cycles 5.229 ns 128 - 155 cycles 38.976 ns - 27 cycles 6.886 ns - 27 cycles 6.892 ns 158 - 132 cycles 33.034 ns - 30 cycles 7.711 ns - 30 cycles 7.728 ns 250 - 130 cycles 32.612 ns - 38 cycles 9.560 ns - 38 cycles 9.549 ns Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Vladimir Davydov <vdavydov@virtuozzo.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-03-16 05:54:00 +08:00
struct page *page;
slub: initial bulk free implementation This implements SLUB specific kmem_cache_free_bulk(). SLUB allocator now both have bulk alloc and free implemented. Choose to reenable local IRQs while calling slowpath __slab_free(). In worst case, where all objects hit slowpath call, the performance should still be faster than fallback function __kmem_cache_free_bulk(), because local_irq_{disable+enable} is very fast (7-cycles), while the fallback invokes this_cpu_cmpxchg() which is slightly slower (9-cycles). Nitpicking, this should be faster for N>=4, due to the entry cost of local_irq_{disable+enable}. Do notice that the save+restore variant is very expensive, this is key to why this optimization works. CPU: i7-4790K CPU @ 4.00GHz * local_irq_{disable,enable}: 7 cycles(tsc) - 1.821 ns * local_irq_{save,restore} : 37 cycles(tsc) - 9.443 ns Measurements on CPU CPU i7-4790K @ 4.00GHz Baseline normal fastpath (alloc+free cost): 43 cycles(tsc) 10.834 ns Bulk- fallback - this-patch 1 - 58 cycles(tsc) 14.542 ns - 43 cycles(tsc) 10.811 ns improved 25.9% 2 - 50 cycles(tsc) 12.659 ns - 27 cycles(tsc) 6.867 ns improved 46.0% 3 - 48 cycles(tsc) 12.168 ns - 21 cycles(tsc) 5.496 ns improved 56.2% 4 - 47 cycles(tsc) 11.987 ns - 24 cycles(tsc) 6.038 ns improved 48.9% 8 - 46 cycles(tsc) 11.518 ns - 17 cycles(tsc) 4.280 ns improved 63.0% 16 - 45 cycles(tsc) 11.366 ns - 17 cycles(tsc) 4.483 ns improved 62.2% 30 - 45 cycles(tsc) 11.433 ns - 18 cycles(tsc) 4.531 ns improved 60.0% 32 - 75 cycles(tsc) 18.983 ns - 58 cycles(tsc) 14.586 ns improved 22.7% 34 - 71 cycles(tsc) 17.940 ns - 53 cycles(tsc) 13.391 ns improved 25.4% 48 - 80 cycles(tsc) 20.077 ns - 65 cycles(tsc) 16.268 ns improved 18.8% 64 - 71 cycles(tsc) 17.799 ns - 53 cycles(tsc) 13.440 ns improved 25.4% 128 - 91 cycles(tsc) 22.980 ns - 79 cycles(tsc) 19.899 ns improved 13.2% 158 - 100 cycles(tsc) 25.241 ns - 90 cycles(tsc) 22.732 ns improved 10.0% 250 - 102 cycles(tsc) 25.583 ns - 95 cycles(tsc) 23.916 ns improved 6.9% Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-05 06:45:43 +08:00
slub: optimize bulk slowpath free by detached freelist This change focus on improving the speed of object freeing in the "slowpath" of kmem_cache_free_bulk. The calls slab_free (fastpath) and __slab_free (slowpath) have been extended with support for bulk free, which amortize the overhead of the (locked) cmpxchg_double. To use the new bulking feature, we build what I call a detached freelist. The detached freelist takes advantage of three properties: 1) the free function call owns the object that is about to be freed, thus writing into this memory is synchronization-free. 2) many freelist's can co-exist side-by-side in the same slab-page each with a separate head pointer. 3) it is the visibility of the head pointer that needs synchronization. Given these properties, the brilliant part is that the detached freelist can be constructed without any need for synchronization. The freelist is constructed directly in the page objects, without any synchronization needed. The detached freelist is allocated on the stack of the function call kmem_cache_free_bulk. Thus, the freelist head pointer is not visible to other CPUs. All objects in a SLUB freelist must belong to the same slab-page. Thus, constructing the detached freelist is about matching objects that belong to the same slab-page. The bulk free array is scanned is a progressive manor with a limited look-ahead facility. Kmem debug support is handled in call of slab_free(). Notice kmem_cache_free_bulk no longer need to disable IRQs. This only slowed down single free bulk with approx 3 cycles. Performance data: Benchmarked[1] obj size 256 bytes on CPU i7-4790K @ 4.00GHz SLUB fastpath single object quick reuse: 47 cycles(tsc) 11.931 ns To get stable and comparable numbers, the kernel have been booted with "slab_merge" (this also improve performance for larger bulk sizes). Performance data, compared against fallback bulking: bulk - fallback bulk - improvement with this patch 1 - 62 cycles(tsc) 15.662 ns - 49 cycles(tsc) 12.407 ns- improved 21.0% 2 - 55 cycles(tsc) 13.935 ns - 30 cycles(tsc) 7.506 ns - improved 45.5% 3 - 53 cycles(tsc) 13.341 ns - 23 cycles(tsc) 5.865 ns - improved 56.6% 4 - 52 cycles(tsc) 13.081 ns - 20 cycles(tsc) 5.048 ns - improved 61.5% 8 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.659 ns - improved 64.0% 16 - 49 cycles(tsc) 12.412 ns - 17 cycles(tsc) 4.495 ns - improved 65.3% 30 - 49 cycles(tsc) 12.484 ns - 18 cycles(tsc) 4.533 ns - improved 63.3% 32 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.707 ns - improved 64.0% 34 - 96 cycles(tsc) 24.243 ns - 23 cycles(tsc) 5.976 ns - improved 76.0% 48 - 83 cycles(tsc) 20.818 ns - 21 cycles(tsc) 5.329 ns - improved 74.7% 64 - 74 cycles(tsc) 18.700 ns - 20 cycles(tsc) 5.127 ns - improved 73.0% 128 - 90 cycles(tsc) 22.734 ns - 27 cycles(tsc) 6.833 ns - improved 70.0% 158 - 99 cycles(tsc) 24.776 ns - 30 cycles(tsc) 7.583 ns - improved 69.7% 250 - 104 cycles(tsc) 26.089 ns - 37 cycles(tsc) 9.280 ns - improved 64.4% Performance data, compared current in-kernel bulking: bulk - curr in-kernel - improvement with this patch 1 - 46 cycles(tsc) - 49 cycles(tsc) - improved (cycles:-3) -6.5% 2 - 27 cycles(tsc) - 30 cycles(tsc) - improved (cycles:-3) -11.1% 3 - 21 cycles(tsc) - 23 cycles(tsc) - improved (cycles:-2) -9.5% 4 - 18 cycles(tsc) - 20 cycles(tsc) - improved (cycles:-2) -11.1% 8 - 17 cycles(tsc) - 18 cycles(tsc) - improved (cycles:-1) -5.9% 16 - 18 cycles(tsc) - 17 cycles(tsc) - improved (cycles: 1) 5.6% 30 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 32 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 34 - 78 cycles(tsc) - 23 cycles(tsc) - improved (cycles:55) 70.5% 48 - 60 cycles(tsc) - 21 cycles(tsc) - improved (cycles:39) 65.0% 64 - 49 cycles(tsc) - 20 cycles(tsc) - improved (cycles:29) 59.2% 128 - 69 cycles(tsc) - 27 cycles(tsc) - improved (cycles:42) 60.9% 158 - 79 cycles(tsc) - 30 cycles(tsc) - improved (cycles:49) 62.0% 250 - 86 cycles(tsc) - 37 cycles(tsc) - improved (cycles:49) 57.0% Performance with normal SLUB merging is significantly slower for larger bulking. This is believed to (primarily) be an effect of not having to share the per-CPU data-structures, as tuning per-CPU size can achieve similar performance. bulk - slab_nomerge - normal SLUB merge 1 - 49 cycles(tsc) - 49 cycles(tsc) - merge slower with cycles:0 2 - 30 cycles(tsc) - 30 cycles(tsc) - merge slower with cycles:0 3 - 23 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:0 4 - 20 cycles(tsc) - 20 cycles(tsc) - merge slower with cycles:0 8 - 18 cycles(tsc) - 18 cycles(tsc) - merge slower with cycles:0 16 - 17 cycles(tsc) - 17 cycles(tsc) - merge slower with cycles:0 30 - 18 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:5 32 - 18 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:4 34 - 23 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:-1 48 - 21 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:1 64 - 20 cycles(tsc) - 48 cycles(tsc) - merge slower with cycles:28 128 - 27 cycles(tsc) - 57 cycles(tsc) - merge slower with cycles:30 158 - 30 cycles(tsc) - 59 cycles(tsc) - merge slower with cycles:29 250 - 37 cycles(tsc) - 56 cycles(tsc) - merge slower with cycles:19 Joint work with Alexander Duyck. [1] https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/mm/slab_bulk_test01.c [akpm@linux-foundation.org: BUG_ON -> WARN_ON;return] Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Alexander Duyck <alexander.h.duyck@redhat.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-21 07:57:49 +08:00
/* Always re-init detached_freelist */
df->page = NULL;
slub: initial bulk free implementation This implements SLUB specific kmem_cache_free_bulk(). SLUB allocator now both have bulk alloc and free implemented. Choose to reenable local IRQs while calling slowpath __slab_free(). In worst case, where all objects hit slowpath call, the performance should still be faster than fallback function __kmem_cache_free_bulk(), because local_irq_{disable+enable} is very fast (7-cycles), while the fallback invokes this_cpu_cmpxchg() which is slightly slower (9-cycles). Nitpicking, this should be faster for N>=4, due to the entry cost of local_irq_{disable+enable}. Do notice that the save+restore variant is very expensive, this is key to why this optimization works. CPU: i7-4790K CPU @ 4.00GHz * local_irq_{disable,enable}: 7 cycles(tsc) - 1.821 ns * local_irq_{save,restore} : 37 cycles(tsc) - 9.443 ns Measurements on CPU CPU i7-4790K @ 4.00GHz Baseline normal fastpath (alloc+free cost): 43 cycles(tsc) 10.834 ns Bulk- fallback - this-patch 1 - 58 cycles(tsc) 14.542 ns - 43 cycles(tsc) 10.811 ns improved 25.9% 2 - 50 cycles(tsc) 12.659 ns - 27 cycles(tsc) 6.867 ns improved 46.0% 3 - 48 cycles(tsc) 12.168 ns - 21 cycles(tsc) 5.496 ns improved 56.2% 4 - 47 cycles(tsc) 11.987 ns - 24 cycles(tsc) 6.038 ns improved 48.9% 8 - 46 cycles(tsc) 11.518 ns - 17 cycles(tsc) 4.280 ns improved 63.0% 16 - 45 cycles(tsc) 11.366 ns - 17 cycles(tsc) 4.483 ns improved 62.2% 30 - 45 cycles(tsc) 11.433 ns - 18 cycles(tsc) 4.531 ns improved 60.0% 32 - 75 cycles(tsc) 18.983 ns - 58 cycles(tsc) 14.586 ns improved 22.7% 34 - 71 cycles(tsc) 17.940 ns - 53 cycles(tsc) 13.391 ns improved 25.4% 48 - 80 cycles(tsc) 20.077 ns - 65 cycles(tsc) 16.268 ns improved 18.8% 64 - 71 cycles(tsc) 17.799 ns - 53 cycles(tsc) 13.440 ns improved 25.4% 128 - 91 cycles(tsc) 22.980 ns - 79 cycles(tsc) 19.899 ns improved 13.2% 158 - 100 cycles(tsc) 25.241 ns - 90 cycles(tsc) 22.732 ns improved 10.0% 250 - 102 cycles(tsc) 25.583 ns - 95 cycles(tsc) 23.916 ns improved 6.9% Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-05 06:45:43 +08:00
slub: optimize bulk slowpath free by detached freelist This change focus on improving the speed of object freeing in the "slowpath" of kmem_cache_free_bulk. The calls slab_free (fastpath) and __slab_free (slowpath) have been extended with support for bulk free, which amortize the overhead of the (locked) cmpxchg_double. To use the new bulking feature, we build what I call a detached freelist. The detached freelist takes advantage of three properties: 1) the free function call owns the object that is about to be freed, thus writing into this memory is synchronization-free. 2) many freelist's can co-exist side-by-side in the same slab-page each with a separate head pointer. 3) it is the visibility of the head pointer that needs synchronization. Given these properties, the brilliant part is that the detached freelist can be constructed without any need for synchronization. The freelist is constructed directly in the page objects, without any synchronization needed. The detached freelist is allocated on the stack of the function call kmem_cache_free_bulk. Thus, the freelist head pointer is not visible to other CPUs. All objects in a SLUB freelist must belong to the same slab-page. Thus, constructing the detached freelist is about matching objects that belong to the same slab-page. The bulk free array is scanned is a progressive manor with a limited look-ahead facility. Kmem debug support is handled in call of slab_free(). Notice kmem_cache_free_bulk no longer need to disable IRQs. This only slowed down single free bulk with approx 3 cycles. Performance data: Benchmarked[1] obj size 256 bytes on CPU i7-4790K @ 4.00GHz SLUB fastpath single object quick reuse: 47 cycles(tsc) 11.931 ns To get stable and comparable numbers, the kernel have been booted with "slab_merge" (this also improve performance for larger bulk sizes). Performance data, compared against fallback bulking: bulk - fallback bulk - improvement with this patch 1 - 62 cycles(tsc) 15.662 ns - 49 cycles(tsc) 12.407 ns- improved 21.0% 2 - 55 cycles(tsc) 13.935 ns - 30 cycles(tsc) 7.506 ns - improved 45.5% 3 - 53 cycles(tsc) 13.341 ns - 23 cycles(tsc) 5.865 ns - improved 56.6% 4 - 52 cycles(tsc) 13.081 ns - 20 cycles(tsc) 5.048 ns - improved 61.5% 8 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.659 ns - improved 64.0% 16 - 49 cycles(tsc) 12.412 ns - 17 cycles(tsc) 4.495 ns - improved 65.3% 30 - 49 cycles(tsc) 12.484 ns - 18 cycles(tsc) 4.533 ns - improved 63.3% 32 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.707 ns - improved 64.0% 34 - 96 cycles(tsc) 24.243 ns - 23 cycles(tsc) 5.976 ns - improved 76.0% 48 - 83 cycles(tsc) 20.818 ns - 21 cycles(tsc) 5.329 ns - improved 74.7% 64 - 74 cycles(tsc) 18.700 ns - 20 cycles(tsc) 5.127 ns - improved 73.0% 128 - 90 cycles(tsc) 22.734 ns - 27 cycles(tsc) 6.833 ns - improved 70.0% 158 - 99 cycles(tsc) 24.776 ns - 30 cycles(tsc) 7.583 ns - improved 69.7% 250 - 104 cycles(tsc) 26.089 ns - 37 cycles(tsc) 9.280 ns - improved 64.4% Performance data, compared current in-kernel bulking: bulk - curr in-kernel - improvement with this patch 1 - 46 cycles(tsc) - 49 cycles(tsc) - improved (cycles:-3) -6.5% 2 - 27 cycles(tsc) - 30 cycles(tsc) - improved (cycles:-3) -11.1% 3 - 21 cycles(tsc) - 23 cycles(tsc) - improved (cycles:-2) -9.5% 4 - 18 cycles(tsc) - 20 cycles(tsc) - improved (cycles:-2) -11.1% 8 - 17 cycles(tsc) - 18 cycles(tsc) - improved (cycles:-1) -5.9% 16 - 18 cycles(tsc) - 17 cycles(tsc) - improved (cycles: 1) 5.6% 30 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 32 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 34 - 78 cycles(tsc) - 23 cycles(tsc) - improved (cycles:55) 70.5% 48 - 60 cycles(tsc) - 21 cycles(tsc) - improved (cycles:39) 65.0% 64 - 49 cycles(tsc) - 20 cycles(tsc) - improved (cycles:29) 59.2% 128 - 69 cycles(tsc) - 27 cycles(tsc) - improved (cycles:42) 60.9% 158 - 79 cycles(tsc) - 30 cycles(tsc) - improved (cycles:49) 62.0% 250 - 86 cycles(tsc) - 37 cycles(tsc) - improved (cycles:49) 57.0% Performance with normal SLUB merging is significantly slower for larger bulking. This is believed to (primarily) be an effect of not having to share the per-CPU data-structures, as tuning per-CPU size can achieve similar performance. bulk - slab_nomerge - normal SLUB merge 1 - 49 cycles(tsc) - 49 cycles(tsc) - merge slower with cycles:0 2 - 30 cycles(tsc) - 30 cycles(tsc) - merge slower with cycles:0 3 - 23 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:0 4 - 20 cycles(tsc) - 20 cycles(tsc) - merge slower with cycles:0 8 - 18 cycles(tsc) - 18 cycles(tsc) - merge slower with cycles:0 16 - 17 cycles(tsc) - 17 cycles(tsc) - merge slower with cycles:0 30 - 18 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:5 32 - 18 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:4 34 - 23 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:-1 48 - 21 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:1 64 - 20 cycles(tsc) - 48 cycles(tsc) - merge slower with cycles:28 128 - 27 cycles(tsc) - 57 cycles(tsc) - merge slower with cycles:30 158 - 30 cycles(tsc) - 59 cycles(tsc) - merge slower with cycles:29 250 - 37 cycles(tsc) - 56 cycles(tsc) - merge slower with cycles:19 Joint work with Alexander Duyck. [1] https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/mm/slab_bulk_test01.c [akpm@linux-foundation.org: BUG_ON -> WARN_ON;return] Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Alexander Duyck <alexander.h.duyck@redhat.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-21 07:57:49 +08:00
do {
object = p[--size];
mm: new API kfree_bulk() for SLAB+SLUB allocators This patch introduce a new API call kfree_bulk() for bulk freeing memory objects not bound to a single kmem_cache. Christoph pointed out that it is possible to implement freeing of objects, without knowing the kmem_cache pointer as that information is available from the object's page->slab_cache. Proposing to remove the kmem_cache argument from the bulk free API. Jesper demonstrated that these extra steps per object comes at a performance cost. It is only in the case CONFIG_MEMCG_KMEM is compiled in and activated runtime that these steps are done anyhow. The extra cost is most visible for SLAB allocator, because the SLUB allocator does the page lookup (virt_to_head_page()) anyhow. Thus, the conclusion was to keep the kmem_cache free bulk API with a kmem_cache pointer, but we can still implement a kfree_bulk() API fairly easily. Simply by handling if kmem_cache_free_bulk() gets called with a kmem_cache NULL pointer. This does increase the code size a bit, but implementing a separate kfree_bulk() call would likely increase code size even more. Below benchmarks cost of alloc+free (obj size 256 bytes) on CPU i7-4790K @ 4.00GHz, no PREEMPT and CONFIG_MEMCG_KMEM=y. Code size increase for SLAB: add/remove: 0/0 grow/shrink: 1/0 up/down: 74/0 (74) function old new delta kmem_cache_free_bulk 660 734 +74 SLAB fastpath: 87 cycles(tsc) 21.814 sz - fallback - kmem_cache_free_bulk - kfree_bulk 1 - 103 cycles 25.878 ns - 41 cycles 10.498 ns - 81 cycles 20.312 ns 2 - 94 cycles 23.673 ns - 26 cycles 6.682 ns - 42 cycles 10.649 ns 3 - 92 cycles 23.181 ns - 21 cycles 5.325 ns - 39 cycles 9.950 ns 4 - 90 cycles 22.727 ns - 18 cycles 4.673 ns - 26 cycles 6.693 ns 8 - 89 cycles 22.270 ns - 14 cycles 3.664 ns - 23 cycles 5.835 ns 16 - 88 cycles 22.038 ns - 14 cycles 3.503 ns - 22 cycles 5.543 ns 30 - 89 cycles 22.284 ns - 13 cycles 3.310 ns - 20 cycles 5.197 ns 32 - 88 cycles 22.249 ns - 13 cycles 3.420 ns - 20 cycles 5.166 ns 34 - 88 cycles 22.224 ns - 14 cycles 3.643 ns - 20 cycles 5.170 ns 48 - 88 cycles 22.088 ns - 14 cycles 3.507 ns - 20 cycles 5.203 ns 64 - 88 cycles 22.063 ns - 13 cycles 3.428 ns - 20 cycles 5.152 ns 128 - 89 cycles 22.483 ns - 15 cycles 3.891 ns - 23 cycles 5.885 ns 158 - 89 cycles 22.381 ns - 15 cycles 3.779 ns - 22 cycles 5.548 ns 250 - 91 cycles 22.798 ns - 16 cycles 4.152 ns - 23 cycles 5.967 ns SLAB when enabling MEMCG_KMEM runtime: - kmemcg fastpath: 130 cycles(tsc) 32.684 ns (step:0) 1 - 148 cycles 37.220 ns - 66 cycles 16.622 ns - 66 cycles 16.583 ns 2 - 141 cycles 35.510 ns - 51 cycles 12.820 ns - 58 cycles 14.625 ns 3 - 140 cycles 35.017 ns - 37 cycles 9.326 ns - 33 cycles 8.474 ns 4 - 137 cycles 34.507 ns - 31 cycles 7.888 ns - 33 cycles 8.300 ns 8 - 140 cycles 35.069 ns - 25 cycles 6.461 ns - 25 cycles 6.436 ns 16 - 138 cycles 34.542 ns - 23 cycles 5.945 ns - 22 cycles 5.670 ns 30 - 136 cycles 34.227 ns - 22 cycles 5.502 ns - 22 cycles 5.587 ns 32 - 136 cycles 34.253 ns - 21 cycles 5.475 ns - 21 cycles 5.324 ns 34 - 136 cycles 34.254 ns - 21 cycles 5.448 ns - 20 cycles 5.194 ns 48 - 136 cycles 34.075 ns - 21 cycles 5.458 ns - 21 cycles 5.367 ns 64 - 135 cycles 33.994 ns - 21 cycles 5.350 ns - 21 cycles 5.259 ns 128 - 137 cycles 34.446 ns - 23 cycles 5.816 ns - 22 cycles 5.688 ns 158 - 137 cycles 34.379 ns - 22 cycles 5.727 ns - 22 cycles 5.602 ns 250 - 138 cycles 34.755 ns - 24 cycles 6.093 ns - 23 cycles 5.986 ns Code size increase for SLUB: function old new delta kmem_cache_free_bulk 717 799 +82 SLUB benchmark: SLUB fastpath: 46 cycles(tsc) 11.691 ns (step:0) sz - fallback - kmem_cache_free_bulk - kfree_bulk 1 - 61 cycles 15.486 ns - 53 cycles 13.364 ns - 57 cycles 14.464 ns 2 - 54 cycles 13.703 ns - 32 cycles 8.110 ns - 33 cycles 8.482 ns 3 - 53 cycles 13.272 ns - 25 cycles 6.362 ns - 27 cycles 6.947 ns 4 - 51 cycles 12.994 ns - 24 cycles 6.087 ns - 24 cycles 6.078 ns 8 - 50 cycles 12.576 ns - 21 cycles 5.354 ns - 22 cycles 5.513 ns 16 - 49 cycles 12.368 ns - 20 cycles 5.054 ns - 20 cycles 5.042 ns 30 - 49 cycles 12.273 ns - 18 cycles 4.748 ns - 19 cycles 4.758 ns 32 - 49 cycles 12.401 ns - 19 cycles 4.821 ns - 19 cycles 4.810 ns 34 - 98 cycles 24.519 ns - 24 cycles 6.154 ns - 24 cycles 6.157 ns 48 - 83 cycles 20.833 ns - 21 cycles 5.446 ns - 21 cycles 5.429 ns 64 - 75 cycles 18.891 ns - 20 cycles 5.247 ns - 20 cycles 5.238 ns 128 - 93 cycles 23.271 ns - 27 cycles 6.856 ns - 27 cycles 6.823 ns 158 - 102 cycles 25.581 ns - 30 cycles 7.714 ns - 30 cycles 7.695 ns 250 - 107 cycles 26.917 ns - 38 cycles 9.514 ns - 38 cycles 9.506 ns SLUB when enabling MEMCG_KMEM runtime: - kmemcg fastpath: 71 cycles(tsc) 17.897 ns (step:0) 1 - 85 cycles 21.484 ns - 78 cycles 19.569 ns - 75 cycles 18.938 ns 2 - 81 cycles 20.363 ns - 45 cycles 11.258 ns - 44 cycles 11.076 ns 3 - 78 cycles 19.709 ns - 33 cycles 8.354 ns - 32 cycles 8.044 ns 4 - 77 cycles 19.430 ns - 28 cycles 7.216 ns - 28 cycles 7.003 ns 8 - 101 cycles 25.288 ns - 23 cycles 5.849 ns - 23 cycles 5.787 ns 16 - 76 cycles 19.148 ns - 20 cycles 5.162 ns - 20 cycles 5.081 ns 30 - 76 cycles 19.067 ns - 19 cycles 4.868 ns - 19 cycles 4.821 ns 32 - 76 cycles 19.052 ns - 19 cycles 4.857 ns - 19 cycles 4.815 ns 34 - 121 cycles 30.291 ns - 25 cycles 6.333 ns - 25 cycles 6.268 ns 48 - 108 cycles 27.111 ns - 21 cycles 5.498 ns - 21 cycles 5.458 ns 64 - 100 cycles 25.164 ns - 20 cycles 5.242 ns - 20 cycles 5.229 ns 128 - 155 cycles 38.976 ns - 27 cycles 6.886 ns - 27 cycles 6.892 ns 158 - 132 cycles 33.034 ns - 30 cycles 7.711 ns - 30 cycles 7.728 ns 250 - 130 cycles 32.612 ns - 38 cycles 9.560 ns - 38 cycles 9.549 ns Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Vladimir Davydov <vdavydov@virtuozzo.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-03-16 05:54:00 +08:00
/* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
slub: optimize bulk slowpath free by detached freelist This change focus on improving the speed of object freeing in the "slowpath" of kmem_cache_free_bulk. The calls slab_free (fastpath) and __slab_free (slowpath) have been extended with support for bulk free, which amortize the overhead of the (locked) cmpxchg_double. To use the new bulking feature, we build what I call a detached freelist. The detached freelist takes advantage of three properties: 1) the free function call owns the object that is about to be freed, thus writing into this memory is synchronization-free. 2) many freelist's can co-exist side-by-side in the same slab-page each with a separate head pointer. 3) it is the visibility of the head pointer that needs synchronization. Given these properties, the brilliant part is that the detached freelist can be constructed without any need for synchronization. The freelist is constructed directly in the page objects, without any synchronization needed. The detached freelist is allocated on the stack of the function call kmem_cache_free_bulk. Thus, the freelist head pointer is not visible to other CPUs. All objects in a SLUB freelist must belong to the same slab-page. Thus, constructing the detached freelist is about matching objects that belong to the same slab-page. The bulk free array is scanned is a progressive manor with a limited look-ahead facility. Kmem debug support is handled in call of slab_free(). Notice kmem_cache_free_bulk no longer need to disable IRQs. This only slowed down single free bulk with approx 3 cycles. Performance data: Benchmarked[1] obj size 256 bytes on CPU i7-4790K @ 4.00GHz SLUB fastpath single object quick reuse: 47 cycles(tsc) 11.931 ns To get stable and comparable numbers, the kernel have been booted with "slab_merge" (this also improve performance for larger bulk sizes). Performance data, compared against fallback bulking: bulk - fallback bulk - improvement with this patch 1 - 62 cycles(tsc) 15.662 ns - 49 cycles(tsc) 12.407 ns- improved 21.0% 2 - 55 cycles(tsc) 13.935 ns - 30 cycles(tsc) 7.506 ns - improved 45.5% 3 - 53 cycles(tsc) 13.341 ns - 23 cycles(tsc) 5.865 ns - improved 56.6% 4 - 52 cycles(tsc) 13.081 ns - 20 cycles(tsc) 5.048 ns - improved 61.5% 8 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.659 ns - improved 64.0% 16 - 49 cycles(tsc) 12.412 ns - 17 cycles(tsc) 4.495 ns - improved 65.3% 30 - 49 cycles(tsc) 12.484 ns - 18 cycles(tsc) 4.533 ns - improved 63.3% 32 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.707 ns - improved 64.0% 34 - 96 cycles(tsc) 24.243 ns - 23 cycles(tsc) 5.976 ns - improved 76.0% 48 - 83 cycles(tsc) 20.818 ns - 21 cycles(tsc) 5.329 ns - improved 74.7% 64 - 74 cycles(tsc) 18.700 ns - 20 cycles(tsc) 5.127 ns - improved 73.0% 128 - 90 cycles(tsc) 22.734 ns - 27 cycles(tsc) 6.833 ns - improved 70.0% 158 - 99 cycles(tsc) 24.776 ns - 30 cycles(tsc) 7.583 ns - improved 69.7% 250 - 104 cycles(tsc) 26.089 ns - 37 cycles(tsc) 9.280 ns - improved 64.4% Performance data, compared current in-kernel bulking: bulk - curr in-kernel - improvement with this patch 1 - 46 cycles(tsc) - 49 cycles(tsc) - improved (cycles:-3) -6.5% 2 - 27 cycles(tsc) - 30 cycles(tsc) - improved (cycles:-3) -11.1% 3 - 21 cycles(tsc) - 23 cycles(tsc) - improved (cycles:-2) -9.5% 4 - 18 cycles(tsc) - 20 cycles(tsc) - improved (cycles:-2) -11.1% 8 - 17 cycles(tsc) - 18 cycles(tsc) - improved (cycles:-1) -5.9% 16 - 18 cycles(tsc) - 17 cycles(tsc) - improved (cycles: 1) 5.6% 30 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 32 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 34 - 78 cycles(tsc) - 23 cycles(tsc) - improved (cycles:55) 70.5% 48 - 60 cycles(tsc) - 21 cycles(tsc) - improved (cycles:39) 65.0% 64 - 49 cycles(tsc) - 20 cycles(tsc) - improved (cycles:29) 59.2% 128 - 69 cycles(tsc) - 27 cycles(tsc) - improved (cycles:42) 60.9% 158 - 79 cycles(tsc) - 30 cycles(tsc) - improved (cycles:49) 62.0% 250 - 86 cycles(tsc) - 37 cycles(tsc) - improved (cycles:49) 57.0% Performance with normal SLUB merging is significantly slower for larger bulking. This is believed to (primarily) be an effect of not having to share the per-CPU data-structures, as tuning per-CPU size can achieve similar performance. bulk - slab_nomerge - normal SLUB merge 1 - 49 cycles(tsc) - 49 cycles(tsc) - merge slower with cycles:0 2 - 30 cycles(tsc) - 30 cycles(tsc) - merge slower with cycles:0 3 - 23 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:0 4 - 20 cycles(tsc) - 20 cycles(tsc) - merge slower with cycles:0 8 - 18 cycles(tsc) - 18 cycles(tsc) - merge slower with cycles:0 16 - 17 cycles(tsc) - 17 cycles(tsc) - merge slower with cycles:0 30 - 18 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:5 32 - 18 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:4 34 - 23 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:-1 48 - 21 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:1 64 - 20 cycles(tsc) - 48 cycles(tsc) - merge slower with cycles:28 128 - 27 cycles(tsc) - 57 cycles(tsc) - merge slower with cycles:30 158 - 30 cycles(tsc) - 59 cycles(tsc) - merge slower with cycles:29 250 - 37 cycles(tsc) - 56 cycles(tsc) - merge slower with cycles:19 Joint work with Alexander Duyck. [1] https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/mm/slab_bulk_test01.c [akpm@linux-foundation.org: BUG_ON -> WARN_ON;return] Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Alexander Duyck <alexander.h.duyck@redhat.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-21 07:57:49 +08:00
} while (!object && size);
slub: add support for kmem_cache_debug in bulk calls Per request of Joonsoo Kim adding kmem debug support. I've tested that when debugging is disabled, then there is almost no performance impact as this code basically gets removed by the compiler. Need some guidance in enabling and testing this. bulk- PREVIOUS - THIS-PATCH 1 - 43 cycles(tsc) 10.811 ns - 44 cycles(tsc) 11.236 ns improved -2.3% 2 - 27 cycles(tsc) 6.867 ns - 28 cycles(tsc) 7.019 ns improved -3.7% 3 - 21 cycles(tsc) 5.496 ns - 22 cycles(tsc) 5.526 ns improved -4.8% 4 - 24 cycles(tsc) 6.038 ns - 19 cycles(tsc) 4.786 ns improved 20.8% 8 - 17 cycles(tsc) 4.280 ns - 18 cycles(tsc) 4.572 ns improved -5.9% 16 - 17 cycles(tsc) 4.483 ns - 18 cycles(tsc) 4.658 ns improved -5.9% 30 - 18 cycles(tsc) 4.531 ns - 18 cycles(tsc) 4.568 ns improved 0.0% 32 - 58 cycles(tsc) 14.586 ns - 65 cycles(tsc) 16.454 ns improved -12.1% 34 - 53 cycles(tsc) 13.391 ns - 63 cycles(tsc) 15.932 ns improved -18.9% 48 - 65 cycles(tsc) 16.268 ns - 50 cycles(tsc) 12.506 ns improved 23.1% 64 - 53 cycles(tsc) 13.440 ns - 63 cycles(tsc) 15.929 ns improved -18.9% 128 - 79 cycles(tsc) 19.899 ns - 86 cycles(tsc) 21.583 ns improved -8.9% 158 - 90 cycles(tsc) 22.732 ns - 90 cycles(tsc) 22.552 ns improved 0.0% 250 - 95 cycles(tsc) 23.916 ns - 98 cycles(tsc) 24.589 ns improved -3.2% Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-05 06:45:45 +08:00
slub: optimize bulk slowpath free by detached freelist This change focus on improving the speed of object freeing in the "slowpath" of kmem_cache_free_bulk. The calls slab_free (fastpath) and __slab_free (slowpath) have been extended with support for bulk free, which amortize the overhead of the (locked) cmpxchg_double. To use the new bulking feature, we build what I call a detached freelist. The detached freelist takes advantage of three properties: 1) the free function call owns the object that is about to be freed, thus writing into this memory is synchronization-free. 2) many freelist's can co-exist side-by-side in the same slab-page each with a separate head pointer. 3) it is the visibility of the head pointer that needs synchronization. Given these properties, the brilliant part is that the detached freelist can be constructed without any need for synchronization. The freelist is constructed directly in the page objects, without any synchronization needed. The detached freelist is allocated on the stack of the function call kmem_cache_free_bulk. Thus, the freelist head pointer is not visible to other CPUs. All objects in a SLUB freelist must belong to the same slab-page. Thus, constructing the detached freelist is about matching objects that belong to the same slab-page. The bulk free array is scanned is a progressive manor with a limited look-ahead facility. Kmem debug support is handled in call of slab_free(). Notice kmem_cache_free_bulk no longer need to disable IRQs. This only slowed down single free bulk with approx 3 cycles. Performance data: Benchmarked[1] obj size 256 bytes on CPU i7-4790K @ 4.00GHz SLUB fastpath single object quick reuse: 47 cycles(tsc) 11.931 ns To get stable and comparable numbers, the kernel have been booted with "slab_merge" (this also improve performance for larger bulk sizes). Performance data, compared against fallback bulking: bulk - fallback bulk - improvement with this patch 1 - 62 cycles(tsc) 15.662 ns - 49 cycles(tsc) 12.407 ns- improved 21.0% 2 - 55 cycles(tsc) 13.935 ns - 30 cycles(tsc) 7.506 ns - improved 45.5% 3 - 53 cycles(tsc) 13.341 ns - 23 cycles(tsc) 5.865 ns - improved 56.6% 4 - 52 cycles(tsc) 13.081 ns - 20 cycles(tsc) 5.048 ns - improved 61.5% 8 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.659 ns - improved 64.0% 16 - 49 cycles(tsc) 12.412 ns - 17 cycles(tsc) 4.495 ns - improved 65.3% 30 - 49 cycles(tsc) 12.484 ns - 18 cycles(tsc) 4.533 ns - improved 63.3% 32 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.707 ns - improved 64.0% 34 - 96 cycles(tsc) 24.243 ns - 23 cycles(tsc) 5.976 ns - improved 76.0% 48 - 83 cycles(tsc) 20.818 ns - 21 cycles(tsc) 5.329 ns - improved 74.7% 64 - 74 cycles(tsc) 18.700 ns - 20 cycles(tsc) 5.127 ns - improved 73.0% 128 - 90 cycles(tsc) 22.734 ns - 27 cycles(tsc) 6.833 ns - improved 70.0% 158 - 99 cycles(tsc) 24.776 ns - 30 cycles(tsc) 7.583 ns - improved 69.7% 250 - 104 cycles(tsc) 26.089 ns - 37 cycles(tsc) 9.280 ns - improved 64.4% Performance data, compared current in-kernel bulking: bulk - curr in-kernel - improvement with this patch 1 - 46 cycles(tsc) - 49 cycles(tsc) - improved (cycles:-3) -6.5% 2 - 27 cycles(tsc) - 30 cycles(tsc) - improved (cycles:-3) -11.1% 3 - 21 cycles(tsc) - 23 cycles(tsc) - improved (cycles:-2) -9.5% 4 - 18 cycles(tsc) - 20 cycles(tsc) - improved (cycles:-2) -11.1% 8 - 17 cycles(tsc) - 18 cycles(tsc) - improved (cycles:-1) -5.9% 16 - 18 cycles(tsc) - 17 cycles(tsc) - improved (cycles: 1) 5.6% 30 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 32 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 34 - 78 cycles(tsc) - 23 cycles(tsc) - improved (cycles:55) 70.5% 48 - 60 cycles(tsc) - 21 cycles(tsc) - improved (cycles:39) 65.0% 64 - 49 cycles(tsc) - 20 cycles(tsc) - improved (cycles:29) 59.2% 128 - 69 cycles(tsc) - 27 cycles(tsc) - improved (cycles:42) 60.9% 158 - 79 cycles(tsc) - 30 cycles(tsc) - improved (cycles:49) 62.0% 250 - 86 cycles(tsc) - 37 cycles(tsc) - improved (cycles:49) 57.0% Performance with normal SLUB merging is significantly slower for larger bulking. This is believed to (primarily) be an effect of not having to share the per-CPU data-structures, as tuning per-CPU size can achieve similar performance. bulk - slab_nomerge - normal SLUB merge 1 - 49 cycles(tsc) - 49 cycles(tsc) - merge slower with cycles:0 2 - 30 cycles(tsc) - 30 cycles(tsc) - merge slower with cycles:0 3 - 23 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:0 4 - 20 cycles(tsc) - 20 cycles(tsc) - merge slower with cycles:0 8 - 18 cycles(tsc) - 18 cycles(tsc) - merge slower with cycles:0 16 - 17 cycles(tsc) - 17 cycles(tsc) - merge slower with cycles:0 30 - 18 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:5 32 - 18 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:4 34 - 23 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:-1 48 - 21 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:1 64 - 20 cycles(tsc) - 48 cycles(tsc) - merge slower with cycles:28 128 - 27 cycles(tsc) - 57 cycles(tsc) - merge slower with cycles:30 158 - 30 cycles(tsc) - 59 cycles(tsc) - merge slower with cycles:29 250 - 37 cycles(tsc) - 56 cycles(tsc) - merge slower with cycles:19 Joint work with Alexander Duyck. [1] https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/mm/slab_bulk_test01.c [akpm@linux-foundation.org: BUG_ON -> WARN_ON;return] Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Alexander Duyck <alexander.h.duyck@redhat.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-21 07:57:49 +08:00
if (!object)
return 0;
slub: initial bulk free implementation This implements SLUB specific kmem_cache_free_bulk(). SLUB allocator now both have bulk alloc and free implemented. Choose to reenable local IRQs while calling slowpath __slab_free(). In worst case, where all objects hit slowpath call, the performance should still be faster than fallback function __kmem_cache_free_bulk(), because local_irq_{disable+enable} is very fast (7-cycles), while the fallback invokes this_cpu_cmpxchg() which is slightly slower (9-cycles). Nitpicking, this should be faster for N>=4, due to the entry cost of local_irq_{disable+enable}. Do notice that the save+restore variant is very expensive, this is key to why this optimization works. CPU: i7-4790K CPU @ 4.00GHz * local_irq_{disable,enable}: 7 cycles(tsc) - 1.821 ns * local_irq_{save,restore} : 37 cycles(tsc) - 9.443 ns Measurements on CPU CPU i7-4790K @ 4.00GHz Baseline normal fastpath (alloc+free cost): 43 cycles(tsc) 10.834 ns Bulk- fallback - this-patch 1 - 58 cycles(tsc) 14.542 ns - 43 cycles(tsc) 10.811 ns improved 25.9% 2 - 50 cycles(tsc) 12.659 ns - 27 cycles(tsc) 6.867 ns improved 46.0% 3 - 48 cycles(tsc) 12.168 ns - 21 cycles(tsc) 5.496 ns improved 56.2% 4 - 47 cycles(tsc) 11.987 ns - 24 cycles(tsc) 6.038 ns improved 48.9% 8 - 46 cycles(tsc) 11.518 ns - 17 cycles(tsc) 4.280 ns improved 63.0% 16 - 45 cycles(tsc) 11.366 ns - 17 cycles(tsc) 4.483 ns improved 62.2% 30 - 45 cycles(tsc) 11.433 ns - 18 cycles(tsc) 4.531 ns improved 60.0% 32 - 75 cycles(tsc) 18.983 ns - 58 cycles(tsc) 14.586 ns improved 22.7% 34 - 71 cycles(tsc) 17.940 ns - 53 cycles(tsc) 13.391 ns improved 25.4% 48 - 80 cycles(tsc) 20.077 ns - 65 cycles(tsc) 16.268 ns improved 18.8% 64 - 71 cycles(tsc) 17.799 ns - 53 cycles(tsc) 13.440 ns improved 25.4% 128 - 91 cycles(tsc) 22.980 ns - 79 cycles(tsc) 19.899 ns improved 13.2% 158 - 100 cycles(tsc) 25.241 ns - 90 cycles(tsc) 22.732 ns improved 10.0% 250 - 102 cycles(tsc) 25.583 ns - 95 cycles(tsc) 23.916 ns improved 6.9% Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-05 06:45:43 +08:00
mm: new API kfree_bulk() for SLAB+SLUB allocators This patch introduce a new API call kfree_bulk() for bulk freeing memory objects not bound to a single kmem_cache. Christoph pointed out that it is possible to implement freeing of objects, without knowing the kmem_cache pointer as that information is available from the object's page->slab_cache. Proposing to remove the kmem_cache argument from the bulk free API. Jesper demonstrated that these extra steps per object comes at a performance cost. It is only in the case CONFIG_MEMCG_KMEM is compiled in and activated runtime that these steps are done anyhow. The extra cost is most visible for SLAB allocator, because the SLUB allocator does the page lookup (virt_to_head_page()) anyhow. Thus, the conclusion was to keep the kmem_cache free bulk API with a kmem_cache pointer, but we can still implement a kfree_bulk() API fairly easily. Simply by handling if kmem_cache_free_bulk() gets called with a kmem_cache NULL pointer. This does increase the code size a bit, but implementing a separate kfree_bulk() call would likely increase code size even more. Below benchmarks cost of alloc+free (obj size 256 bytes) on CPU i7-4790K @ 4.00GHz, no PREEMPT and CONFIG_MEMCG_KMEM=y. Code size increase for SLAB: add/remove: 0/0 grow/shrink: 1/0 up/down: 74/0 (74) function old new delta kmem_cache_free_bulk 660 734 +74 SLAB fastpath: 87 cycles(tsc) 21.814 sz - fallback - kmem_cache_free_bulk - kfree_bulk 1 - 103 cycles 25.878 ns - 41 cycles 10.498 ns - 81 cycles 20.312 ns 2 - 94 cycles 23.673 ns - 26 cycles 6.682 ns - 42 cycles 10.649 ns 3 - 92 cycles 23.181 ns - 21 cycles 5.325 ns - 39 cycles 9.950 ns 4 - 90 cycles 22.727 ns - 18 cycles 4.673 ns - 26 cycles 6.693 ns 8 - 89 cycles 22.270 ns - 14 cycles 3.664 ns - 23 cycles 5.835 ns 16 - 88 cycles 22.038 ns - 14 cycles 3.503 ns - 22 cycles 5.543 ns 30 - 89 cycles 22.284 ns - 13 cycles 3.310 ns - 20 cycles 5.197 ns 32 - 88 cycles 22.249 ns - 13 cycles 3.420 ns - 20 cycles 5.166 ns 34 - 88 cycles 22.224 ns - 14 cycles 3.643 ns - 20 cycles 5.170 ns 48 - 88 cycles 22.088 ns - 14 cycles 3.507 ns - 20 cycles 5.203 ns 64 - 88 cycles 22.063 ns - 13 cycles 3.428 ns - 20 cycles 5.152 ns 128 - 89 cycles 22.483 ns - 15 cycles 3.891 ns - 23 cycles 5.885 ns 158 - 89 cycles 22.381 ns - 15 cycles 3.779 ns - 22 cycles 5.548 ns 250 - 91 cycles 22.798 ns - 16 cycles 4.152 ns - 23 cycles 5.967 ns SLAB when enabling MEMCG_KMEM runtime: - kmemcg fastpath: 130 cycles(tsc) 32.684 ns (step:0) 1 - 148 cycles 37.220 ns - 66 cycles 16.622 ns - 66 cycles 16.583 ns 2 - 141 cycles 35.510 ns - 51 cycles 12.820 ns - 58 cycles 14.625 ns 3 - 140 cycles 35.017 ns - 37 cycles 9.326 ns - 33 cycles 8.474 ns 4 - 137 cycles 34.507 ns - 31 cycles 7.888 ns - 33 cycles 8.300 ns 8 - 140 cycles 35.069 ns - 25 cycles 6.461 ns - 25 cycles 6.436 ns 16 - 138 cycles 34.542 ns - 23 cycles 5.945 ns - 22 cycles 5.670 ns 30 - 136 cycles 34.227 ns - 22 cycles 5.502 ns - 22 cycles 5.587 ns 32 - 136 cycles 34.253 ns - 21 cycles 5.475 ns - 21 cycles 5.324 ns 34 - 136 cycles 34.254 ns - 21 cycles 5.448 ns - 20 cycles 5.194 ns 48 - 136 cycles 34.075 ns - 21 cycles 5.458 ns - 21 cycles 5.367 ns 64 - 135 cycles 33.994 ns - 21 cycles 5.350 ns - 21 cycles 5.259 ns 128 - 137 cycles 34.446 ns - 23 cycles 5.816 ns - 22 cycles 5.688 ns 158 - 137 cycles 34.379 ns - 22 cycles 5.727 ns - 22 cycles 5.602 ns 250 - 138 cycles 34.755 ns - 24 cycles 6.093 ns - 23 cycles 5.986 ns Code size increase for SLUB: function old new delta kmem_cache_free_bulk 717 799 +82 SLUB benchmark: SLUB fastpath: 46 cycles(tsc) 11.691 ns (step:0) sz - fallback - kmem_cache_free_bulk - kfree_bulk 1 - 61 cycles 15.486 ns - 53 cycles 13.364 ns - 57 cycles 14.464 ns 2 - 54 cycles 13.703 ns - 32 cycles 8.110 ns - 33 cycles 8.482 ns 3 - 53 cycles 13.272 ns - 25 cycles 6.362 ns - 27 cycles 6.947 ns 4 - 51 cycles 12.994 ns - 24 cycles 6.087 ns - 24 cycles 6.078 ns 8 - 50 cycles 12.576 ns - 21 cycles 5.354 ns - 22 cycles 5.513 ns 16 - 49 cycles 12.368 ns - 20 cycles 5.054 ns - 20 cycles 5.042 ns 30 - 49 cycles 12.273 ns - 18 cycles 4.748 ns - 19 cycles 4.758 ns 32 - 49 cycles 12.401 ns - 19 cycles 4.821 ns - 19 cycles 4.810 ns 34 - 98 cycles 24.519 ns - 24 cycles 6.154 ns - 24 cycles 6.157 ns 48 - 83 cycles 20.833 ns - 21 cycles 5.446 ns - 21 cycles 5.429 ns 64 - 75 cycles 18.891 ns - 20 cycles 5.247 ns - 20 cycles 5.238 ns 128 - 93 cycles 23.271 ns - 27 cycles 6.856 ns - 27 cycles 6.823 ns 158 - 102 cycles 25.581 ns - 30 cycles 7.714 ns - 30 cycles 7.695 ns 250 - 107 cycles 26.917 ns - 38 cycles 9.514 ns - 38 cycles 9.506 ns SLUB when enabling MEMCG_KMEM runtime: - kmemcg fastpath: 71 cycles(tsc) 17.897 ns (step:0) 1 - 85 cycles 21.484 ns - 78 cycles 19.569 ns - 75 cycles 18.938 ns 2 - 81 cycles 20.363 ns - 45 cycles 11.258 ns - 44 cycles 11.076 ns 3 - 78 cycles 19.709 ns - 33 cycles 8.354 ns - 32 cycles 8.044 ns 4 - 77 cycles 19.430 ns - 28 cycles 7.216 ns - 28 cycles 7.003 ns 8 - 101 cycles 25.288 ns - 23 cycles 5.849 ns - 23 cycles 5.787 ns 16 - 76 cycles 19.148 ns - 20 cycles 5.162 ns - 20 cycles 5.081 ns 30 - 76 cycles 19.067 ns - 19 cycles 4.868 ns - 19 cycles 4.821 ns 32 - 76 cycles 19.052 ns - 19 cycles 4.857 ns - 19 cycles 4.815 ns 34 - 121 cycles 30.291 ns - 25 cycles 6.333 ns - 25 cycles 6.268 ns 48 - 108 cycles 27.111 ns - 21 cycles 5.498 ns - 21 cycles 5.458 ns 64 - 100 cycles 25.164 ns - 20 cycles 5.242 ns - 20 cycles 5.229 ns 128 - 155 cycles 38.976 ns - 27 cycles 6.886 ns - 27 cycles 6.892 ns 158 - 132 cycles 33.034 ns - 30 cycles 7.711 ns - 30 cycles 7.728 ns 250 - 130 cycles 32.612 ns - 38 cycles 9.560 ns - 38 cycles 9.549 ns Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Vladimir Davydov <vdavydov@virtuozzo.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-03-16 05:54:00 +08:00
page = virt_to_head_page(object);
if (!s) {
/* Handle kalloc'ed objects */
if (unlikely(!PageSlab(page))) {
BUG_ON(!PageCompound(page));
kfree_hook(object);
mm: charge/uncharge kmemcg from generic page allocator paths Currently, to charge a non-slab allocation to kmemcg one has to use alloc_kmem_pages helper with __GFP_ACCOUNT flag. A page allocated with this helper should finally be freed using free_kmem_pages, otherwise it won't be uncharged. This API suits its current users fine, but it turns out to be impossible to use along with page reference counting, i.e. when an allocation is supposed to be freed with put_page, as it is the case with pipe or unix socket buffers. To overcome this limitation, this patch moves charging/uncharging to generic page allocator paths, i.e. to __alloc_pages_nodemask and free_pages_prepare, and zaps alloc/free_kmem_pages helpers. This way, one can use any of the available page allocation functions to get the allocated page charged to kmemcg - it's enough to pass __GFP_ACCOUNT, just like in case of kmalloc and friends. A charged page will be automatically uncharged on free. To make it possible, we need to mark pages charged to kmemcg somehow. To avoid introducing a new page flag, we make use of page->_mapcount for marking such pages. Since pages charged to kmemcg are not supposed to be mapped to userspace, it should work just fine. There are other (ab)users of page->_mapcount - buddy and balloon pages - but we don't conflict with them. In case kmemcg is compiled out or not used at runtime, this patch introduces no overhead to generic page allocator paths. If kmemcg is used, it will be plus one gfp flags check on alloc and plus one page->_mapcount check on free, which shouldn't hurt performance, because the data accessed are hot. Link: http://lkml.kernel.org/r/a9736d856f895bcb465d9f257b54efe32eda6f99.1464079538.git.vdavydov@virtuozzo.com Signed-off-by: Vladimir Davydov <vdavydov@virtuozzo.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Eric Dumazet <eric.dumazet@gmail.com> Cc: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-07-27 06:24:24 +08:00
__free_pages(page, compound_order(page));
mm: new API kfree_bulk() for SLAB+SLUB allocators This patch introduce a new API call kfree_bulk() for bulk freeing memory objects not bound to a single kmem_cache. Christoph pointed out that it is possible to implement freeing of objects, without knowing the kmem_cache pointer as that information is available from the object's page->slab_cache. Proposing to remove the kmem_cache argument from the bulk free API. Jesper demonstrated that these extra steps per object comes at a performance cost. It is only in the case CONFIG_MEMCG_KMEM is compiled in and activated runtime that these steps are done anyhow. The extra cost is most visible for SLAB allocator, because the SLUB allocator does the page lookup (virt_to_head_page()) anyhow. Thus, the conclusion was to keep the kmem_cache free bulk API with a kmem_cache pointer, but we can still implement a kfree_bulk() API fairly easily. Simply by handling if kmem_cache_free_bulk() gets called with a kmem_cache NULL pointer. This does increase the code size a bit, but implementing a separate kfree_bulk() call would likely increase code size even more. Below benchmarks cost of alloc+free (obj size 256 bytes) on CPU i7-4790K @ 4.00GHz, no PREEMPT and CONFIG_MEMCG_KMEM=y. Code size increase for SLAB: add/remove: 0/0 grow/shrink: 1/0 up/down: 74/0 (74) function old new delta kmem_cache_free_bulk 660 734 +74 SLAB fastpath: 87 cycles(tsc) 21.814 sz - fallback - kmem_cache_free_bulk - kfree_bulk 1 - 103 cycles 25.878 ns - 41 cycles 10.498 ns - 81 cycles 20.312 ns 2 - 94 cycles 23.673 ns - 26 cycles 6.682 ns - 42 cycles 10.649 ns 3 - 92 cycles 23.181 ns - 21 cycles 5.325 ns - 39 cycles 9.950 ns 4 - 90 cycles 22.727 ns - 18 cycles 4.673 ns - 26 cycles 6.693 ns 8 - 89 cycles 22.270 ns - 14 cycles 3.664 ns - 23 cycles 5.835 ns 16 - 88 cycles 22.038 ns - 14 cycles 3.503 ns - 22 cycles 5.543 ns 30 - 89 cycles 22.284 ns - 13 cycles 3.310 ns - 20 cycles 5.197 ns 32 - 88 cycles 22.249 ns - 13 cycles 3.420 ns - 20 cycles 5.166 ns 34 - 88 cycles 22.224 ns - 14 cycles 3.643 ns - 20 cycles 5.170 ns 48 - 88 cycles 22.088 ns - 14 cycles 3.507 ns - 20 cycles 5.203 ns 64 - 88 cycles 22.063 ns - 13 cycles 3.428 ns - 20 cycles 5.152 ns 128 - 89 cycles 22.483 ns - 15 cycles 3.891 ns - 23 cycles 5.885 ns 158 - 89 cycles 22.381 ns - 15 cycles 3.779 ns - 22 cycles 5.548 ns 250 - 91 cycles 22.798 ns - 16 cycles 4.152 ns - 23 cycles 5.967 ns SLAB when enabling MEMCG_KMEM runtime: - kmemcg fastpath: 130 cycles(tsc) 32.684 ns (step:0) 1 - 148 cycles 37.220 ns - 66 cycles 16.622 ns - 66 cycles 16.583 ns 2 - 141 cycles 35.510 ns - 51 cycles 12.820 ns - 58 cycles 14.625 ns 3 - 140 cycles 35.017 ns - 37 cycles 9.326 ns - 33 cycles 8.474 ns 4 - 137 cycles 34.507 ns - 31 cycles 7.888 ns - 33 cycles 8.300 ns 8 - 140 cycles 35.069 ns - 25 cycles 6.461 ns - 25 cycles 6.436 ns 16 - 138 cycles 34.542 ns - 23 cycles 5.945 ns - 22 cycles 5.670 ns 30 - 136 cycles 34.227 ns - 22 cycles 5.502 ns - 22 cycles 5.587 ns 32 - 136 cycles 34.253 ns - 21 cycles 5.475 ns - 21 cycles 5.324 ns 34 - 136 cycles 34.254 ns - 21 cycles 5.448 ns - 20 cycles 5.194 ns 48 - 136 cycles 34.075 ns - 21 cycles 5.458 ns - 21 cycles 5.367 ns 64 - 135 cycles 33.994 ns - 21 cycles 5.350 ns - 21 cycles 5.259 ns 128 - 137 cycles 34.446 ns - 23 cycles 5.816 ns - 22 cycles 5.688 ns 158 - 137 cycles 34.379 ns - 22 cycles 5.727 ns - 22 cycles 5.602 ns 250 - 138 cycles 34.755 ns - 24 cycles 6.093 ns - 23 cycles 5.986 ns Code size increase for SLUB: function old new delta kmem_cache_free_bulk 717 799 +82 SLUB benchmark: SLUB fastpath: 46 cycles(tsc) 11.691 ns (step:0) sz - fallback - kmem_cache_free_bulk - kfree_bulk 1 - 61 cycles 15.486 ns - 53 cycles 13.364 ns - 57 cycles 14.464 ns 2 - 54 cycles 13.703 ns - 32 cycles 8.110 ns - 33 cycles 8.482 ns 3 - 53 cycles 13.272 ns - 25 cycles 6.362 ns - 27 cycles 6.947 ns 4 - 51 cycles 12.994 ns - 24 cycles 6.087 ns - 24 cycles 6.078 ns 8 - 50 cycles 12.576 ns - 21 cycles 5.354 ns - 22 cycles 5.513 ns 16 - 49 cycles 12.368 ns - 20 cycles 5.054 ns - 20 cycles 5.042 ns 30 - 49 cycles 12.273 ns - 18 cycles 4.748 ns - 19 cycles 4.758 ns 32 - 49 cycles 12.401 ns - 19 cycles 4.821 ns - 19 cycles 4.810 ns 34 - 98 cycles 24.519 ns - 24 cycles 6.154 ns - 24 cycles 6.157 ns 48 - 83 cycles 20.833 ns - 21 cycles 5.446 ns - 21 cycles 5.429 ns 64 - 75 cycles 18.891 ns - 20 cycles 5.247 ns - 20 cycles 5.238 ns 128 - 93 cycles 23.271 ns - 27 cycles 6.856 ns - 27 cycles 6.823 ns 158 - 102 cycles 25.581 ns - 30 cycles 7.714 ns - 30 cycles 7.695 ns 250 - 107 cycles 26.917 ns - 38 cycles 9.514 ns - 38 cycles 9.506 ns SLUB when enabling MEMCG_KMEM runtime: - kmemcg fastpath: 71 cycles(tsc) 17.897 ns (step:0) 1 - 85 cycles 21.484 ns - 78 cycles 19.569 ns - 75 cycles 18.938 ns 2 - 81 cycles 20.363 ns - 45 cycles 11.258 ns - 44 cycles 11.076 ns 3 - 78 cycles 19.709 ns - 33 cycles 8.354 ns - 32 cycles 8.044 ns 4 - 77 cycles 19.430 ns - 28 cycles 7.216 ns - 28 cycles 7.003 ns 8 - 101 cycles 25.288 ns - 23 cycles 5.849 ns - 23 cycles 5.787 ns 16 - 76 cycles 19.148 ns - 20 cycles 5.162 ns - 20 cycles 5.081 ns 30 - 76 cycles 19.067 ns - 19 cycles 4.868 ns - 19 cycles 4.821 ns 32 - 76 cycles 19.052 ns - 19 cycles 4.857 ns - 19 cycles 4.815 ns 34 - 121 cycles 30.291 ns - 25 cycles 6.333 ns - 25 cycles 6.268 ns 48 - 108 cycles 27.111 ns - 21 cycles 5.498 ns - 21 cycles 5.458 ns 64 - 100 cycles 25.164 ns - 20 cycles 5.242 ns - 20 cycles 5.229 ns 128 - 155 cycles 38.976 ns - 27 cycles 6.886 ns - 27 cycles 6.892 ns 158 - 132 cycles 33.034 ns - 30 cycles 7.711 ns - 30 cycles 7.728 ns 250 - 130 cycles 32.612 ns - 38 cycles 9.560 ns - 38 cycles 9.549 ns Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Vladimir Davydov <vdavydov@virtuozzo.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-03-16 05:54:00 +08:00
p[size] = NULL; /* mark object processed */
return size;
}
/* Derive kmem_cache from object */
df->s = page->slab_cache;
} else {
df->s = cache_from_obj(s, object); /* Support for memcg */
}
mm, kfence: insert KFENCE hooks for SLUB Inserts KFENCE hooks into the SLUB allocator. To pass the originally requested size to KFENCE, add an argument 'orig_size' to slab_alloc*(). The additional argument is required to preserve the requested original size for kmalloc() allocations, which uses size classes (e.g. an allocation of 272 bytes will return an object of size 512). Therefore, kmem_cache::size does not represent the kmalloc-caller's requested size, and we must introduce the argument 'orig_size' to propagate the originally requested size to KFENCE. Without the originally requested size, we would not be able to detect out-of-bounds accesses for objects placed at the end of a KFENCE object page if that object is not equal to the kmalloc-size class it was bucketed into. When KFENCE is disabled, there is no additional overhead, since slab_alloc*() functions are __always_inline. Link: https://lkml.kernel.org/r/20201103175841.3495947-6-elver@google.com Signed-off-by: Marco Elver <elver@google.com> Signed-off-by: Alexander Potapenko <glider@google.com> Reviewed-by: Dmitry Vyukov <dvyukov@google.com> Reviewed-by: Jann Horn <jannh@google.com> Co-developed-by: Marco Elver <elver@google.com> Cc: Andrey Konovalov <andreyknvl@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christopher Lameter <cl@linux.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: David Rientjes <rientjes@google.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Hillf Danton <hdanton@sina.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Joern Engel <joern@purestorage.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kees Cook <keescook@chromium.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Paul E. McKenney <paulmck@kernel.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: SeongJae Park <sjpark@amazon.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-26 09:19:16 +08:00
if (is_kfence_address(object)) {
kasan, mm: integrate slab init_on_free with HW_TAGS This change uses the previously added memory initialization feature of HW_TAGS KASAN routines for slab memory when init_on_free is enabled. With this change, memory initialization memset() is no longer called when both HW_TAGS KASAN and init_on_free are enabled. Instead, memory is initialized in KASAN runtime. For SLUB, the memory initialization memset() is moved into slab_free_hook() that currently directly follows the initialization loop. A new argument is added to slab_free_hook() that indicates whether to initialize the memory or not. To avoid discrepancies with which memory gets initialized that can be caused by future changes, both KASAN hook and initialization memset() are put together and a warning comment is added. Combining setting allocation tags with memory initialization improves HW_TAGS KASAN performance when init_on_free is enabled. Link: https://lkml.kernel.org/r/190fd15c1886654afdec0d19ebebd5ade665b601.1615296150.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Reviewed-by: Marco Elver <elver@google.com> Cc: Alexander Potapenko <glider@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Branislav Rankov <Branislav.Rankov@arm.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christoph Lameter <cl@linux.com> Cc: David Rientjes <rientjes@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Evgenii Stepanov <eugenis@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kevin Brodsky <kevin.brodsky@arm.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Collingbourne <pcc@google.com> Cc: Vincenzo Frascino <vincenzo.frascino@arm.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-30 14:00:09 +08:00
slab_free_hook(df->s, object, false);
mm, kfence: insert KFENCE hooks for SLUB Inserts KFENCE hooks into the SLUB allocator. To pass the originally requested size to KFENCE, add an argument 'orig_size' to slab_alloc*(). The additional argument is required to preserve the requested original size for kmalloc() allocations, which uses size classes (e.g. an allocation of 272 bytes will return an object of size 512). Therefore, kmem_cache::size does not represent the kmalloc-caller's requested size, and we must introduce the argument 'orig_size' to propagate the originally requested size to KFENCE. Without the originally requested size, we would not be able to detect out-of-bounds accesses for objects placed at the end of a KFENCE object page if that object is not equal to the kmalloc-size class it was bucketed into. When KFENCE is disabled, there is no additional overhead, since slab_alloc*() functions are __always_inline. Link: https://lkml.kernel.org/r/20201103175841.3495947-6-elver@google.com Signed-off-by: Marco Elver <elver@google.com> Signed-off-by: Alexander Potapenko <glider@google.com> Reviewed-by: Dmitry Vyukov <dvyukov@google.com> Reviewed-by: Jann Horn <jannh@google.com> Co-developed-by: Marco Elver <elver@google.com> Cc: Andrey Konovalov <andreyknvl@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christopher Lameter <cl@linux.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: David Rientjes <rientjes@google.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Hillf Danton <hdanton@sina.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Joern Engel <joern@purestorage.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kees Cook <keescook@chromium.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Paul E. McKenney <paulmck@kernel.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: SeongJae Park <sjpark@amazon.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-26 09:19:16 +08:00
__kfence_free(object);
p[size] = NULL; /* mark object processed */
return size;
}
slub: optimize bulk slowpath free by detached freelist This change focus on improving the speed of object freeing in the "slowpath" of kmem_cache_free_bulk. The calls slab_free (fastpath) and __slab_free (slowpath) have been extended with support for bulk free, which amortize the overhead of the (locked) cmpxchg_double. To use the new bulking feature, we build what I call a detached freelist. The detached freelist takes advantage of three properties: 1) the free function call owns the object that is about to be freed, thus writing into this memory is synchronization-free. 2) many freelist's can co-exist side-by-side in the same slab-page each with a separate head pointer. 3) it is the visibility of the head pointer that needs synchronization. Given these properties, the brilliant part is that the detached freelist can be constructed without any need for synchronization. The freelist is constructed directly in the page objects, without any synchronization needed. The detached freelist is allocated on the stack of the function call kmem_cache_free_bulk. Thus, the freelist head pointer is not visible to other CPUs. All objects in a SLUB freelist must belong to the same slab-page. Thus, constructing the detached freelist is about matching objects that belong to the same slab-page. The bulk free array is scanned is a progressive manor with a limited look-ahead facility. Kmem debug support is handled in call of slab_free(). Notice kmem_cache_free_bulk no longer need to disable IRQs. This only slowed down single free bulk with approx 3 cycles. Performance data: Benchmarked[1] obj size 256 bytes on CPU i7-4790K @ 4.00GHz SLUB fastpath single object quick reuse: 47 cycles(tsc) 11.931 ns To get stable and comparable numbers, the kernel have been booted with "slab_merge" (this also improve performance for larger bulk sizes). Performance data, compared against fallback bulking: bulk - fallback bulk - improvement with this patch 1 - 62 cycles(tsc) 15.662 ns - 49 cycles(tsc) 12.407 ns- improved 21.0% 2 - 55 cycles(tsc) 13.935 ns - 30 cycles(tsc) 7.506 ns - improved 45.5% 3 - 53 cycles(tsc) 13.341 ns - 23 cycles(tsc) 5.865 ns - improved 56.6% 4 - 52 cycles(tsc) 13.081 ns - 20 cycles(tsc) 5.048 ns - improved 61.5% 8 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.659 ns - improved 64.0% 16 - 49 cycles(tsc) 12.412 ns - 17 cycles(tsc) 4.495 ns - improved 65.3% 30 - 49 cycles(tsc) 12.484 ns - 18 cycles(tsc) 4.533 ns - improved 63.3% 32 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.707 ns - improved 64.0% 34 - 96 cycles(tsc) 24.243 ns - 23 cycles(tsc) 5.976 ns - improved 76.0% 48 - 83 cycles(tsc) 20.818 ns - 21 cycles(tsc) 5.329 ns - improved 74.7% 64 - 74 cycles(tsc) 18.700 ns - 20 cycles(tsc) 5.127 ns - improved 73.0% 128 - 90 cycles(tsc) 22.734 ns - 27 cycles(tsc) 6.833 ns - improved 70.0% 158 - 99 cycles(tsc) 24.776 ns - 30 cycles(tsc) 7.583 ns - improved 69.7% 250 - 104 cycles(tsc) 26.089 ns - 37 cycles(tsc) 9.280 ns - improved 64.4% Performance data, compared current in-kernel bulking: bulk - curr in-kernel - improvement with this patch 1 - 46 cycles(tsc) - 49 cycles(tsc) - improved (cycles:-3) -6.5% 2 - 27 cycles(tsc) - 30 cycles(tsc) - improved (cycles:-3) -11.1% 3 - 21 cycles(tsc) - 23 cycles(tsc) - improved (cycles:-2) -9.5% 4 - 18 cycles(tsc) - 20 cycles(tsc) - improved (cycles:-2) -11.1% 8 - 17 cycles(tsc) - 18 cycles(tsc) - improved (cycles:-1) -5.9% 16 - 18 cycles(tsc) - 17 cycles(tsc) - improved (cycles: 1) 5.6% 30 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 32 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 34 - 78 cycles(tsc) - 23 cycles(tsc) - improved (cycles:55) 70.5% 48 - 60 cycles(tsc) - 21 cycles(tsc) - improved (cycles:39) 65.0% 64 - 49 cycles(tsc) - 20 cycles(tsc) - improved (cycles:29) 59.2% 128 - 69 cycles(tsc) - 27 cycles(tsc) - improved (cycles:42) 60.9% 158 - 79 cycles(tsc) - 30 cycles(tsc) - improved (cycles:49) 62.0% 250 - 86 cycles(tsc) - 37 cycles(tsc) - improved (cycles:49) 57.0% Performance with normal SLUB merging is significantly slower for larger bulking. This is believed to (primarily) be an effect of not having to share the per-CPU data-structures, as tuning per-CPU size can achieve similar performance. bulk - slab_nomerge - normal SLUB merge 1 - 49 cycles(tsc) - 49 cycles(tsc) - merge slower with cycles:0 2 - 30 cycles(tsc) - 30 cycles(tsc) - merge slower with cycles:0 3 - 23 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:0 4 - 20 cycles(tsc) - 20 cycles(tsc) - merge slower with cycles:0 8 - 18 cycles(tsc) - 18 cycles(tsc) - merge slower with cycles:0 16 - 17 cycles(tsc) - 17 cycles(tsc) - merge slower with cycles:0 30 - 18 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:5 32 - 18 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:4 34 - 23 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:-1 48 - 21 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:1 64 - 20 cycles(tsc) - 48 cycles(tsc) - merge slower with cycles:28 128 - 27 cycles(tsc) - 57 cycles(tsc) - merge slower with cycles:30 158 - 30 cycles(tsc) - 59 cycles(tsc) - merge slower with cycles:29 250 - 37 cycles(tsc) - 56 cycles(tsc) - merge slower with cycles:19 Joint work with Alexander Duyck. [1] https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/mm/slab_bulk_test01.c [akpm@linux-foundation.org: BUG_ON -> WARN_ON;return] Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Alexander Duyck <alexander.h.duyck@redhat.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-21 07:57:49 +08:00
/* Start new detached freelist */
mm: new API kfree_bulk() for SLAB+SLUB allocators This patch introduce a new API call kfree_bulk() for bulk freeing memory objects not bound to a single kmem_cache. Christoph pointed out that it is possible to implement freeing of objects, without knowing the kmem_cache pointer as that information is available from the object's page->slab_cache. Proposing to remove the kmem_cache argument from the bulk free API. Jesper demonstrated that these extra steps per object comes at a performance cost. It is only in the case CONFIG_MEMCG_KMEM is compiled in and activated runtime that these steps are done anyhow. The extra cost is most visible for SLAB allocator, because the SLUB allocator does the page lookup (virt_to_head_page()) anyhow. Thus, the conclusion was to keep the kmem_cache free bulk API with a kmem_cache pointer, but we can still implement a kfree_bulk() API fairly easily. Simply by handling if kmem_cache_free_bulk() gets called with a kmem_cache NULL pointer. This does increase the code size a bit, but implementing a separate kfree_bulk() call would likely increase code size even more. Below benchmarks cost of alloc+free (obj size 256 bytes) on CPU i7-4790K @ 4.00GHz, no PREEMPT and CONFIG_MEMCG_KMEM=y. Code size increase for SLAB: add/remove: 0/0 grow/shrink: 1/0 up/down: 74/0 (74) function old new delta kmem_cache_free_bulk 660 734 +74 SLAB fastpath: 87 cycles(tsc) 21.814 sz - fallback - kmem_cache_free_bulk - kfree_bulk 1 - 103 cycles 25.878 ns - 41 cycles 10.498 ns - 81 cycles 20.312 ns 2 - 94 cycles 23.673 ns - 26 cycles 6.682 ns - 42 cycles 10.649 ns 3 - 92 cycles 23.181 ns - 21 cycles 5.325 ns - 39 cycles 9.950 ns 4 - 90 cycles 22.727 ns - 18 cycles 4.673 ns - 26 cycles 6.693 ns 8 - 89 cycles 22.270 ns - 14 cycles 3.664 ns - 23 cycles 5.835 ns 16 - 88 cycles 22.038 ns - 14 cycles 3.503 ns - 22 cycles 5.543 ns 30 - 89 cycles 22.284 ns - 13 cycles 3.310 ns - 20 cycles 5.197 ns 32 - 88 cycles 22.249 ns - 13 cycles 3.420 ns - 20 cycles 5.166 ns 34 - 88 cycles 22.224 ns - 14 cycles 3.643 ns - 20 cycles 5.170 ns 48 - 88 cycles 22.088 ns - 14 cycles 3.507 ns - 20 cycles 5.203 ns 64 - 88 cycles 22.063 ns - 13 cycles 3.428 ns - 20 cycles 5.152 ns 128 - 89 cycles 22.483 ns - 15 cycles 3.891 ns - 23 cycles 5.885 ns 158 - 89 cycles 22.381 ns - 15 cycles 3.779 ns - 22 cycles 5.548 ns 250 - 91 cycles 22.798 ns - 16 cycles 4.152 ns - 23 cycles 5.967 ns SLAB when enabling MEMCG_KMEM runtime: - kmemcg fastpath: 130 cycles(tsc) 32.684 ns (step:0) 1 - 148 cycles 37.220 ns - 66 cycles 16.622 ns - 66 cycles 16.583 ns 2 - 141 cycles 35.510 ns - 51 cycles 12.820 ns - 58 cycles 14.625 ns 3 - 140 cycles 35.017 ns - 37 cycles 9.326 ns - 33 cycles 8.474 ns 4 - 137 cycles 34.507 ns - 31 cycles 7.888 ns - 33 cycles 8.300 ns 8 - 140 cycles 35.069 ns - 25 cycles 6.461 ns - 25 cycles 6.436 ns 16 - 138 cycles 34.542 ns - 23 cycles 5.945 ns - 22 cycles 5.670 ns 30 - 136 cycles 34.227 ns - 22 cycles 5.502 ns - 22 cycles 5.587 ns 32 - 136 cycles 34.253 ns - 21 cycles 5.475 ns - 21 cycles 5.324 ns 34 - 136 cycles 34.254 ns - 21 cycles 5.448 ns - 20 cycles 5.194 ns 48 - 136 cycles 34.075 ns - 21 cycles 5.458 ns - 21 cycles 5.367 ns 64 - 135 cycles 33.994 ns - 21 cycles 5.350 ns - 21 cycles 5.259 ns 128 - 137 cycles 34.446 ns - 23 cycles 5.816 ns - 22 cycles 5.688 ns 158 - 137 cycles 34.379 ns - 22 cycles 5.727 ns - 22 cycles 5.602 ns 250 - 138 cycles 34.755 ns - 24 cycles 6.093 ns - 23 cycles 5.986 ns Code size increase for SLUB: function old new delta kmem_cache_free_bulk 717 799 +82 SLUB benchmark: SLUB fastpath: 46 cycles(tsc) 11.691 ns (step:0) sz - fallback - kmem_cache_free_bulk - kfree_bulk 1 - 61 cycles 15.486 ns - 53 cycles 13.364 ns - 57 cycles 14.464 ns 2 - 54 cycles 13.703 ns - 32 cycles 8.110 ns - 33 cycles 8.482 ns 3 - 53 cycles 13.272 ns - 25 cycles 6.362 ns - 27 cycles 6.947 ns 4 - 51 cycles 12.994 ns - 24 cycles 6.087 ns - 24 cycles 6.078 ns 8 - 50 cycles 12.576 ns - 21 cycles 5.354 ns - 22 cycles 5.513 ns 16 - 49 cycles 12.368 ns - 20 cycles 5.054 ns - 20 cycles 5.042 ns 30 - 49 cycles 12.273 ns - 18 cycles 4.748 ns - 19 cycles 4.758 ns 32 - 49 cycles 12.401 ns - 19 cycles 4.821 ns - 19 cycles 4.810 ns 34 - 98 cycles 24.519 ns - 24 cycles 6.154 ns - 24 cycles 6.157 ns 48 - 83 cycles 20.833 ns - 21 cycles 5.446 ns - 21 cycles 5.429 ns 64 - 75 cycles 18.891 ns - 20 cycles 5.247 ns - 20 cycles 5.238 ns 128 - 93 cycles 23.271 ns - 27 cycles 6.856 ns - 27 cycles 6.823 ns 158 - 102 cycles 25.581 ns - 30 cycles 7.714 ns - 30 cycles 7.695 ns 250 - 107 cycles 26.917 ns - 38 cycles 9.514 ns - 38 cycles 9.506 ns SLUB when enabling MEMCG_KMEM runtime: - kmemcg fastpath: 71 cycles(tsc) 17.897 ns (step:0) 1 - 85 cycles 21.484 ns - 78 cycles 19.569 ns - 75 cycles 18.938 ns 2 - 81 cycles 20.363 ns - 45 cycles 11.258 ns - 44 cycles 11.076 ns 3 - 78 cycles 19.709 ns - 33 cycles 8.354 ns - 32 cycles 8.044 ns 4 - 77 cycles 19.430 ns - 28 cycles 7.216 ns - 28 cycles 7.003 ns 8 - 101 cycles 25.288 ns - 23 cycles 5.849 ns - 23 cycles 5.787 ns 16 - 76 cycles 19.148 ns - 20 cycles 5.162 ns - 20 cycles 5.081 ns 30 - 76 cycles 19.067 ns - 19 cycles 4.868 ns - 19 cycles 4.821 ns 32 - 76 cycles 19.052 ns - 19 cycles 4.857 ns - 19 cycles 4.815 ns 34 - 121 cycles 30.291 ns - 25 cycles 6.333 ns - 25 cycles 6.268 ns 48 - 108 cycles 27.111 ns - 21 cycles 5.498 ns - 21 cycles 5.458 ns 64 - 100 cycles 25.164 ns - 20 cycles 5.242 ns - 20 cycles 5.229 ns 128 - 155 cycles 38.976 ns - 27 cycles 6.886 ns - 27 cycles 6.892 ns 158 - 132 cycles 33.034 ns - 30 cycles 7.711 ns - 30 cycles 7.728 ns 250 - 130 cycles 32.612 ns - 38 cycles 9.560 ns - 38 cycles 9.549 ns Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Vladimir Davydov <vdavydov@virtuozzo.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-03-16 05:54:00 +08:00
df->page = page;
set_freepointer(df->s, object, NULL);
slub: optimize bulk slowpath free by detached freelist This change focus on improving the speed of object freeing in the "slowpath" of kmem_cache_free_bulk. The calls slab_free (fastpath) and __slab_free (slowpath) have been extended with support for bulk free, which amortize the overhead of the (locked) cmpxchg_double. To use the new bulking feature, we build what I call a detached freelist. The detached freelist takes advantage of three properties: 1) the free function call owns the object that is about to be freed, thus writing into this memory is synchronization-free. 2) many freelist's can co-exist side-by-side in the same slab-page each with a separate head pointer. 3) it is the visibility of the head pointer that needs synchronization. Given these properties, the brilliant part is that the detached freelist can be constructed without any need for synchronization. The freelist is constructed directly in the page objects, without any synchronization needed. The detached freelist is allocated on the stack of the function call kmem_cache_free_bulk. Thus, the freelist head pointer is not visible to other CPUs. All objects in a SLUB freelist must belong to the same slab-page. Thus, constructing the detached freelist is about matching objects that belong to the same slab-page. The bulk free array is scanned is a progressive manor with a limited look-ahead facility. Kmem debug support is handled in call of slab_free(). Notice kmem_cache_free_bulk no longer need to disable IRQs. This only slowed down single free bulk with approx 3 cycles. Performance data: Benchmarked[1] obj size 256 bytes on CPU i7-4790K @ 4.00GHz SLUB fastpath single object quick reuse: 47 cycles(tsc) 11.931 ns To get stable and comparable numbers, the kernel have been booted with "slab_merge" (this also improve performance for larger bulk sizes). Performance data, compared against fallback bulking: bulk - fallback bulk - improvement with this patch 1 - 62 cycles(tsc) 15.662 ns - 49 cycles(tsc) 12.407 ns- improved 21.0% 2 - 55 cycles(tsc) 13.935 ns - 30 cycles(tsc) 7.506 ns - improved 45.5% 3 - 53 cycles(tsc) 13.341 ns - 23 cycles(tsc) 5.865 ns - improved 56.6% 4 - 52 cycles(tsc) 13.081 ns - 20 cycles(tsc) 5.048 ns - improved 61.5% 8 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.659 ns - improved 64.0% 16 - 49 cycles(tsc) 12.412 ns - 17 cycles(tsc) 4.495 ns - improved 65.3% 30 - 49 cycles(tsc) 12.484 ns - 18 cycles(tsc) 4.533 ns - improved 63.3% 32 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.707 ns - improved 64.0% 34 - 96 cycles(tsc) 24.243 ns - 23 cycles(tsc) 5.976 ns - improved 76.0% 48 - 83 cycles(tsc) 20.818 ns - 21 cycles(tsc) 5.329 ns - improved 74.7% 64 - 74 cycles(tsc) 18.700 ns - 20 cycles(tsc) 5.127 ns - improved 73.0% 128 - 90 cycles(tsc) 22.734 ns - 27 cycles(tsc) 6.833 ns - improved 70.0% 158 - 99 cycles(tsc) 24.776 ns - 30 cycles(tsc) 7.583 ns - improved 69.7% 250 - 104 cycles(tsc) 26.089 ns - 37 cycles(tsc) 9.280 ns - improved 64.4% Performance data, compared current in-kernel bulking: bulk - curr in-kernel - improvement with this patch 1 - 46 cycles(tsc) - 49 cycles(tsc) - improved (cycles:-3) -6.5% 2 - 27 cycles(tsc) - 30 cycles(tsc) - improved (cycles:-3) -11.1% 3 - 21 cycles(tsc) - 23 cycles(tsc) - improved (cycles:-2) -9.5% 4 - 18 cycles(tsc) - 20 cycles(tsc) - improved (cycles:-2) -11.1% 8 - 17 cycles(tsc) - 18 cycles(tsc) - improved (cycles:-1) -5.9% 16 - 18 cycles(tsc) - 17 cycles(tsc) - improved (cycles: 1) 5.6% 30 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 32 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 34 - 78 cycles(tsc) - 23 cycles(tsc) - improved (cycles:55) 70.5% 48 - 60 cycles(tsc) - 21 cycles(tsc) - improved (cycles:39) 65.0% 64 - 49 cycles(tsc) - 20 cycles(tsc) - improved (cycles:29) 59.2% 128 - 69 cycles(tsc) - 27 cycles(tsc) - improved (cycles:42) 60.9% 158 - 79 cycles(tsc) - 30 cycles(tsc) - improved (cycles:49) 62.0% 250 - 86 cycles(tsc) - 37 cycles(tsc) - improved (cycles:49) 57.0% Performance with normal SLUB merging is significantly slower for larger bulking. This is believed to (primarily) be an effect of not having to share the per-CPU data-structures, as tuning per-CPU size can achieve similar performance. bulk - slab_nomerge - normal SLUB merge 1 - 49 cycles(tsc) - 49 cycles(tsc) - merge slower with cycles:0 2 - 30 cycles(tsc) - 30 cycles(tsc) - merge slower with cycles:0 3 - 23 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:0 4 - 20 cycles(tsc) - 20 cycles(tsc) - merge slower with cycles:0 8 - 18 cycles(tsc) - 18 cycles(tsc) - merge slower with cycles:0 16 - 17 cycles(tsc) - 17 cycles(tsc) - merge slower with cycles:0 30 - 18 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:5 32 - 18 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:4 34 - 23 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:-1 48 - 21 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:1 64 - 20 cycles(tsc) - 48 cycles(tsc) - merge slower with cycles:28 128 - 27 cycles(tsc) - 57 cycles(tsc) - merge slower with cycles:30 158 - 30 cycles(tsc) - 59 cycles(tsc) - merge slower with cycles:29 250 - 37 cycles(tsc) - 56 cycles(tsc) - merge slower with cycles:19 Joint work with Alexander Duyck. [1] https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/mm/slab_bulk_test01.c [akpm@linux-foundation.org: BUG_ON -> WARN_ON;return] Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Alexander Duyck <alexander.h.duyck@redhat.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-21 07:57:49 +08:00
df->tail = object;
df->freelist = object;
p[size] = NULL; /* mark object processed */
df->cnt = 1;
while (size) {
object = p[--size];
if (!object)
continue; /* Skip processed objects */
/* df->page is always set at this point */
if (df->page == virt_to_head_page(object)) {
/* Opportunity build freelist */
set_freepointer(df->s, object, df->freelist);
slub: optimize bulk slowpath free by detached freelist This change focus on improving the speed of object freeing in the "slowpath" of kmem_cache_free_bulk. The calls slab_free (fastpath) and __slab_free (slowpath) have been extended with support for bulk free, which amortize the overhead of the (locked) cmpxchg_double. To use the new bulking feature, we build what I call a detached freelist. The detached freelist takes advantage of three properties: 1) the free function call owns the object that is about to be freed, thus writing into this memory is synchronization-free. 2) many freelist's can co-exist side-by-side in the same slab-page each with a separate head pointer. 3) it is the visibility of the head pointer that needs synchronization. Given these properties, the brilliant part is that the detached freelist can be constructed without any need for synchronization. The freelist is constructed directly in the page objects, without any synchronization needed. The detached freelist is allocated on the stack of the function call kmem_cache_free_bulk. Thus, the freelist head pointer is not visible to other CPUs. All objects in a SLUB freelist must belong to the same slab-page. Thus, constructing the detached freelist is about matching objects that belong to the same slab-page. The bulk free array is scanned is a progressive manor with a limited look-ahead facility. Kmem debug support is handled in call of slab_free(). Notice kmem_cache_free_bulk no longer need to disable IRQs. This only slowed down single free bulk with approx 3 cycles. Performance data: Benchmarked[1] obj size 256 bytes on CPU i7-4790K @ 4.00GHz SLUB fastpath single object quick reuse: 47 cycles(tsc) 11.931 ns To get stable and comparable numbers, the kernel have been booted with "slab_merge" (this also improve performance for larger bulk sizes). Performance data, compared against fallback bulking: bulk - fallback bulk - improvement with this patch 1 - 62 cycles(tsc) 15.662 ns - 49 cycles(tsc) 12.407 ns- improved 21.0% 2 - 55 cycles(tsc) 13.935 ns - 30 cycles(tsc) 7.506 ns - improved 45.5% 3 - 53 cycles(tsc) 13.341 ns - 23 cycles(tsc) 5.865 ns - improved 56.6% 4 - 52 cycles(tsc) 13.081 ns - 20 cycles(tsc) 5.048 ns - improved 61.5% 8 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.659 ns - improved 64.0% 16 - 49 cycles(tsc) 12.412 ns - 17 cycles(tsc) 4.495 ns - improved 65.3% 30 - 49 cycles(tsc) 12.484 ns - 18 cycles(tsc) 4.533 ns - improved 63.3% 32 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.707 ns - improved 64.0% 34 - 96 cycles(tsc) 24.243 ns - 23 cycles(tsc) 5.976 ns - improved 76.0% 48 - 83 cycles(tsc) 20.818 ns - 21 cycles(tsc) 5.329 ns - improved 74.7% 64 - 74 cycles(tsc) 18.700 ns - 20 cycles(tsc) 5.127 ns - improved 73.0% 128 - 90 cycles(tsc) 22.734 ns - 27 cycles(tsc) 6.833 ns - improved 70.0% 158 - 99 cycles(tsc) 24.776 ns - 30 cycles(tsc) 7.583 ns - improved 69.7% 250 - 104 cycles(tsc) 26.089 ns - 37 cycles(tsc) 9.280 ns - improved 64.4% Performance data, compared current in-kernel bulking: bulk - curr in-kernel - improvement with this patch 1 - 46 cycles(tsc) - 49 cycles(tsc) - improved (cycles:-3) -6.5% 2 - 27 cycles(tsc) - 30 cycles(tsc) - improved (cycles:-3) -11.1% 3 - 21 cycles(tsc) - 23 cycles(tsc) - improved (cycles:-2) -9.5% 4 - 18 cycles(tsc) - 20 cycles(tsc) - improved (cycles:-2) -11.1% 8 - 17 cycles(tsc) - 18 cycles(tsc) - improved (cycles:-1) -5.9% 16 - 18 cycles(tsc) - 17 cycles(tsc) - improved (cycles: 1) 5.6% 30 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 32 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 34 - 78 cycles(tsc) - 23 cycles(tsc) - improved (cycles:55) 70.5% 48 - 60 cycles(tsc) - 21 cycles(tsc) - improved (cycles:39) 65.0% 64 - 49 cycles(tsc) - 20 cycles(tsc) - improved (cycles:29) 59.2% 128 - 69 cycles(tsc) - 27 cycles(tsc) - improved (cycles:42) 60.9% 158 - 79 cycles(tsc) - 30 cycles(tsc) - improved (cycles:49) 62.0% 250 - 86 cycles(tsc) - 37 cycles(tsc) - improved (cycles:49) 57.0% Performance with normal SLUB merging is significantly slower for larger bulking. This is believed to (primarily) be an effect of not having to share the per-CPU data-structures, as tuning per-CPU size can achieve similar performance. bulk - slab_nomerge - normal SLUB merge 1 - 49 cycles(tsc) - 49 cycles(tsc) - merge slower with cycles:0 2 - 30 cycles(tsc) - 30 cycles(tsc) - merge slower with cycles:0 3 - 23 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:0 4 - 20 cycles(tsc) - 20 cycles(tsc) - merge slower with cycles:0 8 - 18 cycles(tsc) - 18 cycles(tsc) - merge slower with cycles:0 16 - 17 cycles(tsc) - 17 cycles(tsc) - merge slower with cycles:0 30 - 18 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:5 32 - 18 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:4 34 - 23 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:-1 48 - 21 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:1 64 - 20 cycles(tsc) - 48 cycles(tsc) - merge slower with cycles:28 128 - 27 cycles(tsc) - 57 cycles(tsc) - merge slower with cycles:30 158 - 30 cycles(tsc) - 59 cycles(tsc) - merge slower with cycles:29 250 - 37 cycles(tsc) - 56 cycles(tsc) - merge slower with cycles:19 Joint work with Alexander Duyck. [1] https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/mm/slab_bulk_test01.c [akpm@linux-foundation.org: BUG_ON -> WARN_ON;return] Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Alexander Duyck <alexander.h.duyck@redhat.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-21 07:57:49 +08:00
df->freelist = object;
df->cnt++;
p[size] = NULL; /* mark object processed */
continue;
slub: initial bulk free implementation This implements SLUB specific kmem_cache_free_bulk(). SLUB allocator now both have bulk alloc and free implemented. Choose to reenable local IRQs while calling slowpath __slab_free(). In worst case, where all objects hit slowpath call, the performance should still be faster than fallback function __kmem_cache_free_bulk(), because local_irq_{disable+enable} is very fast (7-cycles), while the fallback invokes this_cpu_cmpxchg() which is slightly slower (9-cycles). Nitpicking, this should be faster for N>=4, due to the entry cost of local_irq_{disable+enable}. Do notice that the save+restore variant is very expensive, this is key to why this optimization works. CPU: i7-4790K CPU @ 4.00GHz * local_irq_{disable,enable}: 7 cycles(tsc) - 1.821 ns * local_irq_{save,restore} : 37 cycles(tsc) - 9.443 ns Measurements on CPU CPU i7-4790K @ 4.00GHz Baseline normal fastpath (alloc+free cost): 43 cycles(tsc) 10.834 ns Bulk- fallback - this-patch 1 - 58 cycles(tsc) 14.542 ns - 43 cycles(tsc) 10.811 ns improved 25.9% 2 - 50 cycles(tsc) 12.659 ns - 27 cycles(tsc) 6.867 ns improved 46.0% 3 - 48 cycles(tsc) 12.168 ns - 21 cycles(tsc) 5.496 ns improved 56.2% 4 - 47 cycles(tsc) 11.987 ns - 24 cycles(tsc) 6.038 ns improved 48.9% 8 - 46 cycles(tsc) 11.518 ns - 17 cycles(tsc) 4.280 ns improved 63.0% 16 - 45 cycles(tsc) 11.366 ns - 17 cycles(tsc) 4.483 ns improved 62.2% 30 - 45 cycles(tsc) 11.433 ns - 18 cycles(tsc) 4.531 ns improved 60.0% 32 - 75 cycles(tsc) 18.983 ns - 58 cycles(tsc) 14.586 ns improved 22.7% 34 - 71 cycles(tsc) 17.940 ns - 53 cycles(tsc) 13.391 ns improved 25.4% 48 - 80 cycles(tsc) 20.077 ns - 65 cycles(tsc) 16.268 ns improved 18.8% 64 - 71 cycles(tsc) 17.799 ns - 53 cycles(tsc) 13.440 ns improved 25.4% 128 - 91 cycles(tsc) 22.980 ns - 79 cycles(tsc) 19.899 ns improved 13.2% 158 - 100 cycles(tsc) 25.241 ns - 90 cycles(tsc) 22.732 ns improved 10.0% 250 - 102 cycles(tsc) 25.583 ns - 95 cycles(tsc) 23.916 ns improved 6.9% Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-05 06:45:43 +08:00
}
slub: optimize bulk slowpath free by detached freelist This change focus on improving the speed of object freeing in the "slowpath" of kmem_cache_free_bulk. The calls slab_free (fastpath) and __slab_free (slowpath) have been extended with support for bulk free, which amortize the overhead of the (locked) cmpxchg_double. To use the new bulking feature, we build what I call a detached freelist. The detached freelist takes advantage of three properties: 1) the free function call owns the object that is about to be freed, thus writing into this memory is synchronization-free. 2) many freelist's can co-exist side-by-side in the same slab-page each with a separate head pointer. 3) it is the visibility of the head pointer that needs synchronization. Given these properties, the brilliant part is that the detached freelist can be constructed without any need for synchronization. The freelist is constructed directly in the page objects, without any synchronization needed. The detached freelist is allocated on the stack of the function call kmem_cache_free_bulk. Thus, the freelist head pointer is not visible to other CPUs. All objects in a SLUB freelist must belong to the same slab-page. Thus, constructing the detached freelist is about matching objects that belong to the same slab-page. The bulk free array is scanned is a progressive manor with a limited look-ahead facility. Kmem debug support is handled in call of slab_free(). Notice kmem_cache_free_bulk no longer need to disable IRQs. This only slowed down single free bulk with approx 3 cycles. Performance data: Benchmarked[1] obj size 256 bytes on CPU i7-4790K @ 4.00GHz SLUB fastpath single object quick reuse: 47 cycles(tsc) 11.931 ns To get stable and comparable numbers, the kernel have been booted with "slab_merge" (this also improve performance for larger bulk sizes). Performance data, compared against fallback bulking: bulk - fallback bulk - improvement with this patch 1 - 62 cycles(tsc) 15.662 ns - 49 cycles(tsc) 12.407 ns- improved 21.0% 2 - 55 cycles(tsc) 13.935 ns - 30 cycles(tsc) 7.506 ns - improved 45.5% 3 - 53 cycles(tsc) 13.341 ns - 23 cycles(tsc) 5.865 ns - improved 56.6% 4 - 52 cycles(tsc) 13.081 ns - 20 cycles(tsc) 5.048 ns - improved 61.5% 8 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.659 ns - improved 64.0% 16 - 49 cycles(tsc) 12.412 ns - 17 cycles(tsc) 4.495 ns - improved 65.3% 30 - 49 cycles(tsc) 12.484 ns - 18 cycles(tsc) 4.533 ns - improved 63.3% 32 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.707 ns - improved 64.0% 34 - 96 cycles(tsc) 24.243 ns - 23 cycles(tsc) 5.976 ns - improved 76.0% 48 - 83 cycles(tsc) 20.818 ns - 21 cycles(tsc) 5.329 ns - improved 74.7% 64 - 74 cycles(tsc) 18.700 ns - 20 cycles(tsc) 5.127 ns - improved 73.0% 128 - 90 cycles(tsc) 22.734 ns - 27 cycles(tsc) 6.833 ns - improved 70.0% 158 - 99 cycles(tsc) 24.776 ns - 30 cycles(tsc) 7.583 ns - improved 69.7% 250 - 104 cycles(tsc) 26.089 ns - 37 cycles(tsc) 9.280 ns - improved 64.4% Performance data, compared current in-kernel bulking: bulk - curr in-kernel - improvement with this patch 1 - 46 cycles(tsc) - 49 cycles(tsc) - improved (cycles:-3) -6.5% 2 - 27 cycles(tsc) - 30 cycles(tsc) - improved (cycles:-3) -11.1% 3 - 21 cycles(tsc) - 23 cycles(tsc) - improved (cycles:-2) -9.5% 4 - 18 cycles(tsc) - 20 cycles(tsc) - improved (cycles:-2) -11.1% 8 - 17 cycles(tsc) - 18 cycles(tsc) - improved (cycles:-1) -5.9% 16 - 18 cycles(tsc) - 17 cycles(tsc) - improved (cycles: 1) 5.6% 30 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 32 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 34 - 78 cycles(tsc) - 23 cycles(tsc) - improved (cycles:55) 70.5% 48 - 60 cycles(tsc) - 21 cycles(tsc) - improved (cycles:39) 65.0% 64 - 49 cycles(tsc) - 20 cycles(tsc) - improved (cycles:29) 59.2% 128 - 69 cycles(tsc) - 27 cycles(tsc) - improved (cycles:42) 60.9% 158 - 79 cycles(tsc) - 30 cycles(tsc) - improved (cycles:49) 62.0% 250 - 86 cycles(tsc) - 37 cycles(tsc) - improved (cycles:49) 57.0% Performance with normal SLUB merging is significantly slower for larger bulking. This is believed to (primarily) be an effect of not having to share the per-CPU data-structures, as tuning per-CPU size can achieve similar performance. bulk - slab_nomerge - normal SLUB merge 1 - 49 cycles(tsc) - 49 cycles(tsc) - merge slower with cycles:0 2 - 30 cycles(tsc) - 30 cycles(tsc) - merge slower with cycles:0 3 - 23 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:0 4 - 20 cycles(tsc) - 20 cycles(tsc) - merge slower with cycles:0 8 - 18 cycles(tsc) - 18 cycles(tsc) - merge slower with cycles:0 16 - 17 cycles(tsc) - 17 cycles(tsc) - merge slower with cycles:0 30 - 18 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:5 32 - 18 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:4 34 - 23 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:-1 48 - 21 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:1 64 - 20 cycles(tsc) - 48 cycles(tsc) - merge slower with cycles:28 128 - 27 cycles(tsc) - 57 cycles(tsc) - merge slower with cycles:30 158 - 30 cycles(tsc) - 59 cycles(tsc) - merge slower with cycles:29 250 - 37 cycles(tsc) - 56 cycles(tsc) - merge slower with cycles:19 Joint work with Alexander Duyck. [1] https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/mm/slab_bulk_test01.c [akpm@linux-foundation.org: BUG_ON -> WARN_ON;return] Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Alexander Duyck <alexander.h.duyck@redhat.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-21 07:57:49 +08:00
/* Limit look ahead search */
if (!--lookahead)
break;
if (!first_skipped_index)
first_skipped_index = size + 1;
slub: initial bulk free implementation This implements SLUB specific kmem_cache_free_bulk(). SLUB allocator now both have bulk alloc and free implemented. Choose to reenable local IRQs while calling slowpath __slab_free(). In worst case, where all objects hit slowpath call, the performance should still be faster than fallback function __kmem_cache_free_bulk(), because local_irq_{disable+enable} is very fast (7-cycles), while the fallback invokes this_cpu_cmpxchg() which is slightly slower (9-cycles). Nitpicking, this should be faster for N>=4, due to the entry cost of local_irq_{disable+enable}. Do notice that the save+restore variant is very expensive, this is key to why this optimization works. CPU: i7-4790K CPU @ 4.00GHz * local_irq_{disable,enable}: 7 cycles(tsc) - 1.821 ns * local_irq_{save,restore} : 37 cycles(tsc) - 9.443 ns Measurements on CPU CPU i7-4790K @ 4.00GHz Baseline normal fastpath (alloc+free cost): 43 cycles(tsc) 10.834 ns Bulk- fallback - this-patch 1 - 58 cycles(tsc) 14.542 ns - 43 cycles(tsc) 10.811 ns improved 25.9% 2 - 50 cycles(tsc) 12.659 ns - 27 cycles(tsc) 6.867 ns improved 46.0% 3 - 48 cycles(tsc) 12.168 ns - 21 cycles(tsc) 5.496 ns improved 56.2% 4 - 47 cycles(tsc) 11.987 ns - 24 cycles(tsc) 6.038 ns improved 48.9% 8 - 46 cycles(tsc) 11.518 ns - 17 cycles(tsc) 4.280 ns improved 63.0% 16 - 45 cycles(tsc) 11.366 ns - 17 cycles(tsc) 4.483 ns improved 62.2% 30 - 45 cycles(tsc) 11.433 ns - 18 cycles(tsc) 4.531 ns improved 60.0% 32 - 75 cycles(tsc) 18.983 ns - 58 cycles(tsc) 14.586 ns improved 22.7% 34 - 71 cycles(tsc) 17.940 ns - 53 cycles(tsc) 13.391 ns improved 25.4% 48 - 80 cycles(tsc) 20.077 ns - 65 cycles(tsc) 16.268 ns improved 18.8% 64 - 71 cycles(tsc) 17.799 ns - 53 cycles(tsc) 13.440 ns improved 25.4% 128 - 91 cycles(tsc) 22.980 ns - 79 cycles(tsc) 19.899 ns improved 13.2% 158 - 100 cycles(tsc) 25.241 ns - 90 cycles(tsc) 22.732 ns improved 10.0% 250 - 102 cycles(tsc) 25.583 ns - 95 cycles(tsc) 23.916 ns improved 6.9% Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-05 06:45:43 +08:00
}
slub: optimize bulk slowpath free by detached freelist This change focus on improving the speed of object freeing in the "slowpath" of kmem_cache_free_bulk. The calls slab_free (fastpath) and __slab_free (slowpath) have been extended with support for bulk free, which amortize the overhead of the (locked) cmpxchg_double. To use the new bulking feature, we build what I call a detached freelist. The detached freelist takes advantage of three properties: 1) the free function call owns the object that is about to be freed, thus writing into this memory is synchronization-free. 2) many freelist's can co-exist side-by-side in the same slab-page each with a separate head pointer. 3) it is the visibility of the head pointer that needs synchronization. Given these properties, the brilliant part is that the detached freelist can be constructed without any need for synchronization. The freelist is constructed directly in the page objects, without any synchronization needed. The detached freelist is allocated on the stack of the function call kmem_cache_free_bulk. Thus, the freelist head pointer is not visible to other CPUs. All objects in a SLUB freelist must belong to the same slab-page. Thus, constructing the detached freelist is about matching objects that belong to the same slab-page. The bulk free array is scanned is a progressive manor with a limited look-ahead facility. Kmem debug support is handled in call of slab_free(). Notice kmem_cache_free_bulk no longer need to disable IRQs. This only slowed down single free bulk with approx 3 cycles. Performance data: Benchmarked[1] obj size 256 bytes on CPU i7-4790K @ 4.00GHz SLUB fastpath single object quick reuse: 47 cycles(tsc) 11.931 ns To get stable and comparable numbers, the kernel have been booted with "slab_merge" (this also improve performance for larger bulk sizes). Performance data, compared against fallback bulking: bulk - fallback bulk - improvement with this patch 1 - 62 cycles(tsc) 15.662 ns - 49 cycles(tsc) 12.407 ns- improved 21.0% 2 - 55 cycles(tsc) 13.935 ns - 30 cycles(tsc) 7.506 ns - improved 45.5% 3 - 53 cycles(tsc) 13.341 ns - 23 cycles(tsc) 5.865 ns - improved 56.6% 4 - 52 cycles(tsc) 13.081 ns - 20 cycles(tsc) 5.048 ns - improved 61.5% 8 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.659 ns - improved 64.0% 16 - 49 cycles(tsc) 12.412 ns - 17 cycles(tsc) 4.495 ns - improved 65.3% 30 - 49 cycles(tsc) 12.484 ns - 18 cycles(tsc) 4.533 ns - improved 63.3% 32 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.707 ns - improved 64.0% 34 - 96 cycles(tsc) 24.243 ns - 23 cycles(tsc) 5.976 ns - improved 76.0% 48 - 83 cycles(tsc) 20.818 ns - 21 cycles(tsc) 5.329 ns - improved 74.7% 64 - 74 cycles(tsc) 18.700 ns - 20 cycles(tsc) 5.127 ns - improved 73.0% 128 - 90 cycles(tsc) 22.734 ns - 27 cycles(tsc) 6.833 ns - improved 70.0% 158 - 99 cycles(tsc) 24.776 ns - 30 cycles(tsc) 7.583 ns - improved 69.7% 250 - 104 cycles(tsc) 26.089 ns - 37 cycles(tsc) 9.280 ns - improved 64.4% Performance data, compared current in-kernel bulking: bulk - curr in-kernel - improvement with this patch 1 - 46 cycles(tsc) - 49 cycles(tsc) - improved (cycles:-3) -6.5% 2 - 27 cycles(tsc) - 30 cycles(tsc) - improved (cycles:-3) -11.1% 3 - 21 cycles(tsc) - 23 cycles(tsc) - improved (cycles:-2) -9.5% 4 - 18 cycles(tsc) - 20 cycles(tsc) - improved (cycles:-2) -11.1% 8 - 17 cycles(tsc) - 18 cycles(tsc) - improved (cycles:-1) -5.9% 16 - 18 cycles(tsc) - 17 cycles(tsc) - improved (cycles: 1) 5.6% 30 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 32 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 34 - 78 cycles(tsc) - 23 cycles(tsc) - improved (cycles:55) 70.5% 48 - 60 cycles(tsc) - 21 cycles(tsc) - improved (cycles:39) 65.0% 64 - 49 cycles(tsc) - 20 cycles(tsc) - improved (cycles:29) 59.2% 128 - 69 cycles(tsc) - 27 cycles(tsc) - improved (cycles:42) 60.9% 158 - 79 cycles(tsc) - 30 cycles(tsc) - improved (cycles:49) 62.0% 250 - 86 cycles(tsc) - 37 cycles(tsc) - improved (cycles:49) 57.0% Performance with normal SLUB merging is significantly slower for larger bulking. This is believed to (primarily) be an effect of not having to share the per-CPU data-structures, as tuning per-CPU size can achieve similar performance. bulk - slab_nomerge - normal SLUB merge 1 - 49 cycles(tsc) - 49 cycles(tsc) - merge slower with cycles:0 2 - 30 cycles(tsc) - 30 cycles(tsc) - merge slower with cycles:0 3 - 23 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:0 4 - 20 cycles(tsc) - 20 cycles(tsc) - merge slower with cycles:0 8 - 18 cycles(tsc) - 18 cycles(tsc) - merge slower with cycles:0 16 - 17 cycles(tsc) - 17 cycles(tsc) - merge slower with cycles:0 30 - 18 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:5 32 - 18 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:4 34 - 23 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:-1 48 - 21 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:1 64 - 20 cycles(tsc) - 48 cycles(tsc) - merge slower with cycles:28 128 - 27 cycles(tsc) - 57 cycles(tsc) - merge slower with cycles:30 158 - 30 cycles(tsc) - 59 cycles(tsc) - merge slower with cycles:29 250 - 37 cycles(tsc) - 56 cycles(tsc) - merge slower with cycles:19 Joint work with Alexander Duyck. [1] https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/mm/slab_bulk_test01.c [akpm@linux-foundation.org: BUG_ON -> WARN_ON;return] Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Alexander Duyck <alexander.h.duyck@redhat.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-21 07:57:49 +08:00
return first_skipped_index;
}
/* Note that interrupts must be enabled when calling this function. */
void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
slub: optimize bulk slowpath free by detached freelist This change focus on improving the speed of object freeing in the "slowpath" of kmem_cache_free_bulk. The calls slab_free (fastpath) and __slab_free (slowpath) have been extended with support for bulk free, which amortize the overhead of the (locked) cmpxchg_double. To use the new bulking feature, we build what I call a detached freelist. The detached freelist takes advantage of three properties: 1) the free function call owns the object that is about to be freed, thus writing into this memory is synchronization-free. 2) many freelist's can co-exist side-by-side in the same slab-page each with a separate head pointer. 3) it is the visibility of the head pointer that needs synchronization. Given these properties, the brilliant part is that the detached freelist can be constructed without any need for synchronization. The freelist is constructed directly in the page objects, without any synchronization needed. The detached freelist is allocated on the stack of the function call kmem_cache_free_bulk. Thus, the freelist head pointer is not visible to other CPUs. All objects in a SLUB freelist must belong to the same slab-page. Thus, constructing the detached freelist is about matching objects that belong to the same slab-page. The bulk free array is scanned is a progressive manor with a limited look-ahead facility. Kmem debug support is handled in call of slab_free(). Notice kmem_cache_free_bulk no longer need to disable IRQs. This only slowed down single free bulk with approx 3 cycles. Performance data: Benchmarked[1] obj size 256 bytes on CPU i7-4790K @ 4.00GHz SLUB fastpath single object quick reuse: 47 cycles(tsc) 11.931 ns To get stable and comparable numbers, the kernel have been booted with "slab_merge" (this also improve performance for larger bulk sizes). Performance data, compared against fallback bulking: bulk - fallback bulk - improvement with this patch 1 - 62 cycles(tsc) 15.662 ns - 49 cycles(tsc) 12.407 ns- improved 21.0% 2 - 55 cycles(tsc) 13.935 ns - 30 cycles(tsc) 7.506 ns - improved 45.5% 3 - 53 cycles(tsc) 13.341 ns - 23 cycles(tsc) 5.865 ns - improved 56.6% 4 - 52 cycles(tsc) 13.081 ns - 20 cycles(tsc) 5.048 ns - improved 61.5% 8 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.659 ns - improved 64.0% 16 - 49 cycles(tsc) 12.412 ns - 17 cycles(tsc) 4.495 ns - improved 65.3% 30 - 49 cycles(tsc) 12.484 ns - 18 cycles(tsc) 4.533 ns - improved 63.3% 32 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.707 ns - improved 64.0% 34 - 96 cycles(tsc) 24.243 ns - 23 cycles(tsc) 5.976 ns - improved 76.0% 48 - 83 cycles(tsc) 20.818 ns - 21 cycles(tsc) 5.329 ns - improved 74.7% 64 - 74 cycles(tsc) 18.700 ns - 20 cycles(tsc) 5.127 ns - improved 73.0% 128 - 90 cycles(tsc) 22.734 ns - 27 cycles(tsc) 6.833 ns - improved 70.0% 158 - 99 cycles(tsc) 24.776 ns - 30 cycles(tsc) 7.583 ns - improved 69.7% 250 - 104 cycles(tsc) 26.089 ns - 37 cycles(tsc) 9.280 ns - improved 64.4% Performance data, compared current in-kernel bulking: bulk - curr in-kernel - improvement with this patch 1 - 46 cycles(tsc) - 49 cycles(tsc) - improved (cycles:-3) -6.5% 2 - 27 cycles(tsc) - 30 cycles(tsc) - improved (cycles:-3) -11.1% 3 - 21 cycles(tsc) - 23 cycles(tsc) - improved (cycles:-2) -9.5% 4 - 18 cycles(tsc) - 20 cycles(tsc) - improved (cycles:-2) -11.1% 8 - 17 cycles(tsc) - 18 cycles(tsc) - improved (cycles:-1) -5.9% 16 - 18 cycles(tsc) - 17 cycles(tsc) - improved (cycles: 1) 5.6% 30 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 32 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 34 - 78 cycles(tsc) - 23 cycles(tsc) - improved (cycles:55) 70.5% 48 - 60 cycles(tsc) - 21 cycles(tsc) - improved (cycles:39) 65.0% 64 - 49 cycles(tsc) - 20 cycles(tsc) - improved (cycles:29) 59.2% 128 - 69 cycles(tsc) - 27 cycles(tsc) - improved (cycles:42) 60.9% 158 - 79 cycles(tsc) - 30 cycles(tsc) - improved (cycles:49) 62.0% 250 - 86 cycles(tsc) - 37 cycles(tsc) - improved (cycles:49) 57.0% Performance with normal SLUB merging is significantly slower for larger bulking. This is believed to (primarily) be an effect of not having to share the per-CPU data-structures, as tuning per-CPU size can achieve similar performance. bulk - slab_nomerge - normal SLUB merge 1 - 49 cycles(tsc) - 49 cycles(tsc) - merge slower with cycles:0 2 - 30 cycles(tsc) - 30 cycles(tsc) - merge slower with cycles:0 3 - 23 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:0 4 - 20 cycles(tsc) - 20 cycles(tsc) - merge slower with cycles:0 8 - 18 cycles(tsc) - 18 cycles(tsc) - merge slower with cycles:0 16 - 17 cycles(tsc) - 17 cycles(tsc) - merge slower with cycles:0 30 - 18 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:5 32 - 18 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:4 34 - 23 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:-1 48 - 21 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:1 64 - 20 cycles(tsc) - 48 cycles(tsc) - merge slower with cycles:28 128 - 27 cycles(tsc) - 57 cycles(tsc) - merge slower with cycles:30 158 - 30 cycles(tsc) - 59 cycles(tsc) - merge slower with cycles:29 250 - 37 cycles(tsc) - 56 cycles(tsc) - merge slower with cycles:19 Joint work with Alexander Duyck. [1] https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/mm/slab_bulk_test01.c [akpm@linux-foundation.org: BUG_ON -> WARN_ON;return] Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Alexander Duyck <alexander.h.duyck@redhat.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-21 07:57:49 +08:00
{
if (WARN_ON(!size))
return;
memcg_slab_free_hook(s, p, size);
slub: optimize bulk slowpath free by detached freelist This change focus on improving the speed of object freeing in the "slowpath" of kmem_cache_free_bulk. The calls slab_free (fastpath) and __slab_free (slowpath) have been extended with support for bulk free, which amortize the overhead of the (locked) cmpxchg_double. To use the new bulking feature, we build what I call a detached freelist. The detached freelist takes advantage of three properties: 1) the free function call owns the object that is about to be freed, thus writing into this memory is synchronization-free. 2) many freelist's can co-exist side-by-side in the same slab-page each with a separate head pointer. 3) it is the visibility of the head pointer that needs synchronization. Given these properties, the brilliant part is that the detached freelist can be constructed without any need for synchronization. The freelist is constructed directly in the page objects, without any synchronization needed. The detached freelist is allocated on the stack of the function call kmem_cache_free_bulk. Thus, the freelist head pointer is not visible to other CPUs. All objects in a SLUB freelist must belong to the same slab-page. Thus, constructing the detached freelist is about matching objects that belong to the same slab-page. The bulk free array is scanned is a progressive manor with a limited look-ahead facility. Kmem debug support is handled in call of slab_free(). Notice kmem_cache_free_bulk no longer need to disable IRQs. This only slowed down single free bulk with approx 3 cycles. Performance data: Benchmarked[1] obj size 256 bytes on CPU i7-4790K @ 4.00GHz SLUB fastpath single object quick reuse: 47 cycles(tsc) 11.931 ns To get stable and comparable numbers, the kernel have been booted with "slab_merge" (this also improve performance for larger bulk sizes). Performance data, compared against fallback bulking: bulk - fallback bulk - improvement with this patch 1 - 62 cycles(tsc) 15.662 ns - 49 cycles(tsc) 12.407 ns- improved 21.0% 2 - 55 cycles(tsc) 13.935 ns - 30 cycles(tsc) 7.506 ns - improved 45.5% 3 - 53 cycles(tsc) 13.341 ns - 23 cycles(tsc) 5.865 ns - improved 56.6% 4 - 52 cycles(tsc) 13.081 ns - 20 cycles(tsc) 5.048 ns - improved 61.5% 8 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.659 ns - improved 64.0% 16 - 49 cycles(tsc) 12.412 ns - 17 cycles(tsc) 4.495 ns - improved 65.3% 30 - 49 cycles(tsc) 12.484 ns - 18 cycles(tsc) 4.533 ns - improved 63.3% 32 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.707 ns - improved 64.0% 34 - 96 cycles(tsc) 24.243 ns - 23 cycles(tsc) 5.976 ns - improved 76.0% 48 - 83 cycles(tsc) 20.818 ns - 21 cycles(tsc) 5.329 ns - improved 74.7% 64 - 74 cycles(tsc) 18.700 ns - 20 cycles(tsc) 5.127 ns - improved 73.0% 128 - 90 cycles(tsc) 22.734 ns - 27 cycles(tsc) 6.833 ns - improved 70.0% 158 - 99 cycles(tsc) 24.776 ns - 30 cycles(tsc) 7.583 ns - improved 69.7% 250 - 104 cycles(tsc) 26.089 ns - 37 cycles(tsc) 9.280 ns - improved 64.4% Performance data, compared current in-kernel bulking: bulk - curr in-kernel - improvement with this patch 1 - 46 cycles(tsc) - 49 cycles(tsc) - improved (cycles:-3) -6.5% 2 - 27 cycles(tsc) - 30 cycles(tsc) - improved (cycles:-3) -11.1% 3 - 21 cycles(tsc) - 23 cycles(tsc) - improved (cycles:-2) -9.5% 4 - 18 cycles(tsc) - 20 cycles(tsc) - improved (cycles:-2) -11.1% 8 - 17 cycles(tsc) - 18 cycles(tsc) - improved (cycles:-1) -5.9% 16 - 18 cycles(tsc) - 17 cycles(tsc) - improved (cycles: 1) 5.6% 30 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 32 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 34 - 78 cycles(tsc) - 23 cycles(tsc) - improved (cycles:55) 70.5% 48 - 60 cycles(tsc) - 21 cycles(tsc) - improved (cycles:39) 65.0% 64 - 49 cycles(tsc) - 20 cycles(tsc) - improved (cycles:29) 59.2% 128 - 69 cycles(tsc) - 27 cycles(tsc) - improved (cycles:42) 60.9% 158 - 79 cycles(tsc) - 30 cycles(tsc) - improved (cycles:49) 62.0% 250 - 86 cycles(tsc) - 37 cycles(tsc) - improved (cycles:49) 57.0% Performance with normal SLUB merging is significantly slower for larger bulking. This is believed to (primarily) be an effect of not having to share the per-CPU data-structures, as tuning per-CPU size can achieve similar performance. bulk - slab_nomerge - normal SLUB merge 1 - 49 cycles(tsc) - 49 cycles(tsc) - merge slower with cycles:0 2 - 30 cycles(tsc) - 30 cycles(tsc) - merge slower with cycles:0 3 - 23 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:0 4 - 20 cycles(tsc) - 20 cycles(tsc) - merge slower with cycles:0 8 - 18 cycles(tsc) - 18 cycles(tsc) - merge slower with cycles:0 16 - 17 cycles(tsc) - 17 cycles(tsc) - merge slower with cycles:0 30 - 18 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:5 32 - 18 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:4 34 - 23 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:-1 48 - 21 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:1 64 - 20 cycles(tsc) - 48 cycles(tsc) - merge slower with cycles:28 128 - 27 cycles(tsc) - 57 cycles(tsc) - merge slower with cycles:30 158 - 30 cycles(tsc) - 59 cycles(tsc) - merge slower with cycles:29 250 - 37 cycles(tsc) - 56 cycles(tsc) - merge slower with cycles:19 Joint work with Alexander Duyck. [1] https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/mm/slab_bulk_test01.c [akpm@linux-foundation.org: BUG_ON -> WARN_ON;return] Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Alexander Duyck <alexander.h.duyck@redhat.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-21 07:57:49 +08:00
do {
struct detached_freelist df;
size = build_detached_freelist(s, size, p, &df);
slub: avoid false-postive warning The slub allocator gives us some incorrect warnings when CONFIG_PROFILE_ANNOTATED_BRANCHES is set, as the unlikely() macro prevents it from seeing that the return code matches what it was before: mm/slub.c: In function `kmem_cache_free_bulk': mm/slub.c:262:23: error: `df.s' may be used uninitialized in this function [-Werror=maybe-uninitialized] mm/slub.c:2943:3: error: `df.cnt' may be used uninitialized in this function [-Werror=maybe-uninitialized] mm/slub.c:2933:4470: error: `df.freelist' may be used uninitialized in this function [-Werror=maybe-uninitialized] mm/slub.c:2943:3: error: `df.tail' may be used uninitialized in this function [-Werror=maybe-uninitialized] I have not been able to come up with a perfect way for dealing with this, the three options I see are: - add a bogus initialization, which would increase the runtime overhead - replace unlikely() with unlikely_notrace() - remove the unlikely() annotation completely I checked the object code for a typical x86 configuration and the last two cases produce the same result, so I went for the last one, which is the simplest. Link: http://lkml.kernel.org/r/20161024155704.3114445-1-arnd@arndb.de Signed-off-by: Arnd Bergmann <arnd@arndb.de> Acked-by: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Arnd Bergmann <arnd@arndb.de> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Laura Abbott <labbott@fedoraproject.org> Cc: Alexander Potapenko <glider@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-12-13 08:41:35 +08:00
if (!df.page)
slub: optimize bulk slowpath free by detached freelist This change focus on improving the speed of object freeing in the "slowpath" of kmem_cache_free_bulk. The calls slab_free (fastpath) and __slab_free (slowpath) have been extended with support for bulk free, which amortize the overhead of the (locked) cmpxchg_double. To use the new bulking feature, we build what I call a detached freelist. The detached freelist takes advantage of three properties: 1) the free function call owns the object that is about to be freed, thus writing into this memory is synchronization-free. 2) many freelist's can co-exist side-by-side in the same slab-page each with a separate head pointer. 3) it is the visibility of the head pointer that needs synchronization. Given these properties, the brilliant part is that the detached freelist can be constructed without any need for synchronization. The freelist is constructed directly in the page objects, without any synchronization needed. The detached freelist is allocated on the stack of the function call kmem_cache_free_bulk. Thus, the freelist head pointer is not visible to other CPUs. All objects in a SLUB freelist must belong to the same slab-page. Thus, constructing the detached freelist is about matching objects that belong to the same slab-page. The bulk free array is scanned is a progressive manor with a limited look-ahead facility. Kmem debug support is handled in call of slab_free(). Notice kmem_cache_free_bulk no longer need to disable IRQs. This only slowed down single free bulk with approx 3 cycles. Performance data: Benchmarked[1] obj size 256 bytes on CPU i7-4790K @ 4.00GHz SLUB fastpath single object quick reuse: 47 cycles(tsc) 11.931 ns To get stable and comparable numbers, the kernel have been booted with "slab_merge" (this also improve performance for larger bulk sizes). Performance data, compared against fallback bulking: bulk - fallback bulk - improvement with this patch 1 - 62 cycles(tsc) 15.662 ns - 49 cycles(tsc) 12.407 ns- improved 21.0% 2 - 55 cycles(tsc) 13.935 ns - 30 cycles(tsc) 7.506 ns - improved 45.5% 3 - 53 cycles(tsc) 13.341 ns - 23 cycles(tsc) 5.865 ns - improved 56.6% 4 - 52 cycles(tsc) 13.081 ns - 20 cycles(tsc) 5.048 ns - improved 61.5% 8 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.659 ns - improved 64.0% 16 - 49 cycles(tsc) 12.412 ns - 17 cycles(tsc) 4.495 ns - improved 65.3% 30 - 49 cycles(tsc) 12.484 ns - 18 cycles(tsc) 4.533 ns - improved 63.3% 32 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.707 ns - improved 64.0% 34 - 96 cycles(tsc) 24.243 ns - 23 cycles(tsc) 5.976 ns - improved 76.0% 48 - 83 cycles(tsc) 20.818 ns - 21 cycles(tsc) 5.329 ns - improved 74.7% 64 - 74 cycles(tsc) 18.700 ns - 20 cycles(tsc) 5.127 ns - improved 73.0% 128 - 90 cycles(tsc) 22.734 ns - 27 cycles(tsc) 6.833 ns - improved 70.0% 158 - 99 cycles(tsc) 24.776 ns - 30 cycles(tsc) 7.583 ns - improved 69.7% 250 - 104 cycles(tsc) 26.089 ns - 37 cycles(tsc) 9.280 ns - improved 64.4% Performance data, compared current in-kernel bulking: bulk - curr in-kernel - improvement with this patch 1 - 46 cycles(tsc) - 49 cycles(tsc) - improved (cycles:-3) -6.5% 2 - 27 cycles(tsc) - 30 cycles(tsc) - improved (cycles:-3) -11.1% 3 - 21 cycles(tsc) - 23 cycles(tsc) - improved (cycles:-2) -9.5% 4 - 18 cycles(tsc) - 20 cycles(tsc) - improved (cycles:-2) -11.1% 8 - 17 cycles(tsc) - 18 cycles(tsc) - improved (cycles:-1) -5.9% 16 - 18 cycles(tsc) - 17 cycles(tsc) - improved (cycles: 1) 5.6% 30 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 32 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 34 - 78 cycles(tsc) - 23 cycles(tsc) - improved (cycles:55) 70.5% 48 - 60 cycles(tsc) - 21 cycles(tsc) - improved (cycles:39) 65.0% 64 - 49 cycles(tsc) - 20 cycles(tsc) - improved (cycles:29) 59.2% 128 - 69 cycles(tsc) - 27 cycles(tsc) - improved (cycles:42) 60.9% 158 - 79 cycles(tsc) - 30 cycles(tsc) - improved (cycles:49) 62.0% 250 - 86 cycles(tsc) - 37 cycles(tsc) - improved (cycles:49) 57.0% Performance with normal SLUB merging is significantly slower for larger bulking. This is believed to (primarily) be an effect of not having to share the per-CPU data-structures, as tuning per-CPU size can achieve similar performance. bulk - slab_nomerge - normal SLUB merge 1 - 49 cycles(tsc) - 49 cycles(tsc) - merge slower with cycles:0 2 - 30 cycles(tsc) - 30 cycles(tsc) - merge slower with cycles:0 3 - 23 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:0 4 - 20 cycles(tsc) - 20 cycles(tsc) - merge slower with cycles:0 8 - 18 cycles(tsc) - 18 cycles(tsc) - merge slower with cycles:0 16 - 17 cycles(tsc) - 17 cycles(tsc) - merge slower with cycles:0 30 - 18 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:5 32 - 18 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:4 34 - 23 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:-1 48 - 21 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:1 64 - 20 cycles(tsc) - 48 cycles(tsc) - merge slower with cycles:28 128 - 27 cycles(tsc) - 57 cycles(tsc) - merge slower with cycles:30 158 - 30 cycles(tsc) - 59 cycles(tsc) - merge slower with cycles:29 250 - 37 cycles(tsc) - 56 cycles(tsc) - merge slower with cycles:19 Joint work with Alexander Duyck. [1] https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/mm/slab_bulk_test01.c [akpm@linux-foundation.org: BUG_ON -> WARN_ON;return] Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Alexander Duyck <alexander.h.duyck@redhat.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-21 07:57:49 +08:00
continue;
slab_free(df.s, df.page, df.freelist, df.tail, df.cnt, _RET_IP_);
slub: optimize bulk slowpath free by detached freelist This change focus on improving the speed of object freeing in the "slowpath" of kmem_cache_free_bulk. The calls slab_free (fastpath) and __slab_free (slowpath) have been extended with support for bulk free, which amortize the overhead of the (locked) cmpxchg_double. To use the new bulking feature, we build what I call a detached freelist. The detached freelist takes advantage of three properties: 1) the free function call owns the object that is about to be freed, thus writing into this memory is synchronization-free. 2) many freelist's can co-exist side-by-side in the same slab-page each with a separate head pointer. 3) it is the visibility of the head pointer that needs synchronization. Given these properties, the brilliant part is that the detached freelist can be constructed without any need for synchronization. The freelist is constructed directly in the page objects, without any synchronization needed. The detached freelist is allocated on the stack of the function call kmem_cache_free_bulk. Thus, the freelist head pointer is not visible to other CPUs. All objects in a SLUB freelist must belong to the same slab-page. Thus, constructing the detached freelist is about matching objects that belong to the same slab-page. The bulk free array is scanned is a progressive manor with a limited look-ahead facility. Kmem debug support is handled in call of slab_free(). Notice kmem_cache_free_bulk no longer need to disable IRQs. This only slowed down single free bulk with approx 3 cycles. Performance data: Benchmarked[1] obj size 256 bytes on CPU i7-4790K @ 4.00GHz SLUB fastpath single object quick reuse: 47 cycles(tsc) 11.931 ns To get stable and comparable numbers, the kernel have been booted with "slab_merge" (this also improve performance for larger bulk sizes). Performance data, compared against fallback bulking: bulk - fallback bulk - improvement with this patch 1 - 62 cycles(tsc) 15.662 ns - 49 cycles(tsc) 12.407 ns- improved 21.0% 2 - 55 cycles(tsc) 13.935 ns - 30 cycles(tsc) 7.506 ns - improved 45.5% 3 - 53 cycles(tsc) 13.341 ns - 23 cycles(tsc) 5.865 ns - improved 56.6% 4 - 52 cycles(tsc) 13.081 ns - 20 cycles(tsc) 5.048 ns - improved 61.5% 8 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.659 ns - improved 64.0% 16 - 49 cycles(tsc) 12.412 ns - 17 cycles(tsc) 4.495 ns - improved 65.3% 30 - 49 cycles(tsc) 12.484 ns - 18 cycles(tsc) 4.533 ns - improved 63.3% 32 - 50 cycles(tsc) 12.627 ns - 18 cycles(tsc) 4.707 ns - improved 64.0% 34 - 96 cycles(tsc) 24.243 ns - 23 cycles(tsc) 5.976 ns - improved 76.0% 48 - 83 cycles(tsc) 20.818 ns - 21 cycles(tsc) 5.329 ns - improved 74.7% 64 - 74 cycles(tsc) 18.700 ns - 20 cycles(tsc) 5.127 ns - improved 73.0% 128 - 90 cycles(tsc) 22.734 ns - 27 cycles(tsc) 6.833 ns - improved 70.0% 158 - 99 cycles(tsc) 24.776 ns - 30 cycles(tsc) 7.583 ns - improved 69.7% 250 - 104 cycles(tsc) 26.089 ns - 37 cycles(tsc) 9.280 ns - improved 64.4% Performance data, compared current in-kernel bulking: bulk - curr in-kernel - improvement with this patch 1 - 46 cycles(tsc) - 49 cycles(tsc) - improved (cycles:-3) -6.5% 2 - 27 cycles(tsc) - 30 cycles(tsc) - improved (cycles:-3) -11.1% 3 - 21 cycles(tsc) - 23 cycles(tsc) - improved (cycles:-2) -9.5% 4 - 18 cycles(tsc) - 20 cycles(tsc) - improved (cycles:-2) -11.1% 8 - 17 cycles(tsc) - 18 cycles(tsc) - improved (cycles:-1) -5.9% 16 - 18 cycles(tsc) - 17 cycles(tsc) - improved (cycles: 1) 5.6% 30 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 32 - 18 cycles(tsc) - 18 cycles(tsc) - improved (cycles: 0) 0.0% 34 - 78 cycles(tsc) - 23 cycles(tsc) - improved (cycles:55) 70.5% 48 - 60 cycles(tsc) - 21 cycles(tsc) - improved (cycles:39) 65.0% 64 - 49 cycles(tsc) - 20 cycles(tsc) - improved (cycles:29) 59.2% 128 - 69 cycles(tsc) - 27 cycles(tsc) - improved (cycles:42) 60.9% 158 - 79 cycles(tsc) - 30 cycles(tsc) - improved (cycles:49) 62.0% 250 - 86 cycles(tsc) - 37 cycles(tsc) - improved (cycles:49) 57.0% Performance with normal SLUB merging is significantly slower for larger bulking. This is believed to (primarily) be an effect of not having to share the per-CPU data-structures, as tuning per-CPU size can achieve similar performance. bulk - slab_nomerge - normal SLUB merge 1 - 49 cycles(tsc) - 49 cycles(tsc) - merge slower with cycles:0 2 - 30 cycles(tsc) - 30 cycles(tsc) - merge slower with cycles:0 3 - 23 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:0 4 - 20 cycles(tsc) - 20 cycles(tsc) - merge slower with cycles:0 8 - 18 cycles(tsc) - 18 cycles(tsc) - merge slower with cycles:0 16 - 17 cycles(tsc) - 17 cycles(tsc) - merge slower with cycles:0 30 - 18 cycles(tsc) - 23 cycles(tsc) - merge slower with cycles:5 32 - 18 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:4 34 - 23 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:-1 48 - 21 cycles(tsc) - 22 cycles(tsc) - merge slower with cycles:1 64 - 20 cycles(tsc) - 48 cycles(tsc) - merge slower with cycles:28 128 - 27 cycles(tsc) - 57 cycles(tsc) - merge slower with cycles:30 158 - 30 cycles(tsc) - 59 cycles(tsc) - merge slower with cycles:29 250 - 37 cycles(tsc) - 56 cycles(tsc) - merge slower with cycles:19 Joint work with Alexander Duyck. [1] https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/mm/slab_bulk_test01.c [akpm@linux-foundation.org: BUG_ON -> WARN_ON;return] Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Alexander Duyck <alexander.h.duyck@redhat.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-21 07:57:49 +08:00
} while (likely(size));
}
EXPORT_SYMBOL(kmem_cache_free_bulk);
slub bulk alloc: extract objects from the per cpu slab First piece: acceleration of retrieval of per cpu objects If we are allocating lots of objects then it is advantageous to disable interrupts and avoid the this_cpu_cmpxchg() operation to get these objects faster. Note that we cannot do the fast operation if debugging is enabled, because we would have to add extra code to do all the debugging checks. And it would not be fast anyway. Note also that the requirement of having interrupts disabled avoids having to do processor flag operations. Allocate as many objects as possible in the fast way and then fall back to the generic implementation for the rest of the objects. Measurements on CPU CPU i7-4790K @ 4.00GHz Baseline normal fastpath (alloc+free cost): 42 cycles(tsc) 10.554 ns Bulk- fallback - this-patch 1 - 57 cycles(tsc) 14.432 ns - 48 cycles(tsc) 12.155 ns improved 15.8% 2 - 50 cycles(tsc) 12.746 ns - 37 cycles(tsc) 9.390 ns improved 26.0% 3 - 48 cycles(tsc) 12.180 ns - 33 cycles(tsc) 8.417 ns improved 31.2% 4 - 48 cycles(tsc) 12.015 ns - 32 cycles(tsc) 8.045 ns improved 33.3% 8 - 46 cycles(tsc) 11.526 ns - 30 cycles(tsc) 7.699 ns improved 34.8% 16 - 45 cycles(tsc) 11.418 ns - 32 cycles(tsc) 8.205 ns improved 28.9% 30 - 80 cycles(tsc) 20.246 ns - 73 cycles(tsc) 18.328 ns improved 8.8% 32 - 79 cycles(tsc) 19.946 ns - 72 cycles(tsc) 18.208 ns improved 8.9% 34 - 78 cycles(tsc) 19.659 ns - 71 cycles(tsc) 17.987 ns improved 9.0% 48 - 86 cycles(tsc) 21.516 ns - 82 cycles(tsc) 20.566 ns improved 4.7% 64 - 93 cycles(tsc) 23.423 ns - 89 cycles(tsc) 22.480 ns improved 4.3% 128 - 100 cycles(tsc) 25.170 ns - 99 cycles(tsc) 24.871 ns improved 1.0% 158 - 102 cycles(tsc) 25.549 ns - 101 cycles(tsc) 25.375 ns improved 1.0% 250 - 101 cycles(tsc) 25.344 ns - 100 cycles(tsc) 25.182 ns improved 1.0% Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-05 06:45:37 +08:00
/* Note that interrupts must be enabled when calling this function. */
int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
void **p)
{
slub bulk alloc: extract objects from the per cpu slab First piece: acceleration of retrieval of per cpu objects If we are allocating lots of objects then it is advantageous to disable interrupts and avoid the this_cpu_cmpxchg() operation to get these objects faster. Note that we cannot do the fast operation if debugging is enabled, because we would have to add extra code to do all the debugging checks. And it would not be fast anyway. Note also that the requirement of having interrupts disabled avoids having to do processor flag operations. Allocate as many objects as possible in the fast way and then fall back to the generic implementation for the rest of the objects. Measurements on CPU CPU i7-4790K @ 4.00GHz Baseline normal fastpath (alloc+free cost): 42 cycles(tsc) 10.554 ns Bulk- fallback - this-patch 1 - 57 cycles(tsc) 14.432 ns - 48 cycles(tsc) 12.155 ns improved 15.8% 2 - 50 cycles(tsc) 12.746 ns - 37 cycles(tsc) 9.390 ns improved 26.0% 3 - 48 cycles(tsc) 12.180 ns - 33 cycles(tsc) 8.417 ns improved 31.2% 4 - 48 cycles(tsc) 12.015 ns - 32 cycles(tsc) 8.045 ns improved 33.3% 8 - 46 cycles(tsc) 11.526 ns - 30 cycles(tsc) 7.699 ns improved 34.8% 16 - 45 cycles(tsc) 11.418 ns - 32 cycles(tsc) 8.205 ns improved 28.9% 30 - 80 cycles(tsc) 20.246 ns - 73 cycles(tsc) 18.328 ns improved 8.8% 32 - 79 cycles(tsc) 19.946 ns - 72 cycles(tsc) 18.208 ns improved 8.9% 34 - 78 cycles(tsc) 19.659 ns - 71 cycles(tsc) 17.987 ns improved 9.0% 48 - 86 cycles(tsc) 21.516 ns - 82 cycles(tsc) 20.566 ns improved 4.7% 64 - 93 cycles(tsc) 23.423 ns - 89 cycles(tsc) 22.480 ns improved 4.3% 128 - 100 cycles(tsc) 25.170 ns - 99 cycles(tsc) 24.871 ns improved 1.0% 158 - 102 cycles(tsc) 25.549 ns - 101 cycles(tsc) 25.375 ns improved 1.0% 250 - 101 cycles(tsc) 25.344 ns - 100 cycles(tsc) 25.182 ns improved 1.0% Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-05 06:45:37 +08:00
struct kmem_cache_cpu *c;
int i;
struct obj_cgroup *objcg = NULL;
slub bulk alloc: extract objects from the per cpu slab First piece: acceleration of retrieval of per cpu objects If we are allocating lots of objects then it is advantageous to disable interrupts and avoid the this_cpu_cmpxchg() operation to get these objects faster. Note that we cannot do the fast operation if debugging is enabled, because we would have to add extra code to do all the debugging checks. And it would not be fast anyway. Note also that the requirement of having interrupts disabled avoids having to do processor flag operations. Allocate as many objects as possible in the fast way and then fall back to the generic implementation for the rest of the objects. Measurements on CPU CPU i7-4790K @ 4.00GHz Baseline normal fastpath (alloc+free cost): 42 cycles(tsc) 10.554 ns Bulk- fallback - this-patch 1 - 57 cycles(tsc) 14.432 ns - 48 cycles(tsc) 12.155 ns improved 15.8% 2 - 50 cycles(tsc) 12.746 ns - 37 cycles(tsc) 9.390 ns improved 26.0% 3 - 48 cycles(tsc) 12.180 ns - 33 cycles(tsc) 8.417 ns improved 31.2% 4 - 48 cycles(tsc) 12.015 ns - 32 cycles(tsc) 8.045 ns improved 33.3% 8 - 46 cycles(tsc) 11.526 ns - 30 cycles(tsc) 7.699 ns improved 34.8% 16 - 45 cycles(tsc) 11.418 ns - 32 cycles(tsc) 8.205 ns improved 28.9% 30 - 80 cycles(tsc) 20.246 ns - 73 cycles(tsc) 18.328 ns improved 8.8% 32 - 79 cycles(tsc) 19.946 ns - 72 cycles(tsc) 18.208 ns improved 8.9% 34 - 78 cycles(tsc) 19.659 ns - 71 cycles(tsc) 17.987 ns improved 9.0% 48 - 86 cycles(tsc) 21.516 ns - 82 cycles(tsc) 20.566 ns improved 4.7% 64 - 93 cycles(tsc) 23.423 ns - 89 cycles(tsc) 22.480 ns improved 4.3% 128 - 100 cycles(tsc) 25.170 ns - 99 cycles(tsc) 24.871 ns improved 1.0% 158 - 102 cycles(tsc) 25.549 ns - 101 cycles(tsc) 25.375 ns improved 1.0% 250 - 101 cycles(tsc) 25.344 ns - 100 cycles(tsc) 25.182 ns improved 1.0% Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-05 06:45:37 +08:00
/* memcg and kmem_cache debug support */
s = slab_pre_alloc_hook(s, &objcg, size, flags);
if (unlikely(!s))
return false;
slub bulk alloc: extract objects from the per cpu slab First piece: acceleration of retrieval of per cpu objects If we are allocating lots of objects then it is advantageous to disable interrupts and avoid the this_cpu_cmpxchg() operation to get these objects faster. Note that we cannot do the fast operation if debugging is enabled, because we would have to add extra code to do all the debugging checks. And it would not be fast anyway. Note also that the requirement of having interrupts disabled avoids having to do processor flag operations. Allocate as many objects as possible in the fast way and then fall back to the generic implementation for the rest of the objects. Measurements on CPU CPU i7-4790K @ 4.00GHz Baseline normal fastpath (alloc+free cost): 42 cycles(tsc) 10.554 ns Bulk- fallback - this-patch 1 - 57 cycles(tsc) 14.432 ns - 48 cycles(tsc) 12.155 ns improved 15.8% 2 - 50 cycles(tsc) 12.746 ns - 37 cycles(tsc) 9.390 ns improved 26.0% 3 - 48 cycles(tsc) 12.180 ns - 33 cycles(tsc) 8.417 ns improved 31.2% 4 - 48 cycles(tsc) 12.015 ns - 32 cycles(tsc) 8.045 ns improved 33.3% 8 - 46 cycles(tsc) 11.526 ns - 30 cycles(tsc) 7.699 ns improved 34.8% 16 - 45 cycles(tsc) 11.418 ns - 32 cycles(tsc) 8.205 ns improved 28.9% 30 - 80 cycles(tsc) 20.246 ns - 73 cycles(tsc) 18.328 ns improved 8.8% 32 - 79 cycles(tsc) 19.946 ns - 72 cycles(tsc) 18.208 ns improved 8.9% 34 - 78 cycles(tsc) 19.659 ns - 71 cycles(tsc) 17.987 ns improved 9.0% 48 - 86 cycles(tsc) 21.516 ns - 82 cycles(tsc) 20.566 ns improved 4.7% 64 - 93 cycles(tsc) 23.423 ns - 89 cycles(tsc) 22.480 ns improved 4.3% 128 - 100 cycles(tsc) 25.170 ns - 99 cycles(tsc) 24.871 ns improved 1.0% 158 - 102 cycles(tsc) 25.549 ns - 101 cycles(tsc) 25.375 ns improved 1.0% 250 - 101 cycles(tsc) 25.344 ns - 100 cycles(tsc) 25.182 ns improved 1.0% Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-05 06:45:37 +08:00
/*
* Drain objects in the per cpu slab, while disabling local
* IRQs, which protects against PREEMPT and interrupts
* handlers invoking normal fastpath.
*/
local_irq_disable();
c = this_cpu_ptr(s->cpu_slab);
for (i = 0; i < size; i++) {
mm, kfence: insert KFENCE hooks for SLUB Inserts KFENCE hooks into the SLUB allocator. To pass the originally requested size to KFENCE, add an argument 'orig_size' to slab_alloc*(). The additional argument is required to preserve the requested original size for kmalloc() allocations, which uses size classes (e.g. an allocation of 272 bytes will return an object of size 512). Therefore, kmem_cache::size does not represent the kmalloc-caller's requested size, and we must introduce the argument 'orig_size' to propagate the originally requested size to KFENCE. Without the originally requested size, we would not be able to detect out-of-bounds accesses for objects placed at the end of a KFENCE object page if that object is not equal to the kmalloc-size class it was bucketed into. When KFENCE is disabled, there is no additional overhead, since slab_alloc*() functions are __always_inline. Link: https://lkml.kernel.org/r/20201103175841.3495947-6-elver@google.com Signed-off-by: Marco Elver <elver@google.com> Signed-off-by: Alexander Potapenko <glider@google.com> Reviewed-by: Dmitry Vyukov <dvyukov@google.com> Reviewed-by: Jann Horn <jannh@google.com> Co-developed-by: Marco Elver <elver@google.com> Cc: Andrey Konovalov <andreyknvl@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christopher Lameter <cl@linux.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: David Rientjes <rientjes@google.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Hillf Danton <hdanton@sina.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Joern Engel <joern@purestorage.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kees Cook <keescook@chromium.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Paul E. McKenney <paulmck@kernel.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: SeongJae Park <sjpark@amazon.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-26 09:19:16 +08:00
void *object = kfence_alloc(s, s->object_size, flags);
slub bulk alloc: extract objects from the per cpu slab First piece: acceleration of retrieval of per cpu objects If we are allocating lots of objects then it is advantageous to disable interrupts and avoid the this_cpu_cmpxchg() operation to get these objects faster. Note that we cannot do the fast operation if debugging is enabled, because we would have to add extra code to do all the debugging checks. And it would not be fast anyway. Note also that the requirement of having interrupts disabled avoids having to do processor flag operations. Allocate as many objects as possible in the fast way and then fall back to the generic implementation for the rest of the objects. Measurements on CPU CPU i7-4790K @ 4.00GHz Baseline normal fastpath (alloc+free cost): 42 cycles(tsc) 10.554 ns Bulk- fallback - this-patch 1 - 57 cycles(tsc) 14.432 ns - 48 cycles(tsc) 12.155 ns improved 15.8% 2 - 50 cycles(tsc) 12.746 ns - 37 cycles(tsc) 9.390 ns improved 26.0% 3 - 48 cycles(tsc) 12.180 ns - 33 cycles(tsc) 8.417 ns improved 31.2% 4 - 48 cycles(tsc) 12.015 ns - 32 cycles(tsc) 8.045 ns improved 33.3% 8 - 46 cycles(tsc) 11.526 ns - 30 cycles(tsc) 7.699 ns improved 34.8% 16 - 45 cycles(tsc) 11.418 ns - 32 cycles(tsc) 8.205 ns improved 28.9% 30 - 80 cycles(tsc) 20.246 ns - 73 cycles(tsc) 18.328 ns improved 8.8% 32 - 79 cycles(tsc) 19.946 ns - 72 cycles(tsc) 18.208 ns improved 8.9% 34 - 78 cycles(tsc) 19.659 ns - 71 cycles(tsc) 17.987 ns improved 9.0% 48 - 86 cycles(tsc) 21.516 ns - 82 cycles(tsc) 20.566 ns improved 4.7% 64 - 93 cycles(tsc) 23.423 ns - 89 cycles(tsc) 22.480 ns improved 4.3% 128 - 100 cycles(tsc) 25.170 ns - 99 cycles(tsc) 24.871 ns improved 1.0% 158 - 102 cycles(tsc) 25.549 ns - 101 cycles(tsc) 25.375 ns improved 1.0% 250 - 101 cycles(tsc) 25.344 ns - 100 cycles(tsc) 25.182 ns improved 1.0% Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-05 06:45:37 +08:00
mm, kfence: insert KFENCE hooks for SLUB Inserts KFENCE hooks into the SLUB allocator. To pass the originally requested size to KFENCE, add an argument 'orig_size' to slab_alloc*(). The additional argument is required to preserve the requested original size for kmalloc() allocations, which uses size classes (e.g. an allocation of 272 bytes will return an object of size 512). Therefore, kmem_cache::size does not represent the kmalloc-caller's requested size, and we must introduce the argument 'orig_size' to propagate the originally requested size to KFENCE. Without the originally requested size, we would not be able to detect out-of-bounds accesses for objects placed at the end of a KFENCE object page if that object is not equal to the kmalloc-size class it was bucketed into. When KFENCE is disabled, there is no additional overhead, since slab_alloc*() functions are __always_inline. Link: https://lkml.kernel.org/r/20201103175841.3495947-6-elver@google.com Signed-off-by: Marco Elver <elver@google.com> Signed-off-by: Alexander Potapenko <glider@google.com> Reviewed-by: Dmitry Vyukov <dvyukov@google.com> Reviewed-by: Jann Horn <jannh@google.com> Co-developed-by: Marco Elver <elver@google.com> Cc: Andrey Konovalov <andreyknvl@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christopher Lameter <cl@linux.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: David Rientjes <rientjes@google.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Hillf Danton <hdanton@sina.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Joern Engel <joern@purestorage.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kees Cook <keescook@chromium.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Paul E. McKenney <paulmck@kernel.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: SeongJae Park <sjpark@amazon.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-26 09:19:16 +08:00
if (unlikely(object)) {
p[i] = object;
continue;
}
object = c->freelist;
slub: improve bulk alloc strategy Call slowpath __slab_alloc() from within the bulk loop, as the side-effect of this call likely repopulates c->freelist. Choose to reenable local IRQs while calling slowpath. Saving some optimizations for later. E.g. it is possible to extract parts of __slab_alloc() and avoid the unnecessary and expensive (37 cycles) local_irq_{save,restore}. For now, be happy calling __slab_alloc() this lower icache impact of this func and I don't have to worry about correctness. Measurements on CPU CPU i7-4790K @ 4.00GHz Baseline normal fastpath (alloc+free cost): 42 cycles(tsc) 10.601 ns Bulk- fallback - this-patch 1 - 58 cycles(tsc) 14.516 ns - 49 cycles(tsc) 12.459 ns improved 15.5% 2 - 51 cycles(tsc) 12.930 ns - 38 cycles(tsc) 9.605 ns improved 25.5% 3 - 49 cycles(tsc) 12.274 ns - 34 cycles(tsc) 8.525 ns improved 30.6% 4 - 48 cycles(tsc) 12.058 ns - 32 cycles(tsc) 8.036 ns improved 33.3% 8 - 46 cycles(tsc) 11.609 ns - 31 cycles(tsc) 7.756 ns improved 32.6% 16 - 45 cycles(tsc) 11.451 ns - 32 cycles(tsc) 8.148 ns improved 28.9% 30 - 79 cycles(tsc) 19.865 ns - 68 cycles(tsc) 17.164 ns improved 13.9% 32 - 76 cycles(tsc) 19.212 ns - 66 cycles(tsc) 16.584 ns improved 13.2% 34 - 74 cycles(tsc) 18.600 ns - 63 cycles(tsc) 15.954 ns improved 14.9% 48 - 88 cycles(tsc) 22.092 ns - 77 cycles(tsc) 19.373 ns improved 12.5% 64 - 80 cycles(tsc) 20.043 ns - 68 cycles(tsc) 17.188 ns improved 15.0% 128 - 99 cycles(tsc) 24.818 ns - 89 cycles(tsc) 22.404 ns improved 10.1% 158 - 99 cycles(tsc) 24.977 ns - 92 cycles(tsc) 23.089 ns improved 7.1% 250 - 106 cycles(tsc) 26.552 ns - 99 cycles(tsc) 24.785 ns improved 6.6% Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-05 06:45:40 +08:00
if (unlikely(!object)) {
/*
* We may have removed an object from c->freelist using
* the fastpath in the previous iteration; in that case,
* c->tid has not been bumped yet.
* Since ___slab_alloc() may reenable interrupts while
* allocating memory, we should bump c->tid now.
*/
c->tid = next_tid(c->tid);
slub: improve bulk alloc strategy Call slowpath __slab_alloc() from within the bulk loop, as the side-effect of this call likely repopulates c->freelist. Choose to reenable local IRQs while calling slowpath. Saving some optimizations for later. E.g. it is possible to extract parts of __slab_alloc() and avoid the unnecessary and expensive (37 cycles) local_irq_{save,restore}. For now, be happy calling __slab_alloc() this lower icache impact of this func and I don't have to worry about correctness. Measurements on CPU CPU i7-4790K @ 4.00GHz Baseline normal fastpath (alloc+free cost): 42 cycles(tsc) 10.601 ns Bulk- fallback - this-patch 1 - 58 cycles(tsc) 14.516 ns - 49 cycles(tsc) 12.459 ns improved 15.5% 2 - 51 cycles(tsc) 12.930 ns - 38 cycles(tsc) 9.605 ns improved 25.5% 3 - 49 cycles(tsc) 12.274 ns - 34 cycles(tsc) 8.525 ns improved 30.6% 4 - 48 cycles(tsc) 12.058 ns - 32 cycles(tsc) 8.036 ns improved 33.3% 8 - 46 cycles(tsc) 11.609 ns - 31 cycles(tsc) 7.756 ns improved 32.6% 16 - 45 cycles(tsc) 11.451 ns - 32 cycles(tsc) 8.148 ns improved 28.9% 30 - 79 cycles(tsc) 19.865 ns - 68 cycles(tsc) 17.164 ns improved 13.9% 32 - 76 cycles(tsc) 19.212 ns - 66 cycles(tsc) 16.584 ns improved 13.2% 34 - 74 cycles(tsc) 18.600 ns - 63 cycles(tsc) 15.954 ns improved 14.9% 48 - 88 cycles(tsc) 22.092 ns - 77 cycles(tsc) 19.373 ns improved 12.5% 64 - 80 cycles(tsc) 20.043 ns - 68 cycles(tsc) 17.188 ns improved 15.0% 128 - 99 cycles(tsc) 24.818 ns - 89 cycles(tsc) 22.404 ns improved 10.1% 158 - 99 cycles(tsc) 24.977 ns - 92 cycles(tsc) 23.089 ns improved 7.1% 250 - 106 cycles(tsc) 26.552 ns - 99 cycles(tsc) 24.785 ns improved 6.6% Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-05 06:45:40 +08:00
/*
* Invoking slow path likely have side-effect
* of re-populating per CPU c->freelist
*/
p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
slub: improve bulk alloc strategy Call slowpath __slab_alloc() from within the bulk loop, as the side-effect of this call likely repopulates c->freelist. Choose to reenable local IRQs while calling slowpath. Saving some optimizations for later. E.g. it is possible to extract parts of __slab_alloc() and avoid the unnecessary and expensive (37 cycles) local_irq_{save,restore}. For now, be happy calling __slab_alloc() this lower icache impact of this func and I don't have to worry about correctness. Measurements on CPU CPU i7-4790K @ 4.00GHz Baseline normal fastpath (alloc+free cost): 42 cycles(tsc) 10.601 ns Bulk- fallback - this-patch 1 - 58 cycles(tsc) 14.516 ns - 49 cycles(tsc) 12.459 ns improved 15.5% 2 - 51 cycles(tsc) 12.930 ns - 38 cycles(tsc) 9.605 ns improved 25.5% 3 - 49 cycles(tsc) 12.274 ns - 34 cycles(tsc) 8.525 ns improved 30.6% 4 - 48 cycles(tsc) 12.058 ns - 32 cycles(tsc) 8.036 ns improved 33.3% 8 - 46 cycles(tsc) 11.609 ns - 31 cycles(tsc) 7.756 ns improved 32.6% 16 - 45 cycles(tsc) 11.451 ns - 32 cycles(tsc) 8.148 ns improved 28.9% 30 - 79 cycles(tsc) 19.865 ns - 68 cycles(tsc) 17.164 ns improved 13.9% 32 - 76 cycles(tsc) 19.212 ns - 66 cycles(tsc) 16.584 ns improved 13.2% 34 - 74 cycles(tsc) 18.600 ns - 63 cycles(tsc) 15.954 ns improved 14.9% 48 - 88 cycles(tsc) 22.092 ns - 77 cycles(tsc) 19.373 ns improved 12.5% 64 - 80 cycles(tsc) 20.043 ns - 68 cycles(tsc) 17.188 ns improved 15.0% 128 - 99 cycles(tsc) 24.818 ns - 89 cycles(tsc) 22.404 ns improved 10.1% 158 - 99 cycles(tsc) 24.977 ns - 92 cycles(tsc) 23.089 ns improved 7.1% 250 - 106 cycles(tsc) 26.552 ns - 99 cycles(tsc) 24.785 ns improved 6.6% Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-05 06:45:40 +08:00
_RET_IP_, c);
if (unlikely(!p[i]))
goto error;
slub: improve bulk alloc strategy Call slowpath __slab_alloc() from within the bulk loop, as the side-effect of this call likely repopulates c->freelist. Choose to reenable local IRQs while calling slowpath. Saving some optimizations for later. E.g. it is possible to extract parts of __slab_alloc() and avoid the unnecessary and expensive (37 cycles) local_irq_{save,restore}. For now, be happy calling __slab_alloc() this lower icache impact of this func and I don't have to worry about correctness. Measurements on CPU CPU i7-4790K @ 4.00GHz Baseline normal fastpath (alloc+free cost): 42 cycles(tsc) 10.601 ns Bulk- fallback - this-patch 1 - 58 cycles(tsc) 14.516 ns - 49 cycles(tsc) 12.459 ns improved 15.5% 2 - 51 cycles(tsc) 12.930 ns - 38 cycles(tsc) 9.605 ns improved 25.5% 3 - 49 cycles(tsc) 12.274 ns - 34 cycles(tsc) 8.525 ns improved 30.6% 4 - 48 cycles(tsc) 12.058 ns - 32 cycles(tsc) 8.036 ns improved 33.3% 8 - 46 cycles(tsc) 11.609 ns - 31 cycles(tsc) 7.756 ns improved 32.6% 16 - 45 cycles(tsc) 11.451 ns - 32 cycles(tsc) 8.148 ns improved 28.9% 30 - 79 cycles(tsc) 19.865 ns - 68 cycles(tsc) 17.164 ns improved 13.9% 32 - 76 cycles(tsc) 19.212 ns - 66 cycles(tsc) 16.584 ns improved 13.2% 34 - 74 cycles(tsc) 18.600 ns - 63 cycles(tsc) 15.954 ns improved 14.9% 48 - 88 cycles(tsc) 22.092 ns - 77 cycles(tsc) 19.373 ns improved 12.5% 64 - 80 cycles(tsc) 20.043 ns - 68 cycles(tsc) 17.188 ns improved 15.0% 128 - 99 cycles(tsc) 24.818 ns - 89 cycles(tsc) 22.404 ns improved 10.1% 158 - 99 cycles(tsc) 24.977 ns - 92 cycles(tsc) 23.089 ns improved 7.1% 250 - 106 cycles(tsc) 26.552 ns - 99 cycles(tsc) 24.785 ns improved 6.6% Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-05 06:45:40 +08:00
c = this_cpu_ptr(s->cpu_slab);
maybe_wipe_obj_freeptr(s, p[i]);
slub: improve bulk alloc strategy Call slowpath __slab_alloc() from within the bulk loop, as the side-effect of this call likely repopulates c->freelist. Choose to reenable local IRQs while calling slowpath. Saving some optimizations for later. E.g. it is possible to extract parts of __slab_alloc() and avoid the unnecessary and expensive (37 cycles) local_irq_{save,restore}. For now, be happy calling __slab_alloc() this lower icache impact of this func and I don't have to worry about correctness. Measurements on CPU CPU i7-4790K @ 4.00GHz Baseline normal fastpath (alloc+free cost): 42 cycles(tsc) 10.601 ns Bulk- fallback - this-patch 1 - 58 cycles(tsc) 14.516 ns - 49 cycles(tsc) 12.459 ns improved 15.5% 2 - 51 cycles(tsc) 12.930 ns - 38 cycles(tsc) 9.605 ns improved 25.5% 3 - 49 cycles(tsc) 12.274 ns - 34 cycles(tsc) 8.525 ns improved 30.6% 4 - 48 cycles(tsc) 12.058 ns - 32 cycles(tsc) 8.036 ns improved 33.3% 8 - 46 cycles(tsc) 11.609 ns - 31 cycles(tsc) 7.756 ns improved 32.6% 16 - 45 cycles(tsc) 11.451 ns - 32 cycles(tsc) 8.148 ns improved 28.9% 30 - 79 cycles(tsc) 19.865 ns - 68 cycles(tsc) 17.164 ns improved 13.9% 32 - 76 cycles(tsc) 19.212 ns - 66 cycles(tsc) 16.584 ns improved 13.2% 34 - 74 cycles(tsc) 18.600 ns - 63 cycles(tsc) 15.954 ns improved 14.9% 48 - 88 cycles(tsc) 22.092 ns - 77 cycles(tsc) 19.373 ns improved 12.5% 64 - 80 cycles(tsc) 20.043 ns - 68 cycles(tsc) 17.188 ns improved 15.0% 128 - 99 cycles(tsc) 24.818 ns - 89 cycles(tsc) 22.404 ns improved 10.1% 158 - 99 cycles(tsc) 24.977 ns - 92 cycles(tsc) 23.089 ns improved 7.1% 250 - 106 cycles(tsc) 26.552 ns - 99 cycles(tsc) 24.785 ns improved 6.6% Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-05 06:45:40 +08:00
continue; /* goto for-loop */
}
slub bulk alloc: extract objects from the per cpu slab First piece: acceleration of retrieval of per cpu objects If we are allocating lots of objects then it is advantageous to disable interrupts and avoid the this_cpu_cmpxchg() operation to get these objects faster. Note that we cannot do the fast operation if debugging is enabled, because we would have to add extra code to do all the debugging checks. And it would not be fast anyway. Note also that the requirement of having interrupts disabled avoids having to do processor flag operations. Allocate as many objects as possible in the fast way and then fall back to the generic implementation for the rest of the objects. Measurements on CPU CPU i7-4790K @ 4.00GHz Baseline normal fastpath (alloc+free cost): 42 cycles(tsc) 10.554 ns Bulk- fallback - this-patch 1 - 57 cycles(tsc) 14.432 ns - 48 cycles(tsc) 12.155 ns improved 15.8% 2 - 50 cycles(tsc) 12.746 ns - 37 cycles(tsc) 9.390 ns improved 26.0% 3 - 48 cycles(tsc) 12.180 ns - 33 cycles(tsc) 8.417 ns improved 31.2% 4 - 48 cycles(tsc) 12.015 ns - 32 cycles(tsc) 8.045 ns improved 33.3% 8 - 46 cycles(tsc) 11.526 ns - 30 cycles(tsc) 7.699 ns improved 34.8% 16 - 45 cycles(tsc) 11.418 ns - 32 cycles(tsc) 8.205 ns improved 28.9% 30 - 80 cycles(tsc) 20.246 ns - 73 cycles(tsc) 18.328 ns improved 8.8% 32 - 79 cycles(tsc) 19.946 ns - 72 cycles(tsc) 18.208 ns improved 8.9% 34 - 78 cycles(tsc) 19.659 ns - 71 cycles(tsc) 17.987 ns improved 9.0% 48 - 86 cycles(tsc) 21.516 ns - 82 cycles(tsc) 20.566 ns improved 4.7% 64 - 93 cycles(tsc) 23.423 ns - 89 cycles(tsc) 22.480 ns improved 4.3% 128 - 100 cycles(tsc) 25.170 ns - 99 cycles(tsc) 24.871 ns improved 1.0% 158 - 102 cycles(tsc) 25.549 ns - 101 cycles(tsc) 25.375 ns improved 1.0% 250 - 101 cycles(tsc) 25.344 ns - 100 cycles(tsc) 25.182 ns improved 1.0% Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-05 06:45:37 +08:00
c->freelist = get_freepointer(s, object);
p[i] = object;
maybe_wipe_obj_freeptr(s, p[i]);
slub bulk alloc: extract objects from the per cpu slab First piece: acceleration of retrieval of per cpu objects If we are allocating lots of objects then it is advantageous to disable interrupts and avoid the this_cpu_cmpxchg() operation to get these objects faster. Note that we cannot do the fast operation if debugging is enabled, because we would have to add extra code to do all the debugging checks. And it would not be fast anyway. Note also that the requirement of having interrupts disabled avoids having to do processor flag operations. Allocate as many objects as possible in the fast way and then fall back to the generic implementation for the rest of the objects. Measurements on CPU CPU i7-4790K @ 4.00GHz Baseline normal fastpath (alloc+free cost): 42 cycles(tsc) 10.554 ns Bulk- fallback - this-patch 1 - 57 cycles(tsc) 14.432 ns - 48 cycles(tsc) 12.155 ns improved 15.8% 2 - 50 cycles(tsc) 12.746 ns - 37 cycles(tsc) 9.390 ns improved 26.0% 3 - 48 cycles(tsc) 12.180 ns - 33 cycles(tsc) 8.417 ns improved 31.2% 4 - 48 cycles(tsc) 12.015 ns - 32 cycles(tsc) 8.045 ns improved 33.3% 8 - 46 cycles(tsc) 11.526 ns - 30 cycles(tsc) 7.699 ns improved 34.8% 16 - 45 cycles(tsc) 11.418 ns - 32 cycles(tsc) 8.205 ns improved 28.9% 30 - 80 cycles(tsc) 20.246 ns - 73 cycles(tsc) 18.328 ns improved 8.8% 32 - 79 cycles(tsc) 19.946 ns - 72 cycles(tsc) 18.208 ns improved 8.9% 34 - 78 cycles(tsc) 19.659 ns - 71 cycles(tsc) 17.987 ns improved 9.0% 48 - 86 cycles(tsc) 21.516 ns - 82 cycles(tsc) 20.566 ns improved 4.7% 64 - 93 cycles(tsc) 23.423 ns - 89 cycles(tsc) 22.480 ns improved 4.3% 128 - 100 cycles(tsc) 25.170 ns - 99 cycles(tsc) 24.871 ns improved 1.0% 158 - 102 cycles(tsc) 25.549 ns - 101 cycles(tsc) 25.375 ns improved 1.0% 250 - 101 cycles(tsc) 25.344 ns - 100 cycles(tsc) 25.182 ns improved 1.0% Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-05 06:45:37 +08:00
}
c->tid = next_tid(c->tid);
local_irq_enable();
kasan, mm: integrate slab init_on_alloc with HW_TAGS This change uses the previously added memory initialization feature of HW_TAGS KASAN routines for slab memory when init_on_alloc is enabled. With this change, memory initialization memset() is no longer called when both HW_TAGS KASAN and init_on_alloc are enabled. Instead, memory is initialized in KASAN runtime. The memory initialization memset() is moved into slab_post_alloc_hook() that currently directly follows the initialization loop. A new argument is added to slab_post_alloc_hook() that indicates whether to initialize the memory or not. To avoid discrepancies with which memory gets initialized that can be caused by future changes, both KASAN hook and initialization memset() are put together and a warning comment is added. Combining setting allocation tags with memory initialization improves HW_TAGS KASAN performance when init_on_alloc is enabled. Link: https://lkml.kernel.org/r/c1292aeb5d519da221ec74a0684a949b027d7720.1615296150.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Reviewed-by: Marco Elver <elver@google.com> Cc: Alexander Potapenko <glider@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Branislav Rankov <Branislav.Rankov@arm.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christoph Lameter <cl@linux.com> Cc: David Rientjes <rientjes@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Evgenii Stepanov <eugenis@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kevin Brodsky <kevin.brodsky@arm.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Collingbourne <pcc@google.com> Cc: Vincenzo Frascino <vincenzo.frascino@arm.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-30 14:00:06 +08:00
/*
* memcg and kmem_cache debug support and memory initialization.
* Done outside of the IRQ disabled fastpath loop.
*/
slab_post_alloc_hook(s, objcg, flags, size, p,
slab_want_init_on_alloc(flags, s));
return i;
error:
local_irq_enable();
kasan, mm: integrate slab init_on_alloc with HW_TAGS This change uses the previously added memory initialization feature of HW_TAGS KASAN routines for slab memory when init_on_alloc is enabled. With this change, memory initialization memset() is no longer called when both HW_TAGS KASAN and init_on_alloc are enabled. Instead, memory is initialized in KASAN runtime. The memory initialization memset() is moved into slab_post_alloc_hook() that currently directly follows the initialization loop. A new argument is added to slab_post_alloc_hook() that indicates whether to initialize the memory or not. To avoid discrepancies with which memory gets initialized that can be caused by future changes, both KASAN hook and initialization memset() are put together and a warning comment is added. Combining setting allocation tags with memory initialization improves HW_TAGS KASAN performance when init_on_alloc is enabled. Link: https://lkml.kernel.org/r/c1292aeb5d519da221ec74a0684a949b027d7720.1615296150.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Reviewed-by: Marco Elver <elver@google.com> Cc: Alexander Potapenko <glider@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Branislav Rankov <Branislav.Rankov@arm.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christoph Lameter <cl@linux.com> Cc: David Rientjes <rientjes@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Evgenii Stepanov <eugenis@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kevin Brodsky <kevin.brodsky@arm.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Collingbourne <pcc@google.com> Cc: Vincenzo Frascino <vincenzo.frascino@arm.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-30 14:00:06 +08:00
slab_post_alloc_hook(s, objcg, flags, i, p, false);
__kmem_cache_free_bulk(s, i, p);
return 0;
}
EXPORT_SYMBOL(kmem_cache_alloc_bulk);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/*
* Object placement in a slab is made very easy because we always start at
* offset 0. If we tune the size of the object to the alignment then we can
* get the required alignment by putting one properly sized object after
* another.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*
* Notice that the allocation order determines the sizes of the per cpu
* caches. Each processor has always one slab available for allocations.
* Increasing the allocation order reduces the number of times that slabs
* must be moved on and off the partial lists and is therefore a factor in
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
* locking overhead.
*/
/*
* Minimum / Maximum order of slab pages. This influences locking overhead
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
* and slab fragmentation. A higher order reduces the number of partial slabs
* and increases the number of allocations possible without having to
* take the list_lock.
*/
static unsigned int slub_min_order;
static unsigned int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
static unsigned int slub_min_objects;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/*
* Calculate the order of allocation given an slab object size.
*
* The order of allocation has significant impact on performance and other
* system components. Generally order 0 allocations should be preferred since
* order 0 does not cause fragmentation in the page allocator. Larger objects
* be problematic to put into order 0 slabs because there may be too much
* unused space left. We go to a higher order if more than 1/16th of the slab
* would be wasted.
*
* In order to reach satisfactory performance we must ensure that a minimum
* number of objects is in one slab. Otherwise we may generate too much
* activity on the partial lists which requires taking the list_lock. This is
* less a concern for large slabs though which are rarely used.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*
* slub_max_order specifies the order where we begin to stop considering the
* number of objects in a slab as critical. If we reach slub_max_order then
* we try to keep the page order as low as possible. So we accept more waste
* of space in favor of a small page order.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*
* Higher order allocations also allow the placement of more objects in a
* slab and thereby reduce object handling overhead. If the user has
* requested a higher minimum order then we start with that one instead of
* the smallest order which will fit the object.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*/
static inline unsigned int slab_order(unsigned int size,
unsigned int min_objects, unsigned int max_order,
unsigned int fract_leftover)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
{
unsigned int min_order = slub_min_order;
unsigned int order;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
return get_order(size * MAX_OBJS_PER_PAGE) - 1;
for (order = max(min_order, (unsigned int)get_order(min_objects * size));
order <= max_order; order++) {
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
unsigned int rem;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
rem = slab_size % size;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (rem <= slab_size / fract_leftover)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
break;
}
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
return order;
}
static inline int calculate_order(unsigned int size)
{
unsigned int order;
unsigned int min_objects;
unsigned int max_objects;
mm, slub: better heuristic for number of cpus when calculating slab order When creating a new kmem cache, SLUB determines how large the slab pages will based on number of inputs, including the number of CPUs in the system. Larger slab pages mean that more objects can be allocated/free from per-cpu slabs before accessing shared structures, but also potentially more memory can be wasted due to low slab usage and fragmentation. The rough idea of using number of CPUs is that larger systems will be more likely to benefit from reduced contention, and also should have enough memory to spare. Number of CPUs used to be determined as nr_cpu_ids, which is number of possible cpus, but on some systems many will never be onlined, thus commit 045ab8c9487b ("mm/slub: let number of online CPUs determine the slub page order") changed it to nr_online_cpus(). However, for kmem caches created early before CPUs are onlined, this may lead to permamently low slab page sizes. Vincent reports a regression [1] of hackbench on arm64 systems: "I'm facing significant performances regression on a large arm64 server system (224 CPUs). Regressions is also present on small arm64 system (8 CPUs) but in a far smaller order of magnitude On 224 CPUs system : 9 iterations of hackbench -l 16000 -g 16 v5.11-rc4 : 9.135sec (+/- 0.45%) v5.11-rc4 + revert this patch: 3.173sec (+/- 0.48%) v5.10: 3.136sec (+/- 0.40%)" Mel reports a regression [2] of hackbench on x86_64, with lockstat suggesting page allocator contention: "i.e. the patch incurs a 7% to 32% performance penalty. This bisected cleanly yesterday when I was looking for the regression and then found the thread. Numerous caches change size. For example, kmalloc-512 goes from order-0 (vanilla) to order-2 with the revert. So mostly this is down to the number of times SLUB calls into the page allocator which only caches order-0 pages on a per-cpu basis" Clearly num_online_cpus() doesn't work too early in bootup. We could change the order dynamically in a memory hotplug callback, but runtime order changing for existing kmem caches has been already shown as dangerous, and removed in 32a6f409b693 ("mm, slub: remove runtime allocation order changes"). It could be resurrected in a safe manner with some effort, but to fix the regression we need something simpler. We could use num_present_cpus() that should be the number of physically present CPUs even before they are onlined. That would work for PowerPC [3], which triggered the original commit, but that still doesn't work on arm64 [4] as explained in [5]. So this patch tries to determine the best available value without specific arch knowledge. - num_present_cpus() if the number is larger than 1, as that means the arch is likely setting it properly - nr_cpu_ids otherwise This should fix the reported regressions while also keeping the effect of 045ab8c9487b for PowerPC systems. It's possible there are configurations where num_present_cpus() is 1 during boot while nr_cpu_ids is at the same time bloated, so these (if they exist) would keep the large orders based on nr_cpu_ids as was before 045ab8c9487b. [1] https://lore.kernel.org/linux-mm/CAKfTPtA_JgMf_+zdFbcb_V9rM7JBWNPjAz9irgwFj7Rou=xzZg@mail.gmail.com/ [2] https://lore.kernel.org/linux-mm/20210128134512.GF3592@techsingularity.net/ [3] https://lore.kernel.org/linux-mm/20210123051607.GC2587010@in.ibm.com/ [4] https://lore.kernel.org/linux-mm/CAKfTPtAjyVmS5VYvU6DBxg4-JEo5bdmWbngf-03YsY18cmWv_g@mail.gmail.com/ [5] https://lore.kernel.org/linux-mm/20210126230305.GD30941@willie-the-truck/ Link: https://lkml.kernel.org/r/20210208134108.22286-1-vbabka@suse.cz Fixes: 045ab8c9487b ("mm/slub: let number of online CPUs determine the slub page order") Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Reported-by: Vincent Guittot <vincent.guittot@linaro.org> Reported-by: Mel Gorman <mgorman@techsingularity.net> Tested-by: Mel Gorman <mgorman@techsingularity.net> Tested-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com> Cc: Bharata B Rao <bharata@linux.ibm.com> Cc: Christoph Lameter <cl@linux.com> Cc: Roman Gushchin <guro@fb.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Jann Horn <jannh@google.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Shakeel Butt <shakeelb@google.com> Cc: Will Deacon <will@kernel.org> Cc: <stable@vger.kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-10 05:42:32 +08:00
unsigned int nr_cpus;
/*
* Attempt to find best configuration for a slab. This
* works by first attempting to generate a layout with
* the best configuration and backing off gradually.
*
* First we increase the acceptable waste in a slab. Then
* we reduce the minimum objects required in a slab.
*/
min_objects = slub_min_objects;
mm, slub: better heuristic for number of cpus when calculating slab order When creating a new kmem cache, SLUB determines how large the slab pages will based on number of inputs, including the number of CPUs in the system. Larger slab pages mean that more objects can be allocated/free from per-cpu slabs before accessing shared structures, but also potentially more memory can be wasted due to low slab usage and fragmentation. The rough idea of using number of CPUs is that larger systems will be more likely to benefit from reduced contention, and also should have enough memory to spare. Number of CPUs used to be determined as nr_cpu_ids, which is number of possible cpus, but on some systems many will never be onlined, thus commit 045ab8c9487b ("mm/slub: let number of online CPUs determine the slub page order") changed it to nr_online_cpus(). However, for kmem caches created early before CPUs are onlined, this may lead to permamently low slab page sizes. Vincent reports a regression [1] of hackbench on arm64 systems: "I'm facing significant performances regression on a large arm64 server system (224 CPUs). Regressions is also present on small arm64 system (8 CPUs) but in a far smaller order of magnitude On 224 CPUs system : 9 iterations of hackbench -l 16000 -g 16 v5.11-rc4 : 9.135sec (+/- 0.45%) v5.11-rc4 + revert this patch: 3.173sec (+/- 0.48%) v5.10: 3.136sec (+/- 0.40%)" Mel reports a regression [2] of hackbench on x86_64, with lockstat suggesting page allocator contention: "i.e. the patch incurs a 7% to 32% performance penalty. This bisected cleanly yesterday when I was looking for the regression and then found the thread. Numerous caches change size. For example, kmalloc-512 goes from order-0 (vanilla) to order-2 with the revert. So mostly this is down to the number of times SLUB calls into the page allocator which only caches order-0 pages on a per-cpu basis" Clearly num_online_cpus() doesn't work too early in bootup. We could change the order dynamically in a memory hotplug callback, but runtime order changing for existing kmem caches has been already shown as dangerous, and removed in 32a6f409b693 ("mm, slub: remove runtime allocation order changes"). It could be resurrected in a safe manner with some effort, but to fix the regression we need something simpler. We could use num_present_cpus() that should be the number of physically present CPUs even before they are onlined. That would work for PowerPC [3], which triggered the original commit, but that still doesn't work on arm64 [4] as explained in [5]. So this patch tries to determine the best available value without specific arch knowledge. - num_present_cpus() if the number is larger than 1, as that means the arch is likely setting it properly - nr_cpu_ids otherwise This should fix the reported regressions while also keeping the effect of 045ab8c9487b for PowerPC systems. It's possible there are configurations where num_present_cpus() is 1 during boot while nr_cpu_ids is at the same time bloated, so these (if they exist) would keep the large orders based on nr_cpu_ids as was before 045ab8c9487b. [1] https://lore.kernel.org/linux-mm/CAKfTPtA_JgMf_+zdFbcb_V9rM7JBWNPjAz9irgwFj7Rou=xzZg@mail.gmail.com/ [2] https://lore.kernel.org/linux-mm/20210128134512.GF3592@techsingularity.net/ [3] https://lore.kernel.org/linux-mm/20210123051607.GC2587010@in.ibm.com/ [4] https://lore.kernel.org/linux-mm/CAKfTPtAjyVmS5VYvU6DBxg4-JEo5bdmWbngf-03YsY18cmWv_g@mail.gmail.com/ [5] https://lore.kernel.org/linux-mm/20210126230305.GD30941@willie-the-truck/ Link: https://lkml.kernel.org/r/20210208134108.22286-1-vbabka@suse.cz Fixes: 045ab8c9487b ("mm/slub: let number of online CPUs determine the slub page order") Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Reported-by: Vincent Guittot <vincent.guittot@linaro.org> Reported-by: Mel Gorman <mgorman@techsingularity.net> Tested-by: Mel Gorman <mgorman@techsingularity.net> Tested-by: Vincent Guittot <vincent.guittot@linaro.org> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com> Cc: Bharata B Rao <bharata@linux.ibm.com> Cc: Christoph Lameter <cl@linux.com> Cc: Roman Gushchin <guro@fb.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Jann Horn <jannh@google.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Shakeel Butt <shakeelb@google.com> Cc: Will Deacon <will@kernel.org> Cc: <stable@vger.kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-10 05:42:32 +08:00
if (!min_objects) {
/*
* Some architectures will only update present cpus when
* onlining them, so don't trust the number if it's just 1. But
* we also don't want to use nr_cpu_ids always, as on some other
* architectures, there can be many possible cpus, but never
* onlined. Here we compromise between trying to avoid too high
* order on systems that appear larger than they are, and too
* low order on systems that appear smaller than they are.
*/
nr_cpus = num_present_cpus();
if (nr_cpus <= 1)
nr_cpus = nr_cpu_ids;
min_objects = 4 * (fls(nr_cpus) + 1);
}
max_objects = order_objects(slub_max_order, size);
min_objects = min(min_objects, max_objects);
while (min_objects > 1) {
unsigned int fraction;
fraction = 16;
while (fraction >= 4) {
order = slab_order(size, min_objects,
slub_max_order, fraction);
if (order <= slub_max_order)
return order;
fraction /= 2;
}
min_objects--;
}
/*
* We were unable to place multiple objects in a slab. Now
* lets see if we can place a single object there.
*/
order = slab_order(size, 1, slub_max_order, 1);
if (order <= slub_max_order)
return order;
/*
* Doh this slab cannot be placed using slub_max_order.
*/
order = slab_order(size, 1, MAX_ORDER, 1);
if (order < MAX_ORDER)
return order;
return -ENOSYS;
}
SLUB: dynamic per-cache MIN_PARTIAL This patch changes the static MIN_PARTIAL to a dynamic per-cache ->min_partial value that is calculated from object size. The bigger the object size, the more pages we keep on the partial list. I tested SLAB, SLUB, and SLUB with this patch on Jens Axboe's 'netio' example script of the fio benchmarking tool. The script stresses the networking subsystem which should also give a fairly good beating of kmalloc() et al. To run the test yourself, first clone the fio repository: git clone git://git.kernel.dk/fio.git and then run the following command n times on your machine: time ./fio examples/netio The results on my 2-way 64-bit x86 machine are as follows: [ the minimum, maximum, and average are captured from 50 individual runs ] real time (seconds) min max avg sd SLAB 22.76 23.38 22.98 0.17 SLUB 22.80 25.78 23.46 0.72 SLUB (dynamic) 22.74 23.54 23.00 0.20 sys time (seconds) min max avg sd SLAB 6.90 8.28 7.70 0.28 SLUB 7.42 16.95 8.89 2.28 SLUB (dynamic) 7.17 8.64 7.73 0.29 user time (seconds) min max avg sd SLAB 36.89 38.11 37.50 0.29 SLUB 30.85 37.99 37.06 1.67 SLUB (dynamic) 36.75 38.07 37.59 0.32 As you can see from the above numbers, this patch brings SLUB to the same level as SLAB for this particular workload fixing a ~2% regression. I'd expect this change to help similar workloads that allocate a lot of objects that are close to the size of a page. Cc: Matthew Wilcox <matthew@wil.cx> Cc: Andrew Morton <akpm@linux-foundation.org> Acked-by: Christoph Lameter <cl@linux-foundation.org> Signed-off-by: Pekka Enberg <penberg@cs.helsinki.fi>
2008-08-05 14:28:47 +08:00
static void
init_kmem_cache_node(struct kmem_cache_node *n)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
{
n->nr_partial = 0;
spin_lock_init(&n->list_lock);
INIT_LIST_HEAD(&n->partial);
#ifdef CONFIG_SLUB_DEBUG
atomic_long_set(&n->nr_slabs, 0);
atomic_long_set(&n->total_objects, 0);
INIT_LIST_HEAD(&n->full);
#endif
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
SLUB: Place kmem_cache_cpu structures in a NUMA aware way The kmem_cache_cpu structures introduced are currently an array placed in the kmem_cache struct. Meaning the kmem_cache_cpu structures are overwhelmingly on the wrong node for systems with a higher amount of nodes. These are performance critical structures since the per node information has to be touched for every alloc and free in a slab. In order to place the kmem_cache_cpu structure optimally we put an array of pointers to kmem_cache_cpu structs in kmem_cache (similar to SLAB). However, the kmem_cache_cpu structures can now be allocated in a more intelligent way. We would like to put per cpu structures for the same cpu but different slab caches in cachelines together to save space and decrease the cache footprint. However, the slab allocators itself control only allocations per node. We set up a simple per cpu array for every processor with 100 per cpu structures which is usually enough to get them all set up right. If we run out then we fall back to kmalloc_node. This also solves the bootstrap problem since we do not have to use slab allocator functions early in boot to get memory for the small per cpu structures. Pro: - NUMA aware placement improves memory performance - All global structures in struct kmem_cache become readonly - Dense packing of per cpu structures reduces cacheline footprint in SMP and NUMA. - Potential avoidance of exclusive cacheline fetches on the free and alloc hotpath since multiple kmem_cache_cpu structures are in one cacheline. This is particularly important for the kmalloc array. Cons: - Additional reference to one read only cacheline (per cpu array of pointers to kmem_cache_cpu) in both slab_alloc() and slab_free(). [akinobu.mita@gmail.com: fix cpu hotplug offline/online path] Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: "Pekka Enberg" <penberg@cs.helsinki.fi> Cc: Akinobu Mita <akinobu.mita@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 16:26:08 +08:00
{
BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
SLUB: Place kmem_cache_cpu structures in a NUMA aware way The kmem_cache_cpu structures introduced are currently an array placed in the kmem_cache struct. Meaning the kmem_cache_cpu structures are overwhelmingly on the wrong node for systems with a higher amount of nodes. These are performance critical structures since the per node information has to be touched for every alloc and free in a slab. In order to place the kmem_cache_cpu structure optimally we put an array of pointers to kmem_cache_cpu structs in kmem_cache (similar to SLAB). However, the kmem_cache_cpu structures can now be allocated in a more intelligent way. We would like to put per cpu structures for the same cpu but different slab caches in cachelines together to save space and decrease the cache footprint. However, the slab allocators itself control only allocations per node. We set up a simple per cpu array for every processor with 100 per cpu structures which is usually enough to get them all set up right. If we run out then we fall back to kmalloc_node. This also solves the bootstrap problem since we do not have to use slab allocator functions early in boot to get memory for the small per cpu structures. Pro: - NUMA aware placement improves memory performance - All global structures in struct kmem_cache become readonly - Dense packing of per cpu structures reduces cacheline footprint in SMP and NUMA. - Potential avoidance of exclusive cacheline fetches on the free and alloc hotpath since multiple kmem_cache_cpu structures are in one cacheline. This is particularly important for the kmalloc array. Cons: - Additional reference to one read only cacheline (per cpu array of pointers to kmem_cache_cpu) in both slab_alloc() and slab_free(). [akinobu.mita@gmail.com: fix cpu hotplug offline/online path] Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: "Pekka Enberg" <penberg@cs.helsinki.fi> Cc: Akinobu Mita <akinobu.mita@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 16:26:08 +08:00
Lockless (and preemptless) fastpaths for slub Use the this_cpu_cmpxchg_double functionality to implement a lockless allocation algorithm on arches that support fast this_cpu_ops. Each of the per cpu pointers is paired with a transaction id that ensures that updates of the per cpu information can only occur in sequence on a certain cpu. A transaction id is a "long" integer that is comprised of an event number and the cpu number. The event number is incremented for every change to the per cpu state. This means that the cmpxchg instruction can verify for an update that nothing interfered and that we are updating the percpu structure for the processor where we picked up the information and that we are also currently on that processor when we update the information. This results in a significant decrease of the overhead in the fastpaths. It also makes it easy to adopt the fast path for realtime kernels since this is lockless and does not require the use of the current per cpu area over the critical section. It is only important that the per cpu area is current at the beginning of the critical section and at the end. So there is no need even to disable preemption. Test results show that the fastpath cycle count is reduced by up to ~ 40% (alloc/free test goes from ~140 cycles down to ~80). The slowpath for kfree adds a few cycles. Sadly this does nothing for the slowpath which is where the main issues with performance in slub are but the best case performance rises significantly. (For that see the more complex slub patches that require cmpxchg_double) Kmalloc: alloc/free test Before: 10000 times kmalloc(8)/kfree -> 134 cycles 10000 times kmalloc(16)/kfree -> 152 cycles 10000 times kmalloc(32)/kfree -> 144 cycles 10000 times kmalloc(64)/kfree -> 142 cycles 10000 times kmalloc(128)/kfree -> 142 cycles 10000 times kmalloc(256)/kfree -> 132 cycles 10000 times kmalloc(512)/kfree -> 132 cycles 10000 times kmalloc(1024)/kfree -> 135 cycles 10000 times kmalloc(2048)/kfree -> 135 cycles 10000 times kmalloc(4096)/kfree -> 135 cycles 10000 times kmalloc(8192)/kfree -> 144 cycles 10000 times kmalloc(16384)/kfree -> 754 cycles After: 10000 times kmalloc(8)/kfree -> 78 cycles 10000 times kmalloc(16)/kfree -> 78 cycles 10000 times kmalloc(32)/kfree -> 82 cycles 10000 times kmalloc(64)/kfree -> 88 cycles 10000 times kmalloc(128)/kfree -> 79 cycles 10000 times kmalloc(256)/kfree -> 79 cycles 10000 times kmalloc(512)/kfree -> 85 cycles 10000 times kmalloc(1024)/kfree -> 82 cycles 10000 times kmalloc(2048)/kfree -> 82 cycles 10000 times kmalloc(4096)/kfree -> 85 cycles 10000 times kmalloc(8192)/kfree -> 82 cycles 10000 times kmalloc(16384)/kfree -> 706 cycles Kmalloc: Repeatedly allocate then free test Before: 10000 times kmalloc(8) -> 211 cycles kfree -> 113 cycles 10000 times kmalloc(16) -> 174 cycles kfree -> 115 cycles 10000 times kmalloc(32) -> 235 cycles kfree -> 129 cycles 10000 times kmalloc(64) -> 222 cycles kfree -> 120 cycles 10000 times kmalloc(128) -> 343 cycles kfree -> 139 cycles 10000 times kmalloc(256) -> 827 cycles kfree -> 147 cycles 10000 times kmalloc(512) -> 1048 cycles kfree -> 272 cycles 10000 times kmalloc(1024) -> 2043 cycles kfree -> 528 cycles 10000 times kmalloc(2048) -> 4002 cycles kfree -> 571 cycles 10000 times kmalloc(4096) -> 7740 cycles kfree -> 628 cycles 10000 times kmalloc(8192) -> 8062 cycles kfree -> 850 cycles 10000 times kmalloc(16384) -> 8895 cycles kfree -> 1249 cycles After: 10000 times kmalloc(8) -> 190 cycles kfree -> 129 cycles 10000 times kmalloc(16) -> 76 cycles kfree -> 123 cycles 10000 times kmalloc(32) -> 126 cycles kfree -> 124 cycles 10000 times kmalloc(64) -> 181 cycles kfree -> 128 cycles 10000 times kmalloc(128) -> 310 cycles kfree -> 140 cycles 10000 times kmalloc(256) -> 809 cycles kfree -> 165 cycles 10000 times kmalloc(512) -> 1005 cycles kfree -> 269 cycles 10000 times kmalloc(1024) -> 1999 cycles kfree -> 527 cycles 10000 times kmalloc(2048) -> 3967 cycles kfree -> 570 cycles 10000 times kmalloc(4096) -> 7658 cycles kfree -> 637 cycles 10000 times kmalloc(8192) -> 8111 cycles kfree -> 859 cycles 10000 times kmalloc(16384) -> 8791 cycles kfree -> 1173 cycles Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-02-26 01:38:54 +08:00
/*
* Must align to double word boundary for the double cmpxchg
* instructions to work; see __pcpu_double_call_return_bool().
Lockless (and preemptless) fastpaths for slub Use the this_cpu_cmpxchg_double functionality to implement a lockless allocation algorithm on arches that support fast this_cpu_ops. Each of the per cpu pointers is paired with a transaction id that ensures that updates of the per cpu information can only occur in sequence on a certain cpu. A transaction id is a "long" integer that is comprised of an event number and the cpu number. The event number is incremented for every change to the per cpu state. This means that the cmpxchg instruction can verify for an update that nothing interfered and that we are updating the percpu structure for the processor where we picked up the information and that we are also currently on that processor when we update the information. This results in a significant decrease of the overhead in the fastpaths. It also makes it easy to adopt the fast path for realtime kernels since this is lockless and does not require the use of the current per cpu area over the critical section. It is only important that the per cpu area is current at the beginning of the critical section and at the end. So there is no need even to disable preemption. Test results show that the fastpath cycle count is reduced by up to ~ 40% (alloc/free test goes from ~140 cycles down to ~80). The slowpath for kfree adds a few cycles. Sadly this does nothing for the slowpath which is where the main issues with performance in slub are but the best case performance rises significantly. (For that see the more complex slub patches that require cmpxchg_double) Kmalloc: alloc/free test Before: 10000 times kmalloc(8)/kfree -> 134 cycles 10000 times kmalloc(16)/kfree -> 152 cycles 10000 times kmalloc(32)/kfree -> 144 cycles 10000 times kmalloc(64)/kfree -> 142 cycles 10000 times kmalloc(128)/kfree -> 142 cycles 10000 times kmalloc(256)/kfree -> 132 cycles 10000 times kmalloc(512)/kfree -> 132 cycles 10000 times kmalloc(1024)/kfree -> 135 cycles 10000 times kmalloc(2048)/kfree -> 135 cycles 10000 times kmalloc(4096)/kfree -> 135 cycles 10000 times kmalloc(8192)/kfree -> 144 cycles 10000 times kmalloc(16384)/kfree -> 754 cycles After: 10000 times kmalloc(8)/kfree -> 78 cycles 10000 times kmalloc(16)/kfree -> 78 cycles 10000 times kmalloc(32)/kfree -> 82 cycles 10000 times kmalloc(64)/kfree -> 88 cycles 10000 times kmalloc(128)/kfree -> 79 cycles 10000 times kmalloc(256)/kfree -> 79 cycles 10000 times kmalloc(512)/kfree -> 85 cycles 10000 times kmalloc(1024)/kfree -> 82 cycles 10000 times kmalloc(2048)/kfree -> 82 cycles 10000 times kmalloc(4096)/kfree -> 85 cycles 10000 times kmalloc(8192)/kfree -> 82 cycles 10000 times kmalloc(16384)/kfree -> 706 cycles Kmalloc: Repeatedly allocate then free test Before: 10000 times kmalloc(8) -> 211 cycles kfree -> 113 cycles 10000 times kmalloc(16) -> 174 cycles kfree -> 115 cycles 10000 times kmalloc(32) -> 235 cycles kfree -> 129 cycles 10000 times kmalloc(64) -> 222 cycles kfree -> 120 cycles 10000 times kmalloc(128) -> 343 cycles kfree -> 139 cycles 10000 times kmalloc(256) -> 827 cycles kfree -> 147 cycles 10000 times kmalloc(512) -> 1048 cycles kfree -> 272 cycles 10000 times kmalloc(1024) -> 2043 cycles kfree -> 528 cycles 10000 times kmalloc(2048) -> 4002 cycles kfree -> 571 cycles 10000 times kmalloc(4096) -> 7740 cycles kfree -> 628 cycles 10000 times kmalloc(8192) -> 8062 cycles kfree -> 850 cycles 10000 times kmalloc(16384) -> 8895 cycles kfree -> 1249 cycles After: 10000 times kmalloc(8) -> 190 cycles kfree -> 129 cycles 10000 times kmalloc(16) -> 76 cycles kfree -> 123 cycles 10000 times kmalloc(32) -> 126 cycles kfree -> 124 cycles 10000 times kmalloc(64) -> 181 cycles kfree -> 128 cycles 10000 times kmalloc(128) -> 310 cycles kfree -> 140 cycles 10000 times kmalloc(256) -> 809 cycles kfree -> 165 cycles 10000 times kmalloc(512) -> 1005 cycles kfree -> 269 cycles 10000 times kmalloc(1024) -> 1999 cycles kfree -> 527 cycles 10000 times kmalloc(2048) -> 3967 cycles kfree -> 570 cycles 10000 times kmalloc(4096) -> 7658 cycles kfree -> 637 cycles 10000 times kmalloc(8192) -> 8111 cycles kfree -> 859 cycles 10000 times kmalloc(16384) -> 8791 cycles kfree -> 1173 cycles Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-02-26 01:38:54 +08:00
*/
s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2 * sizeof(void *));
Lockless (and preemptless) fastpaths for slub Use the this_cpu_cmpxchg_double functionality to implement a lockless allocation algorithm on arches that support fast this_cpu_ops. Each of the per cpu pointers is paired with a transaction id that ensures that updates of the per cpu information can only occur in sequence on a certain cpu. A transaction id is a "long" integer that is comprised of an event number and the cpu number. The event number is incremented for every change to the per cpu state. This means that the cmpxchg instruction can verify for an update that nothing interfered and that we are updating the percpu structure for the processor where we picked up the information and that we are also currently on that processor when we update the information. This results in a significant decrease of the overhead in the fastpaths. It also makes it easy to adopt the fast path for realtime kernels since this is lockless and does not require the use of the current per cpu area over the critical section. It is only important that the per cpu area is current at the beginning of the critical section and at the end. So there is no need even to disable preemption. Test results show that the fastpath cycle count is reduced by up to ~ 40% (alloc/free test goes from ~140 cycles down to ~80). The slowpath for kfree adds a few cycles. Sadly this does nothing for the slowpath which is where the main issues with performance in slub are but the best case performance rises significantly. (For that see the more complex slub patches that require cmpxchg_double) Kmalloc: alloc/free test Before: 10000 times kmalloc(8)/kfree -> 134 cycles 10000 times kmalloc(16)/kfree -> 152 cycles 10000 times kmalloc(32)/kfree -> 144 cycles 10000 times kmalloc(64)/kfree -> 142 cycles 10000 times kmalloc(128)/kfree -> 142 cycles 10000 times kmalloc(256)/kfree -> 132 cycles 10000 times kmalloc(512)/kfree -> 132 cycles 10000 times kmalloc(1024)/kfree -> 135 cycles 10000 times kmalloc(2048)/kfree -> 135 cycles 10000 times kmalloc(4096)/kfree -> 135 cycles 10000 times kmalloc(8192)/kfree -> 144 cycles 10000 times kmalloc(16384)/kfree -> 754 cycles After: 10000 times kmalloc(8)/kfree -> 78 cycles 10000 times kmalloc(16)/kfree -> 78 cycles 10000 times kmalloc(32)/kfree -> 82 cycles 10000 times kmalloc(64)/kfree -> 88 cycles 10000 times kmalloc(128)/kfree -> 79 cycles 10000 times kmalloc(256)/kfree -> 79 cycles 10000 times kmalloc(512)/kfree -> 85 cycles 10000 times kmalloc(1024)/kfree -> 82 cycles 10000 times kmalloc(2048)/kfree -> 82 cycles 10000 times kmalloc(4096)/kfree -> 85 cycles 10000 times kmalloc(8192)/kfree -> 82 cycles 10000 times kmalloc(16384)/kfree -> 706 cycles Kmalloc: Repeatedly allocate then free test Before: 10000 times kmalloc(8) -> 211 cycles kfree -> 113 cycles 10000 times kmalloc(16) -> 174 cycles kfree -> 115 cycles 10000 times kmalloc(32) -> 235 cycles kfree -> 129 cycles 10000 times kmalloc(64) -> 222 cycles kfree -> 120 cycles 10000 times kmalloc(128) -> 343 cycles kfree -> 139 cycles 10000 times kmalloc(256) -> 827 cycles kfree -> 147 cycles 10000 times kmalloc(512) -> 1048 cycles kfree -> 272 cycles 10000 times kmalloc(1024) -> 2043 cycles kfree -> 528 cycles 10000 times kmalloc(2048) -> 4002 cycles kfree -> 571 cycles 10000 times kmalloc(4096) -> 7740 cycles kfree -> 628 cycles 10000 times kmalloc(8192) -> 8062 cycles kfree -> 850 cycles 10000 times kmalloc(16384) -> 8895 cycles kfree -> 1249 cycles After: 10000 times kmalloc(8) -> 190 cycles kfree -> 129 cycles 10000 times kmalloc(16) -> 76 cycles kfree -> 123 cycles 10000 times kmalloc(32) -> 126 cycles kfree -> 124 cycles 10000 times kmalloc(64) -> 181 cycles kfree -> 128 cycles 10000 times kmalloc(128) -> 310 cycles kfree -> 140 cycles 10000 times kmalloc(256) -> 809 cycles kfree -> 165 cycles 10000 times kmalloc(512) -> 1005 cycles kfree -> 269 cycles 10000 times kmalloc(1024) -> 1999 cycles kfree -> 527 cycles 10000 times kmalloc(2048) -> 3967 cycles kfree -> 570 cycles 10000 times kmalloc(4096) -> 7658 cycles kfree -> 637 cycles 10000 times kmalloc(8192) -> 8111 cycles kfree -> 859 cycles 10000 times kmalloc(16384) -> 8791 cycles kfree -> 1173 cycles Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-02-26 01:38:54 +08:00
if (!s->cpu_slab)
return 0;
init_kmem_cache_cpus(s);
SLUB: Place kmem_cache_cpu structures in a NUMA aware way The kmem_cache_cpu structures introduced are currently an array placed in the kmem_cache struct. Meaning the kmem_cache_cpu structures are overwhelmingly on the wrong node for systems with a higher amount of nodes. These are performance critical structures since the per node information has to be touched for every alloc and free in a slab. In order to place the kmem_cache_cpu structure optimally we put an array of pointers to kmem_cache_cpu structs in kmem_cache (similar to SLAB). However, the kmem_cache_cpu structures can now be allocated in a more intelligent way. We would like to put per cpu structures for the same cpu but different slab caches in cachelines together to save space and decrease the cache footprint. However, the slab allocators itself control only allocations per node. We set up a simple per cpu array for every processor with 100 per cpu structures which is usually enough to get them all set up right. If we run out then we fall back to kmalloc_node. This also solves the bootstrap problem since we do not have to use slab allocator functions early in boot to get memory for the small per cpu structures. Pro: - NUMA aware placement improves memory performance - All global structures in struct kmem_cache become readonly - Dense packing of per cpu structures reduces cacheline footprint in SMP and NUMA. - Potential avoidance of exclusive cacheline fetches on the free and alloc hotpath since multiple kmem_cache_cpu structures are in one cacheline. This is particularly important for the kmalloc array. Cons: - Additional reference to one read only cacheline (per cpu array of pointers to kmem_cache_cpu) in both slab_alloc() and slab_free(). [akinobu.mita@gmail.com: fix cpu hotplug offline/online path] Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: "Pekka Enberg" <penberg@cs.helsinki.fi> Cc: Akinobu Mita <akinobu.mita@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 16:26:08 +08:00
Lockless (and preemptless) fastpaths for slub Use the this_cpu_cmpxchg_double functionality to implement a lockless allocation algorithm on arches that support fast this_cpu_ops. Each of the per cpu pointers is paired with a transaction id that ensures that updates of the per cpu information can only occur in sequence on a certain cpu. A transaction id is a "long" integer that is comprised of an event number and the cpu number. The event number is incremented for every change to the per cpu state. This means that the cmpxchg instruction can verify for an update that nothing interfered and that we are updating the percpu structure for the processor where we picked up the information and that we are also currently on that processor when we update the information. This results in a significant decrease of the overhead in the fastpaths. It also makes it easy to adopt the fast path for realtime kernels since this is lockless and does not require the use of the current per cpu area over the critical section. It is only important that the per cpu area is current at the beginning of the critical section and at the end. So there is no need even to disable preemption. Test results show that the fastpath cycle count is reduced by up to ~ 40% (alloc/free test goes from ~140 cycles down to ~80). The slowpath for kfree adds a few cycles. Sadly this does nothing for the slowpath which is where the main issues with performance in slub are but the best case performance rises significantly. (For that see the more complex slub patches that require cmpxchg_double) Kmalloc: alloc/free test Before: 10000 times kmalloc(8)/kfree -> 134 cycles 10000 times kmalloc(16)/kfree -> 152 cycles 10000 times kmalloc(32)/kfree -> 144 cycles 10000 times kmalloc(64)/kfree -> 142 cycles 10000 times kmalloc(128)/kfree -> 142 cycles 10000 times kmalloc(256)/kfree -> 132 cycles 10000 times kmalloc(512)/kfree -> 132 cycles 10000 times kmalloc(1024)/kfree -> 135 cycles 10000 times kmalloc(2048)/kfree -> 135 cycles 10000 times kmalloc(4096)/kfree -> 135 cycles 10000 times kmalloc(8192)/kfree -> 144 cycles 10000 times kmalloc(16384)/kfree -> 754 cycles After: 10000 times kmalloc(8)/kfree -> 78 cycles 10000 times kmalloc(16)/kfree -> 78 cycles 10000 times kmalloc(32)/kfree -> 82 cycles 10000 times kmalloc(64)/kfree -> 88 cycles 10000 times kmalloc(128)/kfree -> 79 cycles 10000 times kmalloc(256)/kfree -> 79 cycles 10000 times kmalloc(512)/kfree -> 85 cycles 10000 times kmalloc(1024)/kfree -> 82 cycles 10000 times kmalloc(2048)/kfree -> 82 cycles 10000 times kmalloc(4096)/kfree -> 85 cycles 10000 times kmalloc(8192)/kfree -> 82 cycles 10000 times kmalloc(16384)/kfree -> 706 cycles Kmalloc: Repeatedly allocate then free test Before: 10000 times kmalloc(8) -> 211 cycles kfree -> 113 cycles 10000 times kmalloc(16) -> 174 cycles kfree -> 115 cycles 10000 times kmalloc(32) -> 235 cycles kfree -> 129 cycles 10000 times kmalloc(64) -> 222 cycles kfree -> 120 cycles 10000 times kmalloc(128) -> 343 cycles kfree -> 139 cycles 10000 times kmalloc(256) -> 827 cycles kfree -> 147 cycles 10000 times kmalloc(512) -> 1048 cycles kfree -> 272 cycles 10000 times kmalloc(1024) -> 2043 cycles kfree -> 528 cycles 10000 times kmalloc(2048) -> 4002 cycles kfree -> 571 cycles 10000 times kmalloc(4096) -> 7740 cycles kfree -> 628 cycles 10000 times kmalloc(8192) -> 8062 cycles kfree -> 850 cycles 10000 times kmalloc(16384) -> 8895 cycles kfree -> 1249 cycles After: 10000 times kmalloc(8) -> 190 cycles kfree -> 129 cycles 10000 times kmalloc(16) -> 76 cycles kfree -> 123 cycles 10000 times kmalloc(32) -> 126 cycles kfree -> 124 cycles 10000 times kmalloc(64) -> 181 cycles kfree -> 128 cycles 10000 times kmalloc(128) -> 310 cycles kfree -> 140 cycles 10000 times kmalloc(256) -> 809 cycles kfree -> 165 cycles 10000 times kmalloc(512) -> 1005 cycles kfree -> 269 cycles 10000 times kmalloc(1024) -> 1999 cycles kfree -> 527 cycles 10000 times kmalloc(2048) -> 3967 cycles kfree -> 570 cycles 10000 times kmalloc(4096) -> 7658 cycles kfree -> 637 cycles 10000 times kmalloc(8192) -> 8111 cycles kfree -> 859 cycles 10000 times kmalloc(16384) -> 8791 cycles kfree -> 1173 cycles Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-02-26 01:38:54 +08:00
return 1;
SLUB: Place kmem_cache_cpu structures in a NUMA aware way The kmem_cache_cpu structures introduced are currently an array placed in the kmem_cache struct. Meaning the kmem_cache_cpu structures are overwhelmingly on the wrong node for systems with a higher amount of nodes. These are performance critical structures since the per node information has to be touched for every alloc and free in a slab. In order to place the kmem_cache_cpu structure optimally we put an array of pointers to kmem_cache_cpu structs in kmem_cache (similar to SLAB). However, the kmem_cache_cpu structures can now be allocated in a more intelligent way. We would like to put per cpu structures for the same cpu but different slab caches in cachelines together to save space and decrease the cache footprint. However, the slab allocators itself control only allocations per node. We set up a simple per cpu array for every processor with 100 per cpu structures which is usually enough to get them all set up right. If we run out then we fall back to kmalloc_node. This also solves the bootstrap problem since we do not have to use slab allocator functions early in boot to get memory for the small per cpu structures. Pro: - NUMA aware placement improves memory performance - All global structures in struct kmem_cache become readonly - Dense packing of per cpu structures reduces cacheline footprint in SMP and NUMA. - Potential avoidance of exclusive cacheline fetches on the free and alloc hotpath since multiple kmem_cache_cpu structures are in one cacheline. This is particularly important for the kmalloc array. Cons: - Additional reference to one read only cacheline (per cpu array of pointers to kmem_cache_cpu) in both slab_alloc() and slab_free(). [akinobu.mita@gmail.com: fix cpu hotplug offline/online path] Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: "Pekka Enberg" <penberg@cs.helsinki.fi> Cc: Akinobu Mita <akinobu.mita@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 16:26:08 +08:00
}
static struct kmem_cache *kmem_cache_node;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/*
* No kmalloc_node yet so do it by hand. We know that this is the first
* slab on the node for this slabcache. There are no concurrent accesses
* possible.
*
* Note that this function only works on the kmem_cache_node
* when allocating for the kmem_cache_node. This is used for bootstrapping
SLUB: Place kmem_cache_cpu structures in a NUMA aware way The kmem_cache_cpu structures introduced are currently an array placed in the kmem_cache struct. Meaning the kmem_cache_cpu structures are overwhelmingly on the wrong node for systems with a higher amount of nodes. These are performance critical structures since the per node information has to be touched for every alloc and free in a slab. In order to place the kmem_cache_cpu structure optimally we put an array of pointers to kmem_cache_cpu structs in kmem_cache (similar to SLAB). However, the kmem_cache_cpu structures can now be allocated in a more intelligent way. We would like to put per cpu structures for the same cpu but different slab caches in cachelines together to save space and decrease the cache footprint. However, the slab allocators itself control only allocations per node. We set up a simple per cpu array for every processor with 100 per cpu structures which is usually enough to get them all set up right. If we run out then we fall back to kmalloc_node. This also solves the bootstrap problem since we do not have to use slab allocator functions early in boot to get memory for the small per cpu structures. Pro: - NUMA aware placement improves memory performance - All global structures in struct kmem_cache become readonly - Dense packing of per cpu structures reduces cacheline footprint in SMP and NUMA. - Potential avoidance of exclusive cacheline fetches on the free and alloc hotpath since multiple kmem_cache_cpu structures are in one cacheline. This is particularly important for the kmalloc array. Cons: - Additional reference to one read only cacheline (per cpu array of pointers to kmem_cache_cpu) in both slab_alloc() and slab_free(). [akinobu.mita@gmail.com: fix cpu hotplug offline/online path] Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: "Pekka Enberg" <penberg@cs.helsinki.fi> Cc: Akinobu Mita <akinobu.mita@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 16:26:08 +08:00
* memory on a fresh node that has no slab structures yet.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*/
static void early_kmem_cache_node_alloc(int node)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
{
struct page *page;
struct kmem_cache_node *n;
BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
BUG_ON(!page);
if (page_to_nid(page) != node) {
pr_err("SLUB: Unable to allocate memory from node %d\n", node);
pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
}
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
n = page->freelist;
BUG_ON(!n);
#ifdef CONFIG_SLUB_DEBUG
init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
init_tracking(kmem_cache_node, n);
#endif
kasan, mm: integrate slab init_on_alloc with HW_TAGS This change uses the previously added memory initialization feature of HW_TAGS KASAN routines for slab memory when init_on_alloc is enabled. With this change, memory initialization memset() is no longer called when both HW_TAGS KASAN and init_on_alloc are enabled. Instead, memory is initialized in KASAN runtime. The memory initialization memset() is moved into slab_post_alloc_hook() that currently directly follows the initialization loop. A new argument is added to slab_post_alloc_hook() that indicates whether to initialize the memory or not. To avoid discrepancies with which memory gets initialized that can be caused by future changes, both KASAN hook and initialization memset() are put together and a warning comment is added. Combining setting allocation tags with memory initialization improves HW_TAGS KASAN performance when init_on_alloc is enabled. Link: https://lkml.kernel.org/r/c1292aeb5d519da221ec74a0684a949b027d7720.1615296150.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Reviewed-by: Marco Elver <elver@google.com> Cc: Alexander Potapenko <glider@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Branislav Rankov <Branislav.Rankov@arm.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christoph Lameter <cl@linux.com> Cc: David Rientjes <rientjes@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Evgenii Stepanov <eugenis@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kevin Brodsky <kevin.brodsky@arm.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Collingbourne <pcc@google.com> Cc: Vincenzo Frascino <vincenzo.frascino@arm.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-04-30 14:00:06 +08:00
n = kasan_slab_alloc(kmem_cache_node, n, GFP_KERNEL, false);
page->freelist = get_freepointer(kmem_cache_node, n);
page->inuse = 1;
page->frozen = 0;
kmem_cache_node->node[node] = n;
init_kmem_cache_node(n);
inc_slabs_node(kmem_cache_node, node, page->objects);
/*
slub: do not assert not having lock in removing freed partial Vladimir reported the following issue: Commit c65c1877bd68 ("slub: use lockdep_assert_held") requires remove_partial() to be called with n->list_lock held, but free_partial() called from kmem_cache_close() on cache destruction does not follow this rule, leading to a warning: WARNING: CPU: 0 PID: 2787 at mm/slub.c:1536 __kmem_cache_shutdown+0x1b2/0x1f0() Modules linked in: CPU: 0 PID: 2787 Comm: modprobe Tainted: G W 3.14.0-rc1-mm1+ #1 Hardware name: 0000000000000600 ffff88003ae1dde8 ffffffff816d9583 0000000000000600 0000000000000000 ffff88003ae1de28 ffffffff8107c107 0000000000000000 ffff880037ab2b00 ffff88007c240d30 ffffea0001ee5280 ffffea0001ee52a0 Call Trace: __kmem_cache_shutdown+0x1b2/0x1f0 kmem_cache_destroy+0x43/0xf0 xfs_destroy_zones+0x103/0x110 [xfs] exit_xfs_fs+0x38/0x4e4 [xfs] SyS_delete_module+0x19a/0x1f0 system_call_fastpath+0x16/0x1b His solution was to add a spinlock in order to quiet lockdep. Although there would be no contention to adding the lock, that lock also requires disabling of interrupts which will have a larger impact on the system. Instead of adding a spinlock to a location where it is not needed for lockdep, make a __remove_partial() function that does not test if the list_lock is held, as no one should have it due to it being freed. Also added a __add_partial() function that does not do the lock validation either, as it is not needed for the creation of the cache. Signed-off-by: Steven Rostedt <rostedt@goodmis.org> Reported-by: Vladimir Davydov <vdavydov@parallels.com> Suggested-by: David Rientjes <rientjes@google.com> Acked-by: David Rientjes <rientjes@google.com> Acked-by: Vladimir Davydov <vdavydov@parallels.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-02-11 06:25:46 +08:00
* No locks need to be taken here as it has just been
* initialized and there is no concurrent access.
*/
slub: do not assert not having lock in removing freed partial Vladimir reported the following issue: Commit c65c1877bd68 ("slub: use lockdep_assert_held") requires remove_partial() to be called with n->list_lock held, but free_partial() called from kmem_cache_close() on cache destruction does not follow this rule, leading to a warning: WARNING: CPU: 0 PID: 2787 at mm/slub.c:1536 __kmem_cache_shutdown+0x1b2/0x1f0() Modules linked in: CPU: 0 PID: 2787 Comm: modprobe Tainted: G W 3.14.0-rc1-mm1+ #1 Hardware name: 0000000000000600 ffff88003ae1dde8 ffffffff816d9583 0000000000000600 0000000000000000 ffff88003ae1de28 ffffffff8107c107 0000000000000000 ffff880037ab2b00 ffff88007c240d30 ffffea0001ee5280 ffffea0001ee52a0 Call Trace: __kmem_cache_shutdown+0x1b2/0x1f0 kmem_cache_destroy+0x43/0xf0 xfs_destroy_zones+0x103/0x110 [xfs] exit_xfs_fs+0x38/0x4e4 [xfs] SyS_delete_module+0x19a/0x1f0 system_call_fastpath+0x16/0x1b His solution was to add a spinlock in order to quiet lockdep. Although there would be no contention to adding the lock, that lock also requires disabling of interrupts which will have a larger impact on the system. Instead of adding a spinlock to a location where it is not needed for lockdep, make a __remove_partial() function that does not test if the list_lock is held, as no one should have it due to it being freed. Also added a __add_partial() function that does not do the lock validation either, as it is not needed for the creation of the cache. Signed-off-by: Steven Rostedt <rostedt@goodmis.org> Reported-by: Vladimir Davydov <vdavydov@parallels.com> Suggested-by: David Rientjes <rientjes@google.com> Acked-by: David Rientjes <rientjes@google.com> Acked-by: Vladimir Davydov <vdavydov@parallels.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-02-11 06:25:46 +08:00
__add_partial(n, page, DEACTIVATE_TO_HEAD);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
static void free_kmem_cache_nodes(struct kmem_cache *s)
{
int node;
struct kmem_cache_node *n;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
for_each_kmem_cache_node(s, node, n) {
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
s->node[node] = NULL;
kmem_cache_free(kmem_cache_node, n);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
}
mm: slab: free kmem_cache_node after destroy sysfs file When slub_debug alloc_calls_show is enabled we will try to track location and user of slab object on each online node, kmem_cache_node structure and cpu_cache/cpu_slub shouldn't be freed till there is the last reference to sysfs file. This fixes the following panic: BUG: unable to handle kernel NULL pointer dereference at 0000000000000020 IP: list_locations+0x169/0x4e0 PGD 257304067 PUD 438456067 PMD 0 Oops: 0000 [#1] SMP CPU: 3 PID: 973074 Comm: cat ve: 0 Not tainted 3.10.0-229.7.2.ovz.9.30-00007-japdoll-dirty #2 9.30 Hardware name: DEPO Computers To Be Filled By O.E.M./H67DE3, BIOS L1.60c 07/14/2011 task: ffff88042a5dc5b0 ti: ffff88037f8d8000 task.ti: ffff88037f8d8000 RIP: list_locations+0x169/0x4e0 Call Trace: alloc_calls_show+0x1d/0x30 slab_attr_show+0x1b/0x30 sysfs_read_file+0x9a/0x1a0 vfs_read+0x9c/0x170 SyS_read+0x58/0xb0 system_call_fastpath+0x16/0x1b Code: 5e 07 12 00 b9 00 04 00 00 3d 00 04 00 00 0f 4f c1 3d 00 04 00 00 89 45 b0 0f 84 c3 00 00 00 48 63 45 b0 49 8b 9c c4 f8 00 00 00 <48> 8b 43 20 48 85 c0 74 b6 48 89 df e8 46 37 44 00 48 8b 53 10 CR2: 0000000000000020 Separated __kmem_cache_release from __kmem_cache_shutdown which now called on slab_kmem_cache_release (after the last reference to sysfs file object has dropped). Reintroduced locking in free_partial as sysfs file might access cache's partial list after shutdowning - partial revert of the commit 69cb8e6b7c29 ("slub: free slabs without holding locks"). Zap __remove_partial and use remove_partial (w/o underscores) as free_partial now takes list_lock which s partial revert for commit 1e4dd9461fab ("slub: do not assert not having lock in removing freed partial") Signed-off-by: Dmitry Safonov <dsafonov@virtuozzo.com> Suggested-by: Vladimir Davydov <vdavydov@virtuozzo.com> Acked-by: Vladimir Davydov <vdavydov@virtuozzo.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-02-18 05:11:37 +08:00
void __kmem_cache_release(struct kmem_cache *s)
{
mm: SLUB freelist randomization Implements freelist randomization for the SLUB allocator. It was previous implemented for the SLAB allocator. Both use the same configuration option (CONFIG_SLAB_FREELIST_RANDOM). The list is randomized during initialization of a new set of pages. The order on different freelist sizes is pre-computed at boot for performance. Each kmem_cache has its own randomized freelist. This security feature reduces the predictability of the kernel SLUB allocator against heap overflows rendering attacks much less stable. For example these attacks exploit the predictability of the heap: - Linux Kernel CAN SLUB overflow (https://goo.gl/oMNWkU) - Exploiting Linux Kernel Heap corruptions (http://goo.gl/EXLn95) Performance results: slab_test impact is between 3% to 4% on average for 100000 attempts without smp. It is a very focused testing, kernbench show the overall impact on the system is way lower. Before: Single thread testing ===================== 1. Kmalloc: Repeatedly allocate then free test 100000 times kmalloc(8) -> 49 cycles kfree -> 77 cycles 100000 times kmalloc(16) -> 51 cycles kfree -> 79 cycles 100000 times kmalloc(32) -> 53 cycles kfree -> 83 cycles 100000 times kmalloc(64) -> 62 cycles kfree -> 90 cycles 100000 times kmalloc(128) -> 81 cycles kfree -> 97 cycles 100000 times kmalloc(256) -> 98 cycles kfree -> 121 cycles 100000 times kmalloc(512) -> 95 cycles kfree -> 122 cycles 100000 times kmalloc(1024) -> 96 cycles kfree -> 126 cycles 100000 times kmalloc(2048) -> 115 cycles kfree -> 140 cycles 100000 times kmalloc(4096) -> 149 cycles kfree -> 171 cycles 2. Kmalloc: alloc/free test 100000 times kmalloc(8)/kfree -> 70 cycles 100000 times kmalloc(16)/kfree -> 70 cycles 100000 times kmalloc(32)/kfree -> 70 cycles 100000 times kmalloc(64)/kfree -> 70 cycles 100000 times kmalloc(128)/kfree -> 70 cycles 100000 times kmalloc(256)/kfree -> 69 cycles 100000 times kmalloc(512)/kfree -> 70 cycles 100000 times kmalloc(1024)/kfree -> 73 cycles 100000 times kmalloc(2048)/kfree -> 72 cycles 100000 times kmalloc(4096)/kfree -> 71 cycles After: Single thread testing ===================== 1. Kmalloc: Repeatedly allocate then free test 100000 times kmalloc(8) -> 57 cycles kfree -> 78 cycles 100000 times kmalloc(16) -> 61 cycles kfree -> 81 cycles 100000 times kmalloc(32) -> 76 cycles kfree -> 93 cycles 100000 times kmalloc(64) -> 83 cycles kfree -> 94 cycles 100000 times kmalloc(128) -> 106 cycles kfree -> 107 cycles 100000 times kmalloc(256) -> 118 cycles kfree -> 117 cycles 100000 times kmalloc(512) -> 114 cycles kfree -> 116 cycles 100000 times kmalloc(1024) -> 115 cycles kfree -> 118 cycles 100000 times kmalloc(2048) -> 147 cycles kfree -> 131 cycles 100000 times kmalloc(4096) -> 214 cycles kfree -> 161 cycles 2. Kmalloc: alloc/free test 100000 times kmalloc(8)/kfree -> 66 cycles 100000 times kmalloc(16)/kfree -> 66 cycles 100000 times kmalloc(32)/kfree -> 66 cycles 100000 times kmalloc(64)/kfree -> 66 cycles 100000 times kmalloc(128)/kfree -> 65 cycles 100000 times kmalloc(256)/kfree -> 67 cycles 100000 times kmalloc(512)/kfree -> 67 cycles 100000 times kmalloc(1024)/kfree -> 64 cycles 100000 times kmalloc(2048)/kfree -> 67 cycles 100000 times kmalloc(4096)/kfree -> 67 cycles Kernbench, before: Average Optimal load -j 12 Run (std deviation): Elapsed Time 101.873 (1.16069) User Time 1045.22 (1.60447) System Time 88.969 (0.559195) Percent CPU 1112.9 (13.8279) Context Switches 189140 (2282.15) Sleeps 99008.6 (768.091) After: Average Optimal load -j 12 Run (std deviation): Elapsed Time 102.47 (0.562732) User Time 1045.3 (1.34263) System Time 88.311 (0.342554) Percent CPU 1105.8 (6.49444) Context Switches 189081 (2355.78) Sleeps 99231.5 (800.358) Link: http://lkml.kernel.org/r/1464295031-26375-3-git-send-email-thgarnie@google.com Signed-off-by: Thomas Garnier <thgarnie@google.com> Reviewed-by: Kees Cook <keescook@chromium.org> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-07-27 06:21:59 +08:00
cache_random_seq_destroy(s);
mm: slab: free kmem_cache_node after destroy sysfs file When slub_debug alloc_calls_show is enabled we will try to track location and user of slab object on each online node, kmem_cache_node structure and cpu_cache/cpu_slub shouldn't be freed till there is the last reference to sysfs file. This fixes the following panic: BUG: unable to handle kernel NULL pointer dereference at 0000000000000020 IP: list_locations+0x169/0x4e0 PGD 257304067 PUD 438456067 PMD 0 Oops: 0000 [#1] SMP CPU: 3 PID: 973074 Comm: cat ve: 0 Not tainted 3.10.0-229.7.2.ovz.9.30-00007-japdoll-dirty #2 9.30 Hardware name: DEPO Computers To Be Filled By O.E.M./H67DE3, BIOS L1.60c 07/14/2011 task: ffff88042a5dc5b0 ti: ffff88037f8d8000 task.ti: ffff88037f8d8000 RIP: list_locations+0x169/0x4e0 Call Trace: alloc_calls_show+0x1d/0x30 slab_attr_show+0x1b/0x30 sysfs_read_file+0x9a/0x1a0 vfs_read+0x9c/0x170 SyS_read+0x58/0xb0 system_call_fastpath+0x16/0x1b Code: 5e 07 12 00 b9 00 04 00 00 3d 00 04 00 00 0f 4f c1 3d 00 04 00 00 89 45 b0 0f 84 c3 00 00 00 48 63 45 b0 49 8b 9c c4 f8 00 00 00 <48> 8b 43 20 48 85 c0 74 b6 48 89 df e8 46 37 44 00 48 8b 53 10 CR2: 0000000000000020 Separated __kmem_cache_release from __kmem_cache_shutdown which now called on slab_kmem_cache_release (after the last reference to sysfs file object has dropped). Reintroduced locking in free_partial as sysfs file might access cache's partial list after shutdowning - partial revert of the commit 69cb8e6b7c29 ("slub: free slabs without holding locks"). Zap __remove_partial and use remove_partial (w/o underscores) as free_partial now takes list_lock which s partial revert for commit 1e4dd9461fab ("slub: do not assert not having lock in removing freed partial") Signed-off-by: Dmitry Safonov <dsafonov@virtuozzo.com> Suggested-by: Vladimir Davydov <vdavydov@virtuozzo.com> Acked-by: Vladimir Davydov <vdavydov@virtuozzo.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-02-18 05:11:37 +08:00
free_percpu(s->cpu_slab);
free_kmem_cache_nodes(s);
}
static int init_kmem_cache_nodes(struct kmem_cache *s)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
{
int node;
mm, slab, slub: stop taking memory hotplug lock Since commit 03afc0e25f7f ("slab: get_online_mems for kmem_cache_{create,destroy,shrink}") we are taking memory hotplug lock for SLAB and SLUB when creating, destroying or shrinking a cache. It is quite a heavy lock and it's best to avoid it if possible, as we had several issues with lockdep complaining about ordering in the past, see e.g. e4f8e513c3d3 ("mm/slub: fix a deadlock in show_slab_objects()"). The problem scenario in 03afc0e25f7f (solved by the memory hotplug lock) can be summarized as follows: while there's slab_mutex synchronizing new kmem cache creation and SLUB's MEM_GOING_ONLINE callback slab_mem_going_online_callback(), we may miss creation of kmem_cache_node for the hotplugged node in the new kmem cache, because the hotplug callback doesn't yet see the new cache, and cache creation in init_kmem_cache_nodes() only inits kmem_cache_node for nodes in the N_NORMAL_MEMORY nodemask, which however may not yet include the new node, as that happens only later after the MEM_GOING_ONLINE callback. Instead of using get/put_online_mems(), the problem can be solved by SLUB maintaining its own nodemask of nodes for which it has allocated the per-node kmem_cache_node structures. This nodemask would generally mirror the N_NORMAL_MEMORY nodemask, but would be updated only in under SLUB's control in its memory hotplug callbacks under the slab_mutex. This patch adds such nodemask and its handling. Commit 03afc0e25f7f mentiones "issues like [the one above]", but there don't appear to be further issues. All the paths (shared for SLAB and SLUB) taking the memory hotplug locks are also taking the slab_mutex, except kmem_cache_shrink() where 03afc0e25f7f replaced slab_mutex with get/put_online_mems(). We however cannot simply restore slab_mutex in kmem_cache_shrink(), as SLUB can enters the function from a write to sysfs 'shrink' file, thus holding kernfs lock, and in kmem_cache_create() the kernfs lock is nested within slab_mutex. But on closer inspection we don't actually need to protect kmem_cache_shrink() from hotplug callbacks: While SLUB's __kmem_cache_shrink() does for_each_kmem_cache_node(), missing a new node added in parallel hotplug is not fatal, and parallel hotremove does not free kmem_cache_node's anymore after the previous patch, so use-after free cannot happen. The per-node shrinking itself is protected by n->list_lock. Same is true for SLAB, and SLOB is no-op. SLAB also doesn't need the memory hotplug locking, which it only gained by 03afc0e25f7f through the shared paths in slab_common.c. Its memory hotplug callbacks are also protected by slab_mutex against races with these paths. The problem of SLUB relying on N_NORMAL_MEMORY doesn't apply to SLAB, as its setup_kmem_cache_nodes relies on N_ONLINE, and the new node is already set there during the MEM_GOING_ONLINE callback, so no special care is needed for SLAB. As such, this patch removes all get/put_online_mems() usage by the slab subsystem. Link: https://lkml.kernel.org/r/20210113131634.3671-3-vbabka@suse.cz Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Cc: Christoph Lameter <cl@linux.com> Cc: David Hildenbrand <david@redhat.com> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: Qian Cai <cai@redhat.com> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-25 04:01:12 +08:00
for_each_node_mask(node, slab_nodes) {
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
struct kmem_cache_node *n;
if (slab_state == DOWN) {
early_kmem_cache_node_alloc(node);
continue;
}
n = kmem_cache_alloc_node(kmem_cache_node,
GFP_KERNEL, node);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (!n) {
free_kmem_cache_nodes(s);
return 0;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
init_kmem_cache_node(n);
s->node[node] = n;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
return 1;
}
static void set_min_partial(struct kmem_cache *s, unsigned long min)
{
if (min < MIN_PARTIAL)
min = MIN_PARTIAL;
else if (min > MAX_PARTIAL)
min = MAX_PARTIAL;
s->min_partial = min;
}
static void set_cpu_partial(struct kmem_cache *s)
{
#ifdef CONFIG_SLUB_CPU_PARTIAL
/*
* cpu_partial determined the maximum number of objects kept in the
* per cpu partial lists of a processor.
*
* Per cpu partial lists mainly contain slabs that just have one
* object freed. If they are used for allocation then they can be
* filled up again with minimal effort. The slab will never hit the
* per node partial lists and therefore no locking will be required.
*
* This setting also determines
*
* A) The number of objects from per cpu partial slabs dumped to the
* per node list when we reach the limit.
* B) The number of objects in cpu partial slabs to extract from the
* per node list when we run out of per cpu objects. We only fetch
* 50% to keep some capacity around for frees.
*/
if (!kmem_cache_has_cpu_partial(s))
slub_set_cpu_partial(s, 0);
else if (s->size >= PAGE_SIZE)
slub_set_cpu_partial(s, 2);
else if (s->size >= 1024)
slub_set_cpu_partial(s, 6);
else if (s->size >= 256)
slub_set_cpu_partial(s, 13);
else
slub_set_cpu_partial(s, 30);
#endif
}
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/*
* calculate_sizes() determines the order and the distribution of data within
* a slab object.
*/
static int calculate_sizes(struct kmem_cache *s, int forced_order)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
{
slab_flags_t flags = s->flags;
unsigned int size = s->object_size;
unsigned int freepointer_area;
unsigned int order;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/*
* Round up object size to the next word boundary. We can only
* place the free pointer at word boundaries and this determines
* the possible location of the free pointer.
*/
size = ALIGN(size, sizeof(void *));
/*
* This is the area of the object where a freepointer can be
* safely written. If redzoning adds more to the inuse size, we
* can't use that portion for writing the freepointer, so
* s->offset must be limited within this for the general case.
*/
freepointer_area = size;
#ifdef CONFIG_SLUB_DEBUG
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/*
* Determine if we can poison the object itself. If the user of
* the slab may touch the object after free or before allocation
* then we should never poison the object itself.
*/
if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
!s->ctor)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
s->flags |= __OBJECT_POISON;
else
s->flags &= ~__OBJECT_POISON;
/*
* If we are Redzoning then check if there is some space between the
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
* end of the object and the free pointer. If not then add an
* additional word to have some bytes to store Redzone information.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*/
if ((flags & SLAB_RED_ZONE) && size == s->object_size)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
size += sizeof(void *);
#endif
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/*
* With that we have determined the number of bytes in actual use
* by the object. This is the potential offset to the free pointer.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*/
s->inuse = size;
if (((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
s->ctor)) {
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/*
* Relocate free pointer after the object if it is not
* permitted to overwrite the first word of the object on
* kmem_cache_free.
*
* This is the case if we do RCU, have a constructor or
* destructor or are poisoning the objects.
mm/slub: fix incorrect interpretation of s->offset In a couple of places in the slub memory allocator, the code uses "s->offset" as a check to see if the free pointer is put right after the object. That check is no longer true with commit 3202fa62fb43 ("slub: relocate freelist pointer to middle of object"). As a result, echoing "1" into the validate sysfs file, e.g. of dentry, may cause a bunch of "Freepointer corrupt" error reports like the following to appear with the system in panic afterwards. ============================================================================= BUG dentry(666:pmcd.service) (Tainted: G B): Freepointer corrupt ----------------------------------------------------------------------------- To fix it, use the check "s->offset == s->inuse" in the new helper function freeptr_outside_object() instead. Also add another helper function get_info_end() to return the end of info block (inuse + free pointer if not overlapping with object). Fixes: 3202fa62fb43 ("slub: relocate freelist pointer to middle of object") Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Acked-by: Rafael Aquini <aquini@redhat.com> Cc: Christoph Lameter <cl@linux.com> Cc: Vitaly Nikolenko <vnik@duasynt.com> Cc: Silvio Cesare <silvio.cesare@gmail.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Markus Elfring <Markus.Elfring@web.de> Cc: Changbin Du <changbin.du@gmail.com> Link: http://lkml.kernel.org/r/20200429135328.26976-1-longman@redhat.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-05-08 09:36:06 +08:00
*
* The assumption that s->offset >= s->inuse means free
* pointer is outside of the object is used in the
* freeptr_outside_object() function. If that is no
* longer true, the function needs to be modified.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*/
s->offset = size;
size += sizeof(void *);
} else if (freepointer_area > sizeof(void *)) {
slub: relocate freelist pointer to middle of object In a recent discussion[1] with Vitaly Nikolenko and Silvio Cesare, it became clear that moving the freelist pointer away from the edge of allocations would likely improve the overall defensive posture of the inline freelist pointer. My benchmarks show no meaningful change to performance (they seem to show it being faster), so this looks like a reasonable change to make. Instead of having the freelist pointer at the very beginning of an allocation (offset 0) or at the very end of an allocation (effectively offset -sizeof(void *) from the next allocation), move it away from the edges of the allocation and into the middle. This provides some protection against small-sized neighboring overflows (or underflows), for which the freelist pointer is commonly the target. (Large or well controlled overwrites are much more likely to attack live object contents, instead of attempting freelist corruption.) The vaunted kernel build benchmark, across 5 runs. Before: Mean: 250.05 Std Dev: 1.85 and after, which appears mysteriously faster: Mean: 247.13 Std Dev: 0.76 Attempts at running "sysbench --test=memory" show the change to be well in the noise (sysbench seems to be pretty unstable here -- it's not really measuring allocation). Hackbench is more allocation-heavy, and while the std dev is above the difference, it looks like may manifest as an improvement as well: 20 runs of "hackbench -g 20 -l 1000", before: Mean: 36.322 Std Dev: 0.577 and after: Mean: 36.056 Std Dev: 0.598 [1] https://twitter.com/vnik5287/status/1235113523098685440 Signed-off-by: Kees Cook <keescook@chromium.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Acked-by: Christoph Lameter <cl@linux.com> Cc: Vitaly Nikolenko <vnik@duasynt.com> Cc: Silvio Cesare <silvio.cesare@gmail.com> Cc: Christoph Lameter <cl@linux.com>Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Link: http://lkml.kernel.org/r/202003051624.AAAC9AECC@keescook Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-02 12:04:27 +08:00
/*
* Store freelist pointer near middle of object to keep
* it away from the edges of the object to avoid small
* sized over/underflows from neighboring allocations.
*/
s->offset = ALIGN(freepointer_area / 2, sizeof(void *));
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
#ifdef CONFIG_SLUB_DEBUG
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (flags & SLAB_STORE_USER)
/*
* Need to store information about allocs and frees after
* the object.
*/
size += 2 * sizeof(struct track);
#endif
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
kasan_cache_create(s, &size, &s->flags);
#ifdef CONFIG_SLUB_DEBUG
if (flags & SLAB_RED_ZONE) {
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/*
* Add some empty padding so that we can catch
* overwrites from earlier objects rather than let
* tracking information or the free pointer be
* corrupted if a user writes before the start
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
* of the object.
*/
size += sizeof(void *);
s->red_left_pad = sizeof(void *);
s->red_left_pad = ALIGN(s->red_left_pad, s->align);
size += s->red_left_pad;
}
#endif
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/*
* SLUB stores one object immediately after another beginning from
* offset 0. In order to align the objects we have to simply size
* each object to conform to the alignment.
*/
size = ALIGN(size, s->align);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
s->size = size;
s->reciprocal_size = reciprocal_value(size);
if (forced_order >= 0)
order = forced_order;
else
order = calculate_order(size);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if ((int)order < 0)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
return 0;
s->allocflags = 0;
if (order)
s->allocflags |= __GFP_COMP;
if (s->flags & SLAB_CACHE_DMA)
s->allocflags |= GFP_DMA;
mm: add support for kmem caches in DMA32 zone Patch series "iommu/io-pgtable-arm-v7s: Use DMA32 zone for page tables", v6. This is a followup to the discussion in [1], [2]. IOMMUs using ARMv7 short-descriptor format require page tables (level 1 and 2) to be allocated within the first 4GB of RAM, even on 64-bit systems. For L1 tables that are bigger than a page, we can just use __get_free_pages with GFP_DMA32 (on arm64 systems only, arm would still use GFP_DMA). For L2 tables that only take 1KB, it would be a waste to allocate a full page, so we considered 3 approaches: 1. This series, adding support for GFP_DMA32 slab caches. 2. genalloc, which requires pre-allocating the maximum number of L2 page tables (4096, so 4MB of memory). 3. page_frag, which is not very memory-efficient as it is unable to reuse freed fragments until the whole page is freed. [3] This series is the most memory-efficient approach. stable@ note: We confirmed that this is a regression, and IOMMU errors happen on 4.19 and linux-next/master on MT8173 (elm, Acer Chromebook R13). The issue most likely starts from commit ad67f5a6545f ("arm64: replace ZONE_DMA with ZONE_DMA32"), i.e. 4.15, and presumably breaks a number of Mediatek platforms (and maybe others?). [1] https://lists.linuxfoundation.org/pipermail/iommu/2018-November/030876.html [2] https://lists.linuxfoundation.org/pipermail/iommu/2018-December/031696.html [3] https://patchwork.codeaurora.org/patch/671639/ This patch (of 3): IOMMUs using ARMv7 short-descriptor format require page tables to be allocated within the first 4GB of RAM, even on 64-bit systems. On arm64, this is done by passing GFP_DMA32 flag to memory allocation functions. For IOMMU L2 tables that only take 1KB, it would be a waste to allocate a full page using get_free_pages, so we considered 3 approaches: 1. This patch, adding support for GFP_DMA32 slab caches. 2. genalloc, which requires pre-allocating the maximum number of L2 page tables (4096, so 4MB of memory). 3. page_frag, which is not very memory-efficient as it is unable to reuse freed fragments until the whole page is freed. This change makes it possible to create a custom cache in DMA32 zone using kmem_cache_create, then allocate memory using kmem_cache_alloc. We do not create a DMA32 kmalloc cache array, as there are currently no users of kmalloc(..., GFP_DMA32). These calls will continue to trigger a warning, as we keep GFP_DMA32 in GFP_SLAB_BUG_MASK. This implies that calls to kmem_cache_*alloc on a SLAB_CACHE_DMA32 kmem_cache must _not_ use GFP_DMA32 (it is anyway redundant and unnecessary). Link: http://lkml.kernel.org/r/20181210011504.122604-2-drinkcat@chromium.org Signed-off-by: Nicolas Boichat <drinkcat@chromium.org> Acked-by: Vlastimil Babka <vbabka@suse.cz> Acked-by: Will Deacon <will.deacon@arm.com> Cc: Robin Murphy <robin.murphy@arm.com> Cc: Joerg Roedel <joro@8bytes.org> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Sasha Levin <Alexander.Levin@microsoft.com> Cc: Huaisheng Ye <yehs1@lenovo.com> Cc: Mike Rapoport <rppt@linux.vnet.ibm.com> Cc: Yong Wu <yong.wu@mediatek.com> Cc: Matthias Brugger <matthias.bgg@gmail.com> Cc: Tomasz Figa <tfiga@google.com> Cc: Yingjoe Chen <yingjoe.chen@mediatek.com> Cc: Christoph Hellwig <hch@infradead.org> Cc: Matthew Wilcox <willy@infradead.org> Cc: Hsin-Yi Wang <hsinyi@chromium.org> Cc: <stable@vger.kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-29 11:43:42 +08:00
if (s->flags & SLAB_CACHE_DMA32)
s->allocflags |= GFP_DMA32;
if (s->flags & SLAB_RECLAIM_ACCOUNT)
s->allocflags |= __GFP_RECLAIMABLE;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/*
* Determine the number of objects per slab
*/
s->oo = oo_make(order, size);
s->min = oo_make(get_order(size), size);
if (oo_objects(s->oo) > oo_objects(s->max))
s->max = s->oo;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
return !!oo_objects(s->oo);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
{
s->flags = kmem_cache_flags(s->size, flags, s->name);
mm: add SLUB free list pointer obfuscation This SLUB free list pointer obfuscation code is modified from Brad Spengler/PaX Team's code in the last public patch of grsecurity/PaX based on my understanding of the code. Changes or omissions from the original code are mine and don't reflect the original grsecurity/PaX code. This adds a per-cache random value to SLUB caches that is XORed with their freelist pointer address and value. This adds nearly zero overhead and frustrates the very common heap overflow exploitation method of overwriting freelist pointers. A recent example of the attack is written up here: http://cyseclabs.com/blog/cve-2016-6187-heap-off-by-one-exploit and there is a section dedicated to the technique the book "A Guide to Kernel Exploitation: Attacking the Core". This is based on patches by Daniel Micay, and refactored to minimize the use of #ifdef. With 200-count cycles of "hackbench -g 20 -l 1000" I saw the following run times: before: mean 10.11882499999999999995 variance .03320378329145728642 stdev .18221905304181911048 after: mean 10.12654000000000000014 variance .04700556623115577889 stdev .21680767106160192064 The difference gets lost in the noise, but if the above is to be taken literally, using CONFIG_FREELIST_HARDENED is 0.07% slower. Link: http://lkml.kernel.org/r/20170802180609.GA66807@beast Signed-off-by: Kees Cook <keescook@chromium.org> Suggested-by: Daniel Micay <danielmicay@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Tycho Andersen <tycho@docker.com> Cc: Alexander Popov <alex.popov@linux.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:19:18 +08:00
#ifdef CONFIG_SLAB_FREELIST_HARDENED
s->random = get_random_long();
#endif
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (!calculate_sizes(s, -1))
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
goto error;
if (disable_higher_order_debug) {
/*
* Disable debugging flags that store metadata if the min slab
* order increased.
*/
if (get_order(s->size) > get_order(s->object_size)) {
s->flags &= ~DEBUG_METADATA_FLAGS;
s->offset = 0;
if (!calculate_sizes(s, -1))
goto error;
}
}
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
/* Enable fast mode */
s->flags |= __CMPXCHG_DOUBLE;
#endif
/*
* The larger the object size is, the more pages we want on the partial
* list to avoid pounding the page allocator excessively.
*/
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
set_min_partial(s, ilog2(s->size) / 2);
set_cpu_partial(s);
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
#ifdef CONFIG_NUMA
s->remote_node_defrag_ratio = 1000;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
#endif
mm: SLUB freelist randomization Implements freelist randomization for the SLUB allocator. It was previous implemented for the SLAB allocator. Both use the same configuration option (CONFIG_SLAB_FREELIST_RANDOM). The list is randomized during initialization of a new set of pages. The order on different freelist sizes is pre-computed at boot for performance. Each kmem_cache has its own randomized freelist. This security feature reduces the predictability of the kernel SLUB allocator against heap overflows rendering attacks much less stable. For example these attacks exploit the predictability of the heap: - Linux Kernel CAN SLUB overflow (https://goo.gl/oMNWkU) - Exploiting Linux Kernel Heap corruptions (http://goo.gl/EXLn95) Performance results: slab_test impact is between 3% to 4% on average for 100000 attempts without smp. It is a very focused testing, kernbench show the overall impact on the system is way lower. Before: Single thread testing ===================== 1. Kmalloc: Repeatedly allocate then free test 100000 times kmalloc(8) -> 49 cycles kfree -> 77 cycles 100000 times kmalloc(16) -> 51 cycles kfree -> 79 cycles 100000 times kmalloc(32) -> 53 cycles kfree -> 83 cycles 100000 times kmalloc(64) -> 62 cycles kfree -> 90 cycles 100000 times kmalloc(128) -> 81 cycles kfree -> 97 cycles 100000 times kmalloc(256) -> 98 cycles kfree -> 121 cycles 100000 times kmalloc(512) -> 95 cycles kfree -> 122 cycles 100000 times kmalloc(1024) -> 96 cycles kfree -> 126 cycles 100000 times kmalloc(2048) -> 115 cycles kfree -> 140 cycles 100000 times kmalloc(4096) -> 149 cycles kfree -> 171 cycles 2. Kmalloc: alloc/free test 100000 times kmalloc(8)/kfree -> 70 cycles 100000 times kmalloc(16)/kfree -> 70 cycles 100000 times kmalloc(32)/kfree -> 70 cycles 100000 times kmalloc(64)/kfree -> 70 cycles 100000 times kmalloc(128)/kfree -> 70 cycles 100000 times kmalloc(256)/kfree -> 69 cycles 100000 times kmalloc(512)/kfree -> 70 cycles 100000 times kmalloc(1024)/kfree -> 73 cycles 100000 times kmalloc(2048)/kfree -> 72 cycles 100000 times kmalloc(4096)/kfree -> 71 cycles After: Single thread testing ===================== 1. Kmalloc: Repeatedly allocate then free test 100000 times kmalloc(8) -> 57 cycles kfree -> 78 cycles 100000 times kmalloc(16) -> 61 cycles kfree -> 81 cycles 100000 times kmalloc(32) -> 76 cycles kfree -> 93 cycles 100000 times kmalloc(64) -> 83 cycles kfree -> 94 cycles 100000 times kmalloc(128) -> 106 cycles kfree -> 107 cycles 100000 times kmalloc(256) -> 118 cycles kfree -> 117 cycles 100000 times kmalloc(512) -> 114 cycles kfree -> 116 cycles 100000 times kmalloc(1024) -> 115 cycles kfree -> 118 cycles 100000 times kmalloc(2048) -> 147 cycles kfree -> 131 cycles 100000 times kmalloc(4096) -> 214 cycles kfree -> 161 cycles 2. Kmalloc: alloc/free test 100000 times kmalloc(8)/kfree -> 66 cycles 100000 times kmalloc(16)/kfree -> 66 cycles 100000 times kmalloc(32)/kfree -> 66 cycles 100000 times kmalloc(64)/kfree -> 66 cycles 100000 times kmalloc(128)/kfree -> 65 cycles 100000 times kmalloc(256)/kfree -> 67 cycles 100000 times kmalloc(512)/kfree -> 67 cycles 100000 times kmalloc(1024)/kfree -> 64 cycles 100000 times kmalloc(2048)/kfree -> 67 cycles 100000 times kmalloc(4096)/kfree -> 67 cycles Kernbench, before: Average Optimal load -j 12 Run (std deviation): Elapsed Time 101.873 (1.16069) User Time 1045.22 (1.60447) System Time 88.969 (0.559195) Percent CPU 1112.9 (13.8279) Context Switches 189140 (2282.15) Sleeps 99008.6 (768.091) After: Average Optimal load -j 12 Run (std deviation): Elapsed Time 102.47 (0.562732) User Time 1045.3 (1.34263) System Time 88.311 (0.342554) Percent CPU 1105.8 (6.49444) Context Switches 189081 (2355.78) Sleeps 99231.5 (800.358) Link: http://lkml.kernel.org/r/1464295031-26375-3-git-send-email-thgarnie@google.com Signed-off-by: Thomas Garnier <thgarnie@google.com> Reviewed-by: Kees Cook <keescook@chromium.org> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-07-27 06:21:59 +08:00
/* Initialize the pre-computed randomized freelist if slab is up */
if (slab_state >= UP) {
if (init_cache_random_seq(s))
goto error;
}
if (!init_kmem_cache_nodes(s))
SLUB: Avoid page struct cacheline bouncing due to remote frees to cpu slab A remote free may access the same page struct that also contains the lockless freelist for the cpu slab. If objects have a short lifetime and are freed by a different processor then remote frees back to the slab from which we are currently allocating are frequent. The cacheline with the page struct needs to be repeately acquired in exclusive mode by both the allocating thread and the freeing thread. If this is frequent enough then performance will suffer because of cacheline bouncing. This patchset puts the lockless_freelist pointer in its own cacheline. In order to make that happen we introduce a per cpu structure called kmem_cache_cpu. Instead of keeping an array of pointers to page structs we now keep an array to a per cpu structure that--among other things--contains the pointer to the lockless freelist. The freeing thread can then keep possession of exclusive access to the page struct cacheline while the allocating thread keeps its exclusive access to the cacheline containing the per cpu structure. This works as long as the allocating cpu is able to service its request from the lockless freelist. If the lockless freelist runs empty then the allocating thread needs to acquire exclusive access to the cacheline with the page struct lock the slab. The allocating thread will then check if new objects were freed to the per cpu slab. If so it will keep the slab as the cpu slab and continue with the recently remote freed objects. So the allocating thread can take a series of just freed remote pages and dish them out again. Ideally allocations could be just recycling objects in the same slab this way which will lead to an ideal allocation / remote free pattern. The number of objects that can be handled in this way is limited by the capacity of one slab. Increasing slab size via slub_min_objects/ slub_max_order may increase the number of objects and therefore performance. If the allocating thread runs out of objects and finds that no objects were put back by the remote processor then it will retrieve a new slab (from the partial lists or from the page allocator) and start with a whole new set of objects while the remote thread may still be freeing objects to the old cpu slab. This may then repeat until the new slab is also exhausted. If remote freeing has freed objects in the earlier slab then that earlier slab will now be on the partial freelist and the allocating thread will pick that slab next for allocation. So the loop is extended. However, both threads need to take the list_lock to make the swizzling via the partial list happen. It is likely that this kind of scheme will keep the objects being passed around to a small set that can be kept in the cpu caches leading to increased performance. More code cleanups become possible: - Instead of passing a cpu we can now pass a kmem_cache_cpu structure around. Allows reducing the number of parameters to various functions. - Can define a new node_match() function for NUMA to encapsulate locality checks. Effect on allocations: Cachelines touched before this patch: Write: page cache struct and first cacheline of object Cachelines touched after this patch: Write: kmem_cache_cpu cacheline and first cacheline of object Read: page cache struct (but see later patch that avoids touching that cacheline) The handling when the lockless alloc list runs empty gets to be a bit more complicated since another cacheline has now to be written to. But that is halfway out of the hot path. Effect on freeing: Cachelines touched before this patch: Write: page_struct and first cacheline of object Cachelines touched after this patch depending on how we free: Write(to cpu_slab): kmem_cache_cpu struct and first cacheline of object Write(to other): page struct and first cacheline of object Read(to cpu_slab): page struct to id slab etc. (but see later patch that avoids touching the page struct on free) Read(to other): cpu local kmem_cache_cpu struct to verify its not the cpu slab. Summary: Pro: - Distinct cachelines so that concurrent remote frees and local allocs on a cpuslab can occur without cacheline bouncing. - Avoids potential bouncing cachelines because of neighboring per cpu pointer updates in kmem_cache's cpu_slab structure since it now grows to a cacheline (Therefore remove the comment that talks about that concern). Cons: - Freeing objects now requires the reading of one additional cacheline. That can be mitigated for some cases by the following patches but its not possible to completely eliminate these references. - Memory usage grows slightly. The size of each per cpu object is blown up from one word (pointing to the page_struct) to one cacheline with various data. So this is NR_CPUS*NR_SLABS*L1_BYTES more memory use. Lets say NR_SLABS is 100 and a cache line size of 128 then we have just increased SLAB metadata requirements by 12.8k per cpu. (Another later patch reduces these requirements) Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 16:26:05 +08:00
goto error;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (alloc_kmem_cache_cpus(s))
return 0;
SLUB: Place kmem_cache_cpu structures in a NUMA aware way The kmem_cache_cpu structures introduced are currently an array placed in the kmem_cache struct. Meaning the kmem_cache_cpu structures are overwhelmingly on the wrong node for systems with a higher amount of nodes. These are performance critical structures since the per node information has to be touched for every alloc and free in a slab. In order to place the kmem_cache_cpu structure optimally we put an array of pointers to kmem_cache_cpu structs in kmem_cache (similar to SLAB). However, the kmem_cache_cpu structures can now be allocated in a more intelligent way. We would like to put per cpu structures for the same cpu but different slab caches in cachelines together to save space and decrease the cache footprint. However, the slab allocators itself control only allocations per node. We set up a simple per cpu array for every processor with 100 per cpu structures which is usually enough to get them all set up right. If we run out then we fall back to kmalloc_node. This also solves the bootstrap problem since we do not have to use slab allocator functions early in boot to get memory for the small per cpu structures. Pro: - NUMA aware placement improves memory performance - All global structures in struct kmem_cache become readonly - Dense packing of per cpu structures reduces cacheline footprint in SMP and NUMA. - Potential avoidance of exclusive cacheline fetches on the free and alloc hotpath since multiple kmem_cache_cpu structures are in one cacheline. This is particularly important for the kmalloc array. Cons: - Additional reference to one read only cacheline (per cpu array of pointers to kmem_cache_cpu) in both slab_alloc() and slab_free(). [akinobu.mita@gmail.com: fix cpu hotplug offline/online path] Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: "Pekka Enberg" <penberg@cs.helsinki.fi> Cc: Akinobu Mita <akinobu.mita@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 16:26:08 +08:00
free_kmem_cache_nodes(s);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
error:
return -EINVAL;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
static void list_slab_objects(struct kmem_cache *s, struct page *page,
const char *text)
{
#ifdef CONFIG_SLUB_DEBUG
void *addr = page_address(page);
unsigned long *map;
void *p;
slab_err(s, page, text, s->name);
slab_lock(page);
map = get_map(s, page);
for_each_object(p, s, addr, page->objects) {
if (!test_bit(__obj_to_index(s, addr, p), map)) {
pr_err("Object 0x%p @offset=%tu\n", p, p - addr);
print_tracking(s, p);
}
}
put_map(map);
slab_unlock(page);
#endif
}
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/*
* Attempt to free all partial slabs on a node.
mm: slab: free kmem_cache_node after destroy sysfs file When slub_debug alloc_calls_show is enabled we will try to track location and user of slab object on each online node, kmem_cache_node structure and cpu_cache/cpu_slub shouldn't be freed till there is the last reference to sysfs file. This fixes the following panic: BUG: unable to handle kernel NULL pointer dereference at 0000000000000020 IP: list_locations+0x169/0x4e0 PGD 257304067 PUD 438456067 PMD 0 Oops: 0000 [#1] SMP CPU: 3 PID: 973074 Comm: cat ve: 0 Not tainted 3.10.0-229.7.2.ovz.9.30-00007-japdoll-dirty #2 9.30 Hardware name: DEPO Computers To Be Filled By O.E.M./H67DE3, BIOS L1.60c 07/14/2011 task: ffff88042a5dc5b0 ti: ffff88037f8d8000 task.ti: ffff88037f8d8000 RIP: list_locations+0x169/0x4e0 Call Trace: alloc_calls_show+0x1d/0x30 slab_attr_show+0x1b/0x30 sysfs_read_file+0x9a/0x1a0 vfs_read+0x9c/0x170 SyS_read+0x58/0xb0 system_call_fastpath+0x16/0x1b Code: 5e 07 12 00 b9 00 04 00 00 3d 00 04 00 00 0f 4f c1 3d 00 04 00 00 89 45 b0 0f 84 c3 00 00 00 48 63 45 b0 49 8b 9c c4 f8 00 00 00 <48> 8b 43 20 48 85 c0 74 b6 48 89 df e8 46 37 44 00 48 8b 53 10 CR2: 0000000000000020 Separated __kmem_cache_release from __kmem_cache_shutdown which now called on slab_kmem_cache_release (after the last reference to sysfs file object has dropped). Reintroduced locking in free_partial as sysfs file might access cache's partial list after shutdowning - partial revert of the commit 69cb8e6b7c29 ("slub: free slabs without holding locks"). Zap __remove_partial and use remove_partial (w/o underscores) as free_partial now takes list_lock which s partial revert for commit 1e4dd9461fab ("slub: do not assert not having lock in removing freed partial") Signed-off-by: Dmitry Safonov <dsafonov@virtuozzo.com> Suggested-by: Vladimir Davydov <vdavydov@virtuozzo.com> Acked-by: Vladimir Davydov <vdavydov@virtuozzo.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-02-18 05:11:37 +08:00
* This is called from __kmem_cache_shutdown(). We must take list_lock
* because sysfs file might still access partial list after the shutdowning.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*/
static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
{
mm/slub.c: run free_partial() outside of the kmem_cache_node->list_lock With debugobjects enabled and using SLAB_DESTROY_BY_RCU, when a kmem_cache_node is destroyed the call_rcu() may trigger a slab allocation to fill the debug object pool (__debug_object_init:fill_pool). Everywhere but during kmem_cache_destroy(), discard_slab() is performed outside of the kmem_cache_node->list_lock and avoids a lockdep warning about potential recursion: ============================================= [ INFO: possible recursive locking detected ] 4.8.0-rc1-gfxbench+ #1 Tainted: G U --------------------------------------------- rmmod/8895 is trying to acquire lock: (&(&n->list_lock)->rlock){-.-...}, at: [<ffffffff811c80d7>] get_partial_node.isra.63+0x47/0x430 but task is already holding lock: (&(&n->list_lock)->rlock){-.-...}, at: [<ffffffff811cbda4>] __kmem_cache_shutdown+0x54/0x320 other info that might help us debug this: Possible unsafe locking scenario: CPU0 ---- lock(&(&n->list_lock)->rlock); lock(&(&n->list_lock)->rlock); *** DEADLOCK *** May be due to missing lock nesting notation 5 locks held by rmmod/8895: #0: (&dev->mutex){......}, at: driver_detach+0x42/0xc0 #1: (&dev->mutex){......}, at: driver_detach+0x50/0xc0 #2: (cpu_hotplug.dep_map){++++++}, at: get_online_cpus+0x2d/0x80 #3: (slab_mutex){+.+.+.}, at: kmem_cache_destroy+0x3c/0x220 #4: (&(&n->list_lock)->rlock){-.-...}, at: __kmem_cache_shutdown+0x54/0x320 stack backtrace: CPU: 6 PID: 8895 Comm: rmmod Tainted: G U 4.8.0-rc1-gfxbench+ #1 Hardware name: Gigabyte Technology Co., Ltd. H87M-D3H/H87M-D3H, BIOS F11 08/18/2015 Call Trace: __lock_acquire+0x1646/0x1ad0 lock_acquire+0xb2/0x200 _raw_spin_lock+0x36/0x50 get_partial_node.isra.63+0x47/0x430 ___slab_alloc.constprop.67+0x1a7/0x3b0 __slab_alloc.isra.64.constprop.66+0x43/0x80 kmem_cache_alloc+0x236/0x2d0 __debug_object_init+0x2de/0x400 debug_object_activate+0x109/0x1e0 __call_rcu.constprop.63+0x32/0x2f0 call_rcu+0x12/0x20 discard_slab+0x3d/0x40 __kmem_cache_shutdown+0xdb/0x320 shutdown_cache+0x19/0x60 kmem_cache_destroy+0x1ae/0x220 i915_gem_load_cleanup+0x14/0x40 [i915] i915_driver_unload+0x151/0x180 [i915] i915_pci_remove+0x14/0x20 [i915] pci_device_remove+0x34/0xb0 __device_release_driver+0x95/0x140 driver_detach+0xb6/0xc0 bus_remove_driver+0x53/0xd0 driver_unregister+0x27/0x50 pci_unregister_driver+0x25/0x70 i915_exit+0x1a/0x1e2 [i915] SyS_delete_module+0x193/0x1f0 entry_SYSCALL_64_fastpath+0x1c/0xac Fixes: 52b4b950b507 ("mm: slab: free kmem_cache_node after destroy sysfs file") Link: http://lkml.kernel.org/r/1470759070-18743-1-git-send-email-chris@chris-wilson.co.uk Reported-by: Dave Gordon <david.s.gordon@intel.com> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Reviewed-by: Vladimir Davydov <vdavydov@virtuozzo.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Dmitry Safonov <dsafonov@virtuozzo.com> Cc: Daniel Vetter <daniel.vetter@ffwll.ch> Cc: Dave Gordon <david.s.gordon@intel.com> Cc: <stable@vger.kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-08-11 07:27:58 +08:00
LIST_HEAD(discard);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
struct page *page, *h;
mm: slab: free kmem_cache_node after destroy sysfs file When slub_debug alloc_calls_show is enabled we will try to track location and user of slab object on each online node, kmem_cache_node structure and cpu_cache/cpu_slub shouldn't be freed till there is the last reference to sysfs file. This fixes the following panic: BUG: unable to handle kernel NULL pointer dereference at 0000000000000020 IP: list_locations+0x169/0x4e0 PGD 257304067 PUD 438456067 PMD 0 Oops: 0000 [#1] SMP CPU: 3 PID: 973074 Comm: cat ve: 0 Not tainted 3.10.0-229.7.2.ovz.9.30-00007-japdoll-dirty #2 9.30 Hardware name: DEPO Computers To Be Filled By O.E.M./H67DE3, BIOS L1.60c 07/14/2011 task: ffff88042a5dc5b0 ti: ffff88037f8d8000 task.ti: ffff88037f8d8000 RIP: list_locations+0x169/0x4e0 Call Trace: alloc_calls_show+0x1d/0x30 slab_attr_show+0x1b/0x30 sysfs_read_file+0x9a/0x1a0 vfs_read+0x9c/0x170 SyS_read+0x58/0xb0 system_call_fastpath+0x16/0x1b Code: 5e 07 12 00 b9 00 04 00 00 3d 00 04 00 00 0f 4f c1 3d 00 04 00 00 89 45 b0 0f 84 c3 00 00 00 48 63 45 b0 49 8b 9c c4 f8 00 00 00 <48> 8b 43 20 48 85 c0 74 b6 48 89 df e8 46 37 44 00 48 8b 53 10 CR2: 0000000000000020 Separated __kmem_cache_release from __kmem_cache_shutdown which now called on slab_kmem_cache_release (after the last reference to sysfs file object has dropped). Reintroduced locking in free_partial as sysfs file might access cache's partial list after shutdowning - partial revert of the commit 69cb8e6b7c29 ("slub: free slabs without holding locks"). Zap __remove_partial and use remove_partial (w/o underscores) as free_partial now takes list_lock which s partial revert for commit 1e4dd9461fab ("slub: do not assert not having lock in removing freed partial") Signed-off-by: Dmitry Safonov <dsafonov@virtuozzo.com> Suggested-by: Vladimir Davydov <vdavydov@virtuozzo.com> Acked-by: Vladimir Davydov <vdavydov@virtuozzo.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-02-18 05:11:37 +08:00
BUG_ON(irqs_disabled());
spin_lock_irq(&n->list_lock);
list_for_each_entry_safe(page, h, &n->partial, slab_list) {
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (!page->inuse) {
mm: slab: free kmem_cache_node after destroy sysfs file When slub_debug alloc_calls_show is enabled we will try to track location and user of slab object on each online node, kmem_cache_node structure and cpu_cache/cpu_slub shouldn't be freed till there is the last reference to sysfs file. This fixes the following panic: BUG: unable to handle kernel NULL pointer dereference at 0000000000000020 IP: list_locations+0x169/0x4e0 PGD 257304067 PUD 438456067 PMD 0 Oops: 0000 [#1] SMP CPU: 3 PID: 973074 Comm: cat ve: 0 Not tainted 3.10.0-229.7.2.ovz.9.30-00007-japdoll-dirty #2 9.30 Hardware name: DEPO Computers To Be Filled By O.E.M./H67DE3, BIOS L1.60c 07/14/2011 task: ffff88042a5dc5b0 ti: ffff88037f8d8000 task.ti: ffff88037f8d8000 RIP: list_locations+0x169/0x4e0 Call Trace: alloc_calls_show+0x1d/0x30 slab_attr_show+0x1b/0x30 sysfs_read_file+0x9a/0x1a0 vfs_read+0x9c/0x170 SyS_read+0x58/0xb0 system_call_fastpath+0x16/0x1b Code: 5e 07 12 00 b9 00 04 00 00 3d 00 04 00 00 0f 4f c1 3d 00 04 00 00 89 45 b0 0f 84 c3 00 00 00 48 63 45 b0 49 8b 9c c4 f8 00 00 00 <48> 8b 43 20 48 85 c0 74 b6 48 89 df e8 46 37 44 00 48 8b 53 10 CR2: 0000000000000020 Separated __kmem_cache_release from __kmem_cache_shutdown which now called on slab_kmem_cache_release (after the last reference to sysfs file object has dropped). Reintroduced locking in free_partial as sysfs file might access cache's partial list after shutdowning - partial revert of the commit 69cb8e6b7c29 ("slub: free slabs without holding locks"). Zap __remove_partial and use remove_partial (w/o underscores) as free_partial now takes list_lock which s partial revert for commit 1e4dd9461fab ("slub: do not assert not having lock in removing freed partial") Signed-off-by: Dmitry Safonov <dsafonov@virtuozzo.com> Suggested-by: Vladimir Davydov <vdavydov@virtuozzo.com> Acked-by: Vladimir Davydov <vdavydov@virtuozzo.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-02-18 05:11:37 +08:00
remove_partial(n, page);
list_add(&page->slab_list, &discard);
} else {
list_slab_objects(s, page,
"Objects remaining in %s on __kmem_cache_shutdown()");
}
}
mm: slab: free kmem_cache_node after destroy sysfs file When slub_debug alloc_calls_show is enabled we will try to track location and user of slab object on each online node, kmem_cache_node structure and cpu_cache/cpu_slub shouldn't be freed till there is the last reference to sysfs file. This fixes the following panic: BUG: unable to handle kernel NULL pointer dereference at 0000000000000020 IP: list_locations+0x169/0x4e0 PGD 257304067 PUD 438456067 PMD 0 Oops: 0000 [#1] SMP CPU: 3 PID: 973074 Comm: cat ve: 0 Not tainted 3.10.0-229.7.2.ovz.9.30-00007-japdoll-dirty #2 9.30 Hardware name: DEPO Computers To Be Filled By O.E.M./H67DE3, BIOS L1.60c 07/14/2011 task: ffff88042a5dc5b0 ti: ffff88037f8d8000 task.ti: ffff88037f8d8000 RIP: list_locations+0x169/0x4e0 Call Trace: alloc_calls_show+0x1d/0x30 slab_attr_show+0x1b/0x30 sysfs_read_file+0x9a/0x1a0 vfs_read+0x9c/0x170 SyS_read+0x58/0xb0 system_call_fastpath+0x16/0x1b Code: 5e 07 12 00 b9 00 04 00 00 3d 00 04 00 00 0f 4f c1 3d 00 04 00 00 89 45 b0 0f 84 c3 00 00 00 48 63 45 b0 49 8b 9c c4 f8 00 00 00 <48> 8b 43 20 48 85 c0 74 b6 48 89 df e8 46 37 44 00 48 8b 53 10 CR2: 0000000000000020 Separated __kmem_cache_release from __kmem_cache_shutdown which now called on slab_kmem_cache_release (after the last reference to sysfs file object has dropped). Reintroduced locking in free_partial as sysfs file might access cache's partial list after shutdowning - partial revert of the commit 69cb8e6b7c29 ("slub: free slabs without holding locks"). Zap __remove_partial and use remove_partial (w/o underscores) as free_partial now takes list_lock which s partial revert for commit 1e4dd9461fab ("slub: do not assert not having lock in removing freed partial") Signed-off-by: Dmitry Safonov <dsafonov@virtuozzo.com> Suggested-by: Vladimir Davydov <vdavydov@virtuozzo.com> Acked-by: Vladimir Davydov <vdavydov@virtuozzo.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-02-18 05:11:37 +08:00
spin_unlock_irq(&n->list_lock);
mm/slub.c: run free_partial() outside of the kmem_cache_node->list_lock With debugobjects enabled and using SLAB_DESTROY_BY_RCU, when a kmem_cache_node is destroyed the call_rcu() may trigger a slab allocation to fill the debug object pool (__debug_object_init:fill_pool). Everywhere but during kmem_cache_destroy(), discard_slab() is performed outside of the kmem_cache_node->list_lock and avoids a lockdep warning about potential recursion: ============================================= [ INFO: possible recursive locking detected ] 4.8.0-rc1-gfxbench+ #1 Tainted: G U --------------------------------------------- rmmod/8895 is trying to acquire lock: (&(&n->list_lock)->rlock){-.-...}, at: [<ffffffff811c80d7>] get_partial_node.isra.63+0x47/0x430 but task is already holding lock: (&(&n->list_lock)->rlock){-.-...}, at: [<ffffffff811cbda4>] __kmem_cache_shutdown+0x54/0x320 other info that might help us debug this: Possible unsafe locking scenario: CPU0 ---- lock(&(&n->list_lock)->rlock); lock(&(&n->list_lock)->rlock); *** DEADLOCK *** May be due to missing lock nesting notation 5 locks held by rmmod/8895: #0: (&dev->mutex){......}, at: driver_detach+0x42/0xc0 #1: (&dev->mutex){......}, at: driver_detach+0x50/0xc0 #2: (cpu_hotplug.dep_map){++++++}, at: get_online_cpus+0x2d/0x80 #3: (slab_mutex){+.+.+.}, at: kmem_cache_destroy+0x3c/0x220 #4: (&(&n->list_lock)->rlock){-.-...}, at: __kmem_cache_shutdown+0x54/0x320 stack backtrace: CPU: 6 PID: 8895 Comm: rmmod Tainted: G U 4.8.0-rc1-gfxbench+ #1 Hardware name: Gigabyte Technology Co., Ltd. H87M-D3H/H87M-D3H, BIOS F11 08/18/2015 Call Trace: __lock_acquire+0x1646/0x1ad0 lock_acquire+0xb2/0x200 _raw_spin_lock+0x36/0x50 get_partial_node.isra.63+0x47/0x430 ___slab_alloc.constprop.67+0x1a7/0x3b0 __slab_alloc.isra.64.constprop.66+0x43/0x80 kmem_cache_alloc+0x236/0x2d0 __debug_object_init+0x2de/0x400 debug_object_activate+0x109/0x1e0 __call_rcu.constprop.63+0x32/0x2f0 call_rcu+0x12/0x20 discard_slab+0x3d/0x40 __kmem_cache_shutdown+0xdb/0x320 shutdown_cache+0x19/0x60 kmem_cache_destroy+0x1ae/0x220 i915_gem_load_cleanup+0x14/0x40 [i915] i915_driver_unload+0x151/0x180 [i915] i915_pci_remove+0x14/0x20 [i915] pci_device_remove+0x34/0xb0 __device_release_driver+0x95/0x140 driver_detach+0xb6/0xc0 bus_remove_driver+0x53/0xd0 driver_unregister+0x27/0x50 pci_unregister_driver+0x25/0x70 i915_exit+0x1a/0x1e2 [i915] SyS_delete_module+0x193/0x1f0 entry_SYSCALL_64_fastpath+0x1c/0xac Fixes: 52b4b950b507 ("mm: slab: free kmem_cache_node after destroy sysfs file") Link: http://lkml.kernel.org/r/1470759070-18743-1-git-send-email-chris@chris-wilson.co.uk Reported-by: Dave Gordon <david.s.gordon@intel.com> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Reviewed-by: Vladimir Davydov <vdavydov@virtuozzo.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Dmitry Safonov <dsafonov@virtuozzo.com> Cc: Daniel Vetter <daniel.vetter@ffwll.ch> Cc: Dave Gordon <david.s.gordon@intel.com> Cc: <stable@vger.kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-08-11 07:27:58 +08:00
list_for_each_entry_safe(page, h, &discard, slab_list)
mm/slub.c: run free_partial() outside of the kmem_cache_node->list_lock With debugobjects enabled and using SLAB_DESTROY_BY_RCU, when a kmem_cache_node is destroyed the call_rcu() may trigger a slab allocation to fill the debug object pool (__debug_object_init:fill_pool). Everywhere but during kmem_cache_destroy(), discard_slab() is performed outside of the kmem_cache_node->list_lock and avoids a lockdep warning about potential recursion: ============================================= [ INFO: possible recursive locking detected ] 4.8.0-rc1-gfxbench+ #1 Tainted: G U --------------------------------------------- rmmod/8895 is trying to acquire lock: (&(&n->list_lock)->rlock){-.-...}, at: [<ffffffff811c80d7>] get_partial_node.isra.63+0x47/0x430 but task is already holding lock: (&(&n->list_lock)->rlock){-.-...}, at: [<ffffffff811cbda4>] __kmem_cache_shutdown+0x54/0x320 other info that might help us debug this: Possible unsafe locking scenario: CPU0 ---- lock(&(&n->list_lock)->rlock); lock(&(&n->list_lock)->rlock); *** DEADLOCK *** May be due to missing lock nesting notation 5 locks held by rmmod/8895: #0: (&dev->mutex){......}, at: driver_detach+0x42/0xc0 #1: (&dev->mutex){......}, at: driver_detach+0x50/0xc0 #2: (cpu_hotplug.dep_map){++++++}, at: get_online_cpus+0x2d/0x80 #3: (slab_mutex){+.+.+.}, at: kmem_cache_destroy+0x3c/0x220 #4: (&(&n->list_lock)->rlock){-.-...}, at: __kmem_cache_shutdown+0x54/0x320 stack backtrace: CPU: 6 PID: 8895 Comm: rmmod Tainted: G U 4.8.0-rc1-gfxbench+ #1 Hardware name: Gigabyte Technology Co., Ltd. H87M-D3H/H87M-D3H, BIOS F11 08/18/2015 Call Trace: __lock_acquire+0x1646/0x1ad0 lock_acquire+0xb2/0x200 _raw_spin_lock+0x36/0x50 get_partial_node.isra.63+0x47/0x430 ___slab_alloc.constprop.67+0x1a7/0x3b0 __slab_alloc.isra.64.constprop.66+0x43/0x80 kmem_cache_alloc+0x236/0x2d0 __debug_object_init+0x2de/0x400 debug_object_activate+0x109/0x1e0 __call_rcu.constprop.63+0x32/0x2f0 call_rcu+0x12/0x20 discard_slab+0x3d/0x40 __kmem_cache_shutdown+0xdb/0x320 shutdown_cache+0x19/0x60 kmem_cache_destroy+0x1ae/0x220 i915_gem_load_cleanup+0x14/0x40 [i915] i915_driver_unload+0x151/0x180 [i915] i915_pci_remove+0x14/0x20 [i915] pci_device_remove+0x34/0xb0 __device_release_driver+0x95/0x140 driver_detach+0xb6/0xc0 bus_remove_driver+0x53/0xd0 driver_unregister+0x27/0x50 pci_unregister_driver+0x25/0x70 i915_exit+0x1a/0x1e2 [i915] SyS_delete_module+0x193/0x1f0 entry_SYSCALL_64_fastpath+0x1c/0xac Fixes: 52b4b950b507 ("mm: slab: free kmem_cache_node after destroy sysfs file") Link: http://lkml.kernel.org/r/1470759070-18743-1-git-send-email-chris@chris-wilson.co.uk Reported-by: Dave Gordon <david.s.gordon@intel.com> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Reviewed-by: Vladimir Davydov <vdavydov@virtuozzo.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Dmitry Safonov <dsafonov@virtuozzo.com> Cc: Daniel Vetter <daniel.vetter@ffwll.ch> Cc: Dave Gordon <david.s.gordon@intel.com> Cc: <stable@vger.kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-08-11 07:27:58 +08:00
discard_slab(s, page);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
slab, slub: skip unnecessary kasan_cache_shutdown() The kasan quarantine is designed to delay freeing slab objects to catch use-after-free. The quarantine can be large (several percent of machine memory size). When kmem_caches are deleted related objects are flushed from the quarantine but this requires scanning the entire quarantine which can be very slow. We have seen the kernel busily working on this while holding slab_mutex and badly affecting cache_reaper, slabinfo readers and memcg kmem cache creations. It can easily reproduced by following script: yes . | head -1000000 | xargs stat > /dev/null for i in `seq 1 10`; do seq 500 | (cd /cg/memory && xargs mkdir) seq 500 | xargs -I{} sh -c 'echo $BASHPID > \ /cg/memory/{}/tasks && exec stat .' > /dev/null seq 500 | (cd /cg/memory && xargs rmdir) done The busy stack: kasan_cache_shutdown shutdown_cache memcg_destroy_kmem_caches mem_cgroup_css_free css_free_rwork_fn process_one_work worker_thread kthread ret_from_fork This patch is based on the observation that if the kmem_cache to be destroyed is empty then there should not be any objects of this cache in the quarantine. Without the patch the script got stuck for couple of hours. With the patch the script completed within a second. Link: http://lkml.kernel.org/r/20180327230603.54721-1-shakeelb@google.com Signed-off-by: Shakeel Butt <shakeelb@google.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Acked-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Cc: Alexander Potapenko <glider@google.com> Cc: Greg Thelen <gthelen@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-04-06 07:21:57 +08:00
bool __kmem_cache_empty(struct kmem_cache *s)
{
int node;
struct kmem_cache_node *n;
for_each_kmem_cache_node(s, node, n)
if (n->nr_partial || slabs_node(s, node))
return false;
return true;
}
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/*
* Release all resources used by a slab cache.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*/
mm: slab: free kmem_cache_node after destroy sysfs file When slub_debug alloc_calls_show is enabled we will try to track location and user of slab object on each online node, kmem_cache_node structure and cpu_cache/cpu_slub shouldn't be freed till there is the last reference to sysfs file. This fixes the following panic: BUG: unable to handle kernel NULL pointer dereference at 0000000000000020 IP: list_locations+0x169/0x4e0 PGD 257304067 PUD 438456067 PMD 0 Oops: 0000 [#1] SMP CPU: 3 PID: 973074 Comm: cat ve: 0 Not tainted 3.10.0-229.7.2.ovz.9.30-00007-japdoll-dirty #2 9.30 Hardware name: DEPO Computers To Be Filled By O.E.M./H67DE3, BIOS L1.60c 07/14/2011 task: ffff88042a5dc5b0 ti: ffff88037f8d8000 task.ti: ffff88037f8d8000 RIP: list_locations+0x169/0x4e0 Call Trace: alloc_calls_show+0x1d/0x30 slab_attr_show+0x1b/0x30 sysfs_read_file+0x9a/0x1a0 vfs_read+0x9c/0x170 SyS_read+0x58/0xb0 system_call_fastpath+0x16/0x1b Code: 5e 07 12 00 b9 00 04 00 00 3d 00 04 00 00 0f 4f c1 3d 00 04 00 00 89 45 b0 0f 84 c3 00 00 00 48 63 45 b0 49 8b 9c c4 f8 00 00 00 <48> 8b 43 20 48 85 c0 74 b6 48 89 df e8 46 37 44 00 48 8b 53 10 CR2: 0000000000000020 Separated __kmem_cache_release from __kmem_cache_shutdown which now called on slab_kmem_cache_release (after the last reference to sysfs file object has dropped). Reintroduced locking in free_partial as sysfs file might access cache's partial list after shutdowning - partial revert of the commit 69cb8e6b7c29 ("slub: free slabs without holding locks"). Zap __remove_partial and use remove_partial (w/o underscores) as free_partial now takes list_lock which s partial revert for commit 1e4dd9461fab ("slub: do not assert not having lock in removing freed partial") Signed-off-by: Dmitry Safonov <dsafonov@virtuozzo.com> Suggested-by: Vladimir Davydov <vdavydov@virtuozzo.com> Acked-by: Vladimir Davydov <vdavydov@virtuozzo.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-02-18 05:11:37 +08:00
int __kmem_cache_shutdown(struct kmem_cache *s)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
{
int node;
struct kmem_cache_node *n;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
flush_all(s);
/* Attempt to free all objects */
for_each_kmem_cache_node(s, node, n) {
free_partial(s, n);
if (n->nr_partial || slabs_node(s, node))
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
return 1;
}
return 0;
}
#ifdef CONFIG_PRINTK
mm: Add mem_dump_obj() to print source of memory block There are kernel facilities such as per-CPU reference counts that give error messages in generic handlers or callbacks, whose messages are unenlightening. In the case of per-CPU reference-count underflow, this is not a problem when creating a new use of this facility because in that case the bug is almost certainly in the code implementing that new use. However, trouble arises when deploying across many systems, which might exercise corner cases that were not seen during development and testing. Here, it would be really nice to get some kind of hint as to which of several uses the underflow was caused by. This commit therefore exposes a mem_dump_obj() function that takes a pointer to memory (which must still be allocated if it has been dynamically allocated) and prints available information on where that memory came from. This pointer can reference the middle of the block as well as the beginning of the block, as needed by things like RCU callback functions and timer handlers that might not know where the beginning of the memory block is. These functions and handlers can use mem_dump_obj() to print out better hints as to where the problem might lie. The information printed can depend on kernel configuration. For example, the allocation return address can be printed only for slab and slub, and even then only when the necessary debug has been enabled. For slab, build with CONFIG_DEBUG_SLAB=y, and either use sizes with ample space to the next power of two or use the SLAB_STORE_USER when creating the kmem_cache structure. For slub, build with CONFIG_SLUB_DEBUG=y and boot with slub_debug=U, or pass SLAB_STORE_USER to kmem_cache_create() if more focused use is desired. Also for slub, use CONFIG_STACKTRACE to enable printing of the allocation-time stack trace. Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: <linux-mm@kvack.org> Reported-by: Andrii Nakryiko <andrii@kernel.org> [ paulmck: Convert to printing and change names per Joonsoo Kim. ] [ paulmck: Move slab definition per Stephen Rothwell and kbuild test robot. ] [ paulmck: Handle CONFIG_MMU=n case where vmalloc() is kmalloc(). ] [ paulmck: Apply Vlastimil Babka feedback on slab.c kmem_provenance(). ] [ paulmck: Extract more info from !SLUB_DEBUG per Joonsoo Kim. ] [ paulmck: Explicitly check for small pointers per Naresh Kamboju. ] Acked-by: Joonsoo Kim <iamjoonsoo.kim@lge.com> Acked-by: Vlastimil Babka <vbabka@suse.cz> Tested-by: Naresh Kamboju <naresh.kamboju@linaro.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-12-08 09:41:02 +08:00
void kmem_obj_info(struct kmem_obj_info *kpp, void *object, struct page *page)
{
void *base;
int __maybe_unused i;
unsigned int objnr;
void *objp;
void *objp0;
struct kmem_cache *s = page->slab_cache;
struct track __maybe_unused *trackp;
kpp->kp_ptr = object;
kpp->kp_page = page;
kpp->kp_slab_cache = s;
base = page_address(page);
objp0 = kasan_reset_tag(object);
#ifdef CONFIG_SLUB_DEBUG
objp = restore_red_left(s, objp0);
#else
objp = objp0;
#endif
objnr = obj_to_index(s, page, objp);
kpp->kp_data_offset = (unsigned long)((char *)objp0 - (char *)objp);
objp = base + s->size * objnr;
kpp->kp_objp = objp;
if (WARN_ON_ONCE(objp < base || objp >= base + page->objects * s->size || (objp - base) % s->size) ||
!(s->flags & SLAB_STORE_USER))
return;
#ifdef CONFIG_SLUB_DEBUG
trackp = get_track(s, objp, TRACK_ALLOC);
kpp->kp_ret = (void *)trackp->addr;
#ifdef CONFIG_STACKTRACE
for (i = 0; i < KS_ADDRS_COUNT && i < TRACK_ADDRS_COUNT; i++) {
kpp->kp_stack[i] = (void *)trackp->addrs[i];
if (!kpp->kp_stack[i])
break;
}
#endif
#endif
}
#endif
mm: Add mem_dump_obj() to print source of memory block There are kernel facilities such as per-CPU reference counts that give error messages in generic handlers or callbacks, whose messages are unenlightening. In the case of per-CPU reference-count underflow, this is not a problem when creating a new use of this facility because in that case the bug is almost certainly in the code implementing that new use. However, trouble arises when deploying across many systems, which might exercise corner cases that were not seen during development and testing. Here, it would be really nice to get some kind of hint as to which of several uses the underflow was caused by. This commit therefore exposes a mem_dump_obj() function that takes a pointer to memory (which must still be allocated if it has been dynamically allocated) and prints available information on where that memory came from. This pointer can reference the middle of the block as well as the beginning of the block, as needed by things like RCU callback functions and timer handlers that might not know where the beginning of the memory block is. These functions and handlers can use mem_dump_obj() to print out better hints as to where the problem might lie. The information printed can depend on kernel configuration. For example, the allocation return address can be printed only for slab and slub, and even then only when the necessary debug has been enabled. For slab, build with CONFIG_DEBUG_SLAB=y, and either use sizes with ample space to the next power of two or use the SLAB_STORE_USER when creating the kmem_cache structure. For slub, build with CONFIG_SLUB_DEBUG=y and boot with slub_debug=U, or pass SLAB_STORE_USER to kmem_cache_create() if more focused use is desired. Also for slub, use CONFIG_STACKTRACE to enable printing of the allocation-time stack trace. Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: <linux-mm@kvack.org> Reported-by: Andrii Nakryiko <andrii@kernel.org> [ paulmck: Convert to printing and change names per Joonsoo Kim. ] [ paulmck: Move slab definition per Stephen Rothwell and kbuild test robot. ] [ paulmck: Handle CONFIG_MMU=n case where vmalloc() is kmalloc(). ] [ paulmck: Apply Vlastimil Babka feedback on slab.c kmem_provenance(). ] [ paulmck: Extract more info from !SLUB_DEBUG per Joonsoo Kim. ] [ paulmck: Explicitly check for small pointers per Naresh Kamboju. ] Acked-by: Joonsoo Kim <iamjoonsoo.kim@lge.com> Acked-by: Vlastimil Babka <vbabka@suse.cz> Tested-by: Naresh Kamboju <naresh.kamboju@linaro.org> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2020-12-08 09:41:02 +08:00
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/********************************************************************
* Kmalloc subsystem
*******************************************************************/
static int __init setup_slub_min_order(char *str)
{
get_option(&str, (int *)&slub_min_order);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
return 1;
}
__setup("slub_min_order=", setup_slub_min_order);
static int __init setup_slub_max_order(char *str)
{
get_option(&str, (int *)&slub_max_order);
slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
return 1;
}
__setup("slub_max_order=", setup_slub_max_order);
static int __init setup_slub_min_objects(char *str)
{
get_option(&str, (int *)&slub_min_objects);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
return 1;
}
__setup("slub_min_objects=", setup_slub_min_objects);
void *__kmalloc(size_t size, gfp_t flags)
{
SLUB: direct pass through of page size or higher kmalloc requests This gets rid of all kmalloc caches larger than page size. A kmalloc request larger than PAGE_SIZE > 2 is going to be passed through to the page allocator. This works both inline where we will call __get_free_pages instead of kmem_cache_alloc and in __kmalloc. kfree is modified to check if the object is in a slab page. If not then the page is freed via the page allocator instead. Roughly similar to what SLOB does. Advantages: - Reduces memory overhead for kmalloc array - Large kmalloc operations are faster since they do not need to pass through the slab allocator to get to the page allocator. - Performance increase of 10%-20% on alloc and 50% on free for PAGE_SIZEd allocations. SLUB must call page allocator for each alloc anyways since the higher order pages which that allowed avoiding the page alloc calls are not available in a reliable way anymore. So we are basically removing useless slab allocator overhead. - Large kmallocs yields page aligned object which is what SLAB did. Bad things like using page sized kmalloc allocations to stand in for page allocate allocs can be transparently handled and are not distinguishable from page allocator uses. - Checking for too large objects can be removed since it is done by the page allocator. Drawbacks: - No accounting for large kmalloc slab allocations anymore - No debugging of large kmalloc slab allocations. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 16:24:38 +08:00
struct kmem_cache *s;
void *ret;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
return kmalloc_large(size, flags);
SLUB: direct pass through of page size or higher kmalloc requests This gets rid of all kmalloc caches larger than page size. A kmalloc request larger than PAGE_SIZE > 2 is going to be passed through to the page allocator. This works both inline where we will call __get_free_pages instead of kmem_cache_alloc and in __kmalloc. kfree is modified to check if the object is in a slab page. If not then the page is freed via the page allocator instead. Roughly similar to what SLOB does. Advantages: - Reduces memory overhead for kmalloc array - Large kmalloc operations are faster since they do not need to pass through the slab allocator to get to the page allocator. - Performance increase of 10%-20% on alloc and 50% on free for PAGE_SIZEd allocations. SLUB must call page allocator for each alloc anyways since the higher order pages which that allowed avoiding the page alloc calls are not available in a reliable way anymore. So we are basically removing useless slab allocator overhead. - Large kmallocs yields page aligned object which is what SLAB did. Bad things like using page sized kmalloc allocations to stand in for page allocate allocs can be transparently handled and are not distinguishable from page allocator uses. - Checking for too large objects can be removed since it is done by the page allocator. Drawbacks: - No accounting for large kmalloc slab allocations anymore - No debugging of large kmalloc slab allocations. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 16:24:38 +08:00
s = kmalloc_slab(size, flags);
SLUB: direct pass through of page size or higher kmalloc requests This gets rid of all kmalloc caches larger than page size. A kmalloc request larger than PAGE_SIZE > 2 is going to be passed through to the page allocator. This works both inline where we will call __get_free_pages instead of kmem_cache_alloc and in __kmalloc. kfree is modified to check if the object is in a slab page. If not then the page is freed via the page allocator instead. Roughly similar to what SLOB does. Advantages: - Reduces memory overhead for kmalloc array - Large kmalloc operations are faster since they do not need to pass through the slab allocator to get to the page allocator. - Performance increase of 10%-20% on alloc and 50% on free for PAGE_SIZEd allocations. SLUB must call page allocator for each alloc anyways since the higher order pages which that allowed avoiding the page alloc calls are not available in a reliable way anymore. So we are basically removing useless slab allocator overhead. - Large kmallocs yields page aligned object which is what SLAB did. Bad things like using page sized kmalloc allocations to stand in for page allocate allocs can be transparently handled and are not distinguishable from page allocator uses. - Checking for too large objects can be removed since it is done by the page allocator. Drawbacks: - No accounting for large kmalloc slab allocations anymore - No debugging of large kmalloc slab allocations. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 16:24:38 +08:00
if (unlikely(ZERO_OR_NULL_PTR(s)))
return s;
mm, kfence: insert KFENCE hooks for SLUB Inserts KFENCE hooks into the SLUB allocator. To pass the originally requested size to KFENCE, add an argument 'orig_size' to slab_alloc*(). The additional argument is required to preserve the requested original size for kmalloc() allocations, which uses size classes (e.g. an allocation of 272 bytes will return an object of size 512). Therefore, kmem_cache::size does not represent the kmalloc-caller's requested size, and we must introduce the argument 'orig_size' to propagate the originally requested size to KFENCE. Without the originally requested size, we would not be able to detect out-of-bounds accesses for objects placed at the end of a KFENCE object page if that object is not equal to the kmalloc-size class it was bucketed into. When KFENCE is disabled, there is no additional overhead, since slab_alloc*() functions are __always_inline. Link: https://lkml.kernel.org/r/20201103175841.3495947-6-elver@google.com Signed-off-by: Marco Elver <elver@google.com> Signed-off-by: Alexander Potapenko <glider@google.com> Reviewed-by: Dmitry Vyukov <dvyukov@google.com> Reviewed-by: Jann Horn <jannh@google.com> Co-developed-by: Marco Elver <elver@google.com> Cc: Andrey Konovalov <andreyknvl@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christopher Lameter <cl@linux.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: David Rientjes <rientjes@google.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Hillf Danton <hdanton@sina.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Joern Engel <joern@purestorage.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kees Cook <keescook@chromium.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Paul E. McKenney <paulmck@kernel.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: SeongJae Park <sjpark@amazon.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-26 09:19:16 +08:00
ret = slab_alloc(s, flags, _RET_IP_, size);
trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
kasan, mm: change hooks signatures Patch series "kasan: add software tag-based mode for arm64", v13. This patchset adds a new software tag-based mode to KASAN [1]. (Initially this mode was called KHWASAN, but it got renamed, see the naming rationale at the end of this section). The plan is to implement HWASan [2] for the kernel with the incentive, that it's going to have comparable to KASAN performance, but in the same time consume much less memory, trading that off for somewhat imprecise bug detection and being supported only for arm64. The underlying ideas of the approach used by software tag-based KASAN are: 1. By using the Top Byte Ignore (TBI) arm64 CPU feature, we can store pointer tags in the top byte of each kernel pointer. 2. Using shadow memory, we can store memory tags for each chunk of kernel memory. 3. On each memory allocation, we can generate a random tag, embed it into the returned pointer and set the memory tags that correspond to this chunk of memory to the same value. 4. By using compiler instrumentation, before each memory access we can add a check that the pointer tag matches the tag of the memory that is being accessed. 5. On a tag mismatch we report an error. With this patchset the existing KASAN mode gets renamed to generic KASAN, with the word "generic" meaning that the implementation can be supported by any architecture as it is purely software. The new mode this patchset adds is called software tag-based KASAN. The word "tag-based" refers to the fact that this mode uses tags embedded into the top byte of kernel pointers and the TBI arm64 CPU feature that allows to dereference such pointers. The word "software" here means that shadow memory manipulation and tag checking on pointer dereference is done in software. As it is the only tag-based implementation right now, "software tag-based" KASAN is sometimes referred to as simply "tag-based" in this patchset. A potential expansion of this mode is a hardware tag-based mode, which would use hardware memory tagging support (announced by Arm [3]) instead of compiler instrumentation and manual shadow memory manipulation. Same as generic KASAN, software tag-based KASAN is strictly a debugging feature. [1] https://www.kernel.org/doc/html/latest/dev-tools/kasan.html [2] http://clang.llvm.org/docs/HardwareAssistedAddressSanitizerDesign.html [3] https://community.arm.com/processors/b/blog/posts/arm-a-profile-architecture-2018-developments-armv85a ====== Rationale On mobile devices generic KASAN's memory usage is significant problem. One of the main reasons to have tag-based KASAN is to be able to perform a similar set of checks as the generic one does, but with lower memory requirements. Comment from Vishwath Mohan <vishwath@google.com>: I don't have data on-hand, but anecdotally both ASAN and KASAN have proven problematic to enable for environments that don't tolerate the increased memory pressure well. This includes (a) Low-memory form factors - Wear, TV, Things, lower-tier phones like Go, (c) Connected components like Pixel's visual core [1]. These are both places I'd love to have a low(er) memory footprint option at my disposal. Comment from Evgenii Stepanov <eugenis@google.com>: Looking at a live Android device under load, slab (according to /proc/meminfo) + kernel stack take 8-10% available RAM (~350MB). KASAN's overhead of 2x - 3x on top of it is not insignificant. Not having this overhead enables near-production use - ex. running KASAN/KHWASAN kernel on a personal, daily-use device to catch bugs that do not reproduce in test configuration. These are the ones that often cost the most engineering time to track down. CPU overhead is bad, but generally tolerable. RAM is critical, in our experience. Once it gets low enough, OOM-killer makes your life miserable. [1] https://www.blog.google/products/pixel/pixel-visual-core-image-processing-and-machine-learning-pixel-2/ ====== Technical details Software tag-based KASAN mode is implemented in a very similar way to the generic one. This patchset essentially does the following: 1. TCR_TBI1 is set to enable Top Byte Ignore. 2. Shadow memory is used (with a different scale, 1:16, so each shadow byte corresponds to 16 bytes of kernel memory) to store memory tags. 3. All slab objects are aligned to shadow scale, which is 16 bytes. 4. All pointers returned from the slab allocator are tagged with a random tag and the corresponding shadow memory is poisoned with the same value. 5. Compiler instrumentation is used to insert tag checks. Either by calling callbacks or by inlining them (CONFIG_KASAN_OUTLINE and CONFIG_KASAN_INLINE flags are reused). 6. When a tag mismatch is detected in callback instrumentation mode KASAN simply prints a bug report. In case of inline instrumentation, clang inserts a brk instruction, and KASAN has it's own brk handler, which reports the bug. 7. The memory in between slab objects is marked with a reserved tag, and acts as a redzone. 8. When a slab object is freed it's marked with a reserved tag. Bug detection is imprecise for two reasons: 1. We won't catch some small out-of-bounds accesses, that fall into the same shadow cell, as the last byte of a slab object. 2. We only have 1 byte to store tags, which means we have a 1/256 probability of a tag match for an incorrect access (actually even slightly less due to reserved tag values). Despite that there's a particular type of bugs that tag-based KASAN can detect compared to generic KASAN: use-after-free after the object has been allocated by someone else. ====== Testing Some kernel developers voiced a concern that changing the top byte of kernel pointers may lead to subtle bugs that are difficult to discover. To address this concern deliberate testing has been performed. It doesn't seem feasible to do some kind of static checking to find potential issues with pointer tagging, so a dynamic approach was taken. All pointer comparisons/subtractions have been instrumented in an LLVM compiler pass and a kernel module that would print a bug report whenever two pointers with different tags are being compared/subtracted (ignoring comparisons with NULL pointers and with pointers obtained by casting an error code to a pointer type) has been used. Then the kernel has been booted in QEMU and on an Odroid C2 board and syzkaller has been run. This yielded the following results. The two places that look interesting are: is_vmalloc_addr in include/linux/mm.h is_kernel_rodata in mm/util.c Here we compare a pointer with some fixed untagged values to make sure that the pointer lies in a particular part of the kernel address space. Since tag-based KASAN doesn't add tags to pointers that belong to rodata or vmalloc regions, this should work as is. To make sure debug checks to those two functions that check that the result doesn't change whether we operate on pointers with or without untagging has been added. A few other cases that don't look that interesting: Comparing pointers to achieve unique sorting order of pointee objects (e.g. sorting locks addresses before performing a double lock): tty_ldisc_lock_pair_timeout in drivers/tty/tty_ldisc.c pipe_double_lock in fs/pipe.c unix_state_double_lock in net/unix/af_unix.c lock_two_nondirectories in fs/inode.c mutex_lock_double in kernel/events/core.c ep_cmp_ffd in fs/eventpoll.c fsnotify_compare_groups fs/notify/mark.c Nothing needs to be done here, since the tags embedded into pointers don't change, so the sorting order would still be unique. Checks that a pointer belongs to some particular allocation: is_sibling_entry in lib/radix-tree.c object_is_on_stack in include/linux/sched/task_stack.h Nothing needs to be done here either, since two pointers can only belong to the same allocation if they have the same tag. Overall, since the kernel boots and works, there are no critical bugs. As for the rest, the traditional kernel testing way (use until fails) is the only one that looks feasible. Another point here is that tag-based KASAN is available under a separate config option that needs to be deliberately enabled. Even though it might be used in a "near-production" environment to find bugs that are not found during fuzzing or running tests, it is still a debug tool. ====== Benchmarks The following numbers were collected on Odroid C2 board. Both generic and tag-based KASAN were used in inline instrumentation mode. Boot time [1]: * ~1.7 sec for clean kernel * ~5.0 sec for generic KASAN * ~5.0 sec for tag-based KASAN Network performance [2]: * 8.33 Gbits/sec for clean kernel * 3.17 Gbits/sec for generic KASAN * 2.85 Gbits/sec for tag-based KASAN Slab memory usage after boot [3]: * ~40 kb for clean kernel * ~105 kb (~260% overhead) for generic KASAN * ~47 kb (~20% overhead) for tag-based KASAN KASAN memory overhead consists of three main parts: 1. Increased slab memory usage due to redzones. 2. Shadow memory (the whole reserved once during boot). 3. Quaratine (grows gradually until some preset limit; the more the limit, the more the chance to detect a use-after-free). Comparing tag-based vs generic KASAN for each of these points: 1. 20% vs 260% overhead. 2. 1/16th vs 1/8th of physical memory. 3. Tag-based KASAN doesn't require quarantine. [1] Time before the ext4 driver is initialized. [2] Measured as `iperf -s & iperf -c 127.0.0.1 -t 30`. [3] Measured as `cat /proc/meminfo | grep Slab`. ====== Some notes A few notes: 1. The patchset can be found here: https://github.com/xairy/kasan-prototype/tree/khwasan 2. Building requires a recent Clang version (7.0.0 or later). 3. Stack instrumentation is not supported yet and will be added later. This patch (of 25): Tag-based KASAN changes the value of the top byte of pointers returned from the kernel allocation functions (such as kmalloc). This patch updates KASAN hooks signatures and their usage in SLAB and SLUB code to reflect that. Link: http://lkml.kernel.org/r/aec2b5e3973781ff8a6bb6760f8543643202c451.1544099024.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Reviewed-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Reviewed-by: Dmitry Vyukov <dvyukov@google.com> Cc: Christoph Lameter <cl@linux.com> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 16:29:37 +08:00
ret = kasan_kmalloc(s, ret, size, flags);
mm: slub: add kernel address sanitizer support for slub allocator With this patch kasan will be able to catch bugs in memory allocated by slub. Initially all objects in newly allocated slab page, marked as redzone. Later, when allocation of slub object happens, requested by caller number of bytes marked as accessible, and the rest of the object (including slub's metadata) marked as redzone (inaccessible). We also mark object as accessible if ksize was called for this object. There is some places in kernel where ksize function is called to inquire size of really allocated area. Such callers could validly access whole allocated memory, so it should be marked as accessible. Code in slub.c and slab_common.c files could validly access to object's metadata, so instrumentation for this files are disabled. Signed-off-by: Andrey Ryabinin <a.ryabinin@samsung.com> Signed-off-by: Dmitry Chernenkov <dmitryc@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Konstantin Serebryany <kcc@google.com> Signed-off-by: Andrey Konovalov <adech.fo@gmail.com> Cc: Yuri Gribov <tetra2005@gmail.com> Cc: Konstantin Khlebnikov <koct9i@gmail.com> Cc: Sasha Levin <sasha.levin@oracle.com> Cc: Christoph Lameter <cl@linux.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Ingo Molnar <mingo@elte.hu> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-14 06:39:42 +08:00
return ret;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
EXPORT_SYMBOL(__kmalloc);
#ifdef CONFIG_NUMA
static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
{
struct page *page;
void *ptr = NULL;
mm, sl[ou]b: improve memory accounting Patch series "guarantee natural alignment for kmalloc()", v2. This patch (of 2): SLOB currently doesn't account its pages at all, so in /proc/meminfo the Slab field shows zero. Modifying a counter on page allocation and freeing should be acceptable even for the small system scenarios SLOB is intended for. Since reclaimable caches are not separated in SLOB, account everything as unreclaimable. SLUB currently doesn't account kmalloc() and kmalloc_node() allocations larger than order-1 page, that are passed directly to the page allocator. As they also don't appear in /proc/slabinfo, it might look like a memory leak. For consistency, account them as well. (SLAB doesn't actually use page allocator directly, so no change there). Ideally SLOB and SLUB would be handled in separate patches, but due to the shared kmalloc_order() function and different kfree() implementations, it's easier to patch both at once to prevent inconsistencies. Link: http://lkml.kernel.org/r/20190826111627.7505-2-vbabka@suse.cz Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Ming Lei <ming.lei@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: "Darrick J . Wong" <darrick.wong@oracle.com> Cc: Christoph Hellwig <hch@lst.de> Cc: James Bottomley <James.Bottomley@HansenPartnership.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-10-07 08:58:42 +08:00
unsigned int order = get_order(size);
flags |= __GFP_COMP;
mm, sl[ou]b: improve memory accounting Patch series "guarantee natural alignment for kmalloc()", v2. This patch (of 2): SLOB currently doesn't account its pages at all, so in /proc/meminfo the Slab field shows zero. Modifying a counter on page allocation and freeing should be acceptable even for the small system scenarios SLOB is intended for. Since reclaimable caches are not separated in SLOB, account everything as unreclaimable. SLUB currently doesn't account kmalloc() and kmalloc_node() allocations larger than order-1 page, that are passed directly to the page allocator. As they also don't appear in /proc/slabinfo, it might look like a memory leak. For consistency, account them as well. (SLAB doesn't actually use page allocator directly, so no change there). Ideally SLOB and SLUB would be handled in separate patches, but due to the shared kmalloc_order() function and different kfree() implementations, it's easier to patch both at once to prevent inconsistencies. Link: http://lkml.kernel.org/r/20190826111627.7505-2-vbabka@suse.cz Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Ming Lei <ming.lei@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: "Darrick J . Wong" <darrick.wong@oracle.com> Cc: Christoph Hellwig <hch@lst.de> Cc: James Bottomley <James.Bottomley@HansenPartnership.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-10-07 08:58:42 +08:00
page = alloc_pages_node(node, flags, order);
if (page) {
ptr = page_address(page);
mod_lruvec_page_state(page, NR_SLAB_UNRECLAIMABLE_B,
PAGE_SIZE << order);
mm, sl[ou]b: improve memory accounting Patch series "guarantee natural alignment for kmalloc()", v2. This patch (of 2): SLOB currently doesn't account its pages at all, so in /proc/meminfo the Slab field shows zero. Modifying a counter on page allocation and freeing should be acceptable even for the small system scenarios SLOB is intended for. Since reclaimable caches are not separated in SLOB, account everything as unreclaimable. SLUB currently doesn't account kmalloc() and kmalloc_node() allocations larger than order-1 page, that are passed directly to the page allocator. As they also don't appear in /proc/slabinfo, it might look like a memory leak. For consistency, account them as well. (SLAB doesn't actually use page allocator directly, so no change there). Ideally SLOB and SLUB would be handled in separate patches, but due to the shared kmalloc_order() function and different kfree() implementations, it's easier to patch both at once to prevent inconsistencies. Link: http://lkml.kernel.org/r/20190826111627.7505-2-vbabka@suse.cz Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Ming Lei <ming.lei@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: "Darrick J . Wong" <darrick.wong@oracle.com> Cc: Christoph Hellwig <hch@lst.de> Cc: James Bottomley <James.Bottomley@HansenPartnership.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-10-07 08:58:42 +08:00
}
kasan, mm: change hooks signatures Patch series "kasan: add software tag-based mode for arm64", v13. This patchset adds a new software tag-based mode to KASAN [1]. (Initially this mode was called KHWASAN, but it got renamed, see the naming rationale at the end of this section). The plan is to implement HWASan [2] for the kernel with the incentive, that it's going to have comparable to KASAN performance, but in the same time consume much less memory, trading that off for somewhat imprecise bug detection and being supported only for arm64. The underlying ideas of the approach used by software tag-based KASAN are: 1. By using the Top Byte Ignore (TBI) arm64 CPU feature, we can store pointer tags in the top byte of each kernel pointer. 2. Using shadow memory, we can store memory tags for each chunk of kernel memory. 3. On each memory allocation, we can generate a random tag, embed it into the returned pointer and set the memory tags that correspond to this chunk of memory to the same value. 4. By using compiler instrumentation, before each memory access we can add a check that the pointer tag matches the tag of the memory that is being accessed. 5. On a tag mismatch we report an error. With this patchset the existing KASAN mode gets renamed to generic KASAN, with the word "generic" meaning that the implementation can be supported by any architecture as it is purely software. The new mode this patchset adds is called software tag-based KASAN. The word "tag-based" refers to the fact that this mode uses tags embedded into the top byte of kernel pointers and the TBI arm64 CPU feature that allows to dereference such pointers. The word "software" here means that shadow memory manipulation and tag checking on pointer dereference is done in software. As it is the only tag-based implementation right now, "software tag-based" KASAN is sometimes referred to as simply "tag-based" in this patchset. A potential expansion of this mode is a hardware tag-based mode, which would use hardware memory tagging support (announced by Arm [3]) instead of compiler instrumentation and manual shadow memory manipulation. Same as generic KASAN, software tag-based KASAN is strictly a debugging feature. [1] https://www.kernel.org/doc/html/latest/dev-tools/kasan.html [2] http://clang.llvm.org/docs/HardwareAssistedAddressSanitizerDesign.html [3] https://community.arm.com/processors/b/blog/posts/arm-a-profile-architecture-2018-developments-armv85a ====== Rationale On mobile devices generic KASAN's memory usage is significant problem. One of the main reasons to have tag-based KASAN is to be able to perform a similar set of checks as the generic one does, but with lower memory requirements. Comment from Vishwath Mohan <vishwath@google.com>: I don't have data on-hand, but anecdotally both ASAN and KASAN have proven problematic to enable for environments that don't tolerate the increased memory pressure well. This includes (a) Low-memory form factors - Wear, TV, Things, lower-tier phones like Go, (c) Connected components like Pixel's visual core [1]. These are both places I'd love to have a low(er) memory footprint option at my disposal. Comment from Evgenii Stepanov <eugenis@google.com>: Looking at a live Android device under load, slab (according to /proc/meminfo) + kernel stack take 8-10% available RAM (~350MB). KASAN's overhead of 2x - 3x on top of it is not insignificant. Not having this overhead enables near-production use - ex. running KASAN/KHWASAN kernel on a personal, daily-use device to catch bugs that do not reproduce in test configuration. These are the ones that often cost the most engineering time to track down. CPU overhead is bad, but generally tolerable. RAM is critical, in our experience. Once it gets low enough, OOM-killer makes your life miserable. [1] https://www.blog.google/products/pixel/pixel-visual-core-image-processing-and-machine-learning-pixel-2/ ====== Technical details Software tag-based KASAN mode is implemented in a very similar way to the generic one. This patchset essentially does the following: 1. TCR_TBI1 is set to enable Top Byte Ignore. 2. Shadow memory is used (with a different scale, 1:16, so each shadow byte corresponds to 16 bytes of kernel memory) to store memory tags. 3. All slab objects are aligned to shadow scale, which is 16 bytes. 4. All pointers returned from the slab allocator are tagged with a random tag and the corresponding shadow memory is poisoned with the same value. 5. Compiler instrumentation is used to insert tag checks. Either by calling callbacks or by inlining them (CONFIG_KASAN_OUTLINE and CONFIG_KASAN_INLINE flags are reused). 6. When a tag mismatch is detected in callback instrumentation mode KASAN simply prints a bug report. In case of inline instrumentation, clang inserts a brk instruction, and KASAN has it's own brk handler, which reports the bug. 7. The memory in between slab objects is marked with a reserved tag, and acts as a redzone. 8. When a slab object is freed it's marked with a reserved tag. Bug detection is imprecise for two reasons: 1. We won't catch some small out-of-bounds accesses, that fall into the same shadow cell, as the last byte of a slab object. 2. We only have 1 byte to store tags, which means we have a 1/256 probability of a tag match for an incorrect access (actually even slightly less due to reserved tag values). Despite that there's a particular type of bugs that tag-based KASAN can detect compared to generic KASAN: use-after-free after the object has been allocated by someone else. ====== Testing Some kernel developers voiced a concern that changing the top byte of kernel pointers may lead to subtle bugs that are difficult to discover. To address this concern deliberate testing has been performed. It doesn't seem feasible to do some kind of static checking to find potential issues with pointer tagging, so a dynamic approach was taken. All pointer comparisons/subtractions have been instrumented in an LLVM compiler pass and a kernel module that would print a bug report whenever two pointers with different tags are being compared/subtracted (ignoring comparisons with NULL pointers and with pointers obtained by casting an error code to a pointer type) has been used. Then the kernel has been booted in QEMU and on an Odroid C2 board and syzkaller has been run. This yielded the following results. The two places that look interesting are: is_vmalloc_addr in include/linux/mm.h is_kernel_rodata in mm/util.c Here we compare a pointer with some fixed untagged values to make sure that the pointer lies in a particular part of the kernel address space. Since tag-based KASAN doesn't add tags to pointers that belong to rodata or vmalloc regions, this should work as is. To make sure debug checks to those two functions that check that the result doesn't change whether we operate on pointers with or without untagging has been added. A few other cases that don't look that interesting: Comparing pointers to achieve unique sorting order of pointee objects (e.g. sorting locks addresses before performing a double lock): tty_ldisc_lock_pair_timeout in drivers/tty/tty_ldisc.c pipe_double_lock in fs/pipe.c unix_state_double_lock in net/unix/af_unix.c lock_two_nondirectories in fs/inode.c mutex_lock_double in kernel/events/core.c ep_cmp_ffd in fs/eventpoll.c fsnotify_compare_groups fs/notify/mark.c Nothing needs to be done here, since the tags embedded into pointers don't change, so the sorting order would still be unique. Checks that a pointer belongs to some particular allocation: is_sibling_entry in lib/radix-tree.c object_is_on_stack in include/linux/sched/task_stack.h Nothing needs to be done here either, since two pointers can only belong to the same allocation if they have the same tag. Overall, since the kernel boots and works, there are no critical bugs. As for the rest, the traditional kernel testing way (use until fails) is the only one that looks feasible. Another point here is that tag-based KASAN is available under a separate config option that needs to be deliberately enabled. Even though it might be used in a "near-production" environment to find bugs that are not found during fuzzing or running tests, it is still a debug tool. ====== Benchmarks The following numbers were collected on Odroid C2 board. Both generic and tag-based KASAN were used in inline instrumentation mode. Boot time [1]: * ~1.7 sec for clean kernel * ~5.0 sec for generic KASAN * ~5.0 sec for tag-based KASAN Network performance [2]: * 8.33 Gbits/sec for clean kernel * 3.17 Gbits/sec for generic KASAN * 2.85 Gbits/sec for tag-based KASAN Slab memory usage after boot [3]: * ~40 kb for clean kernel * ~105 kb (~260% overhead) for generic KASAN * ~47 kb (~20% overhead) for tag-based KASAN KASAN memory overhead consists of three main parts: 1. Increased slab memory usage due to redzones. 2. Shadow memory (the whole reserved once during boot). 3. Quaratine (grows gradually until some preset limit; the more the limit, the more the chance to detect a use-after-free). Comparing tag-based vs generic KASAN for each of these points: 1. 20% vs 260% overhead. 2. 1/16th vs 1/8th of physical memory. 3. Tag-based KASAN doesn't require quarantine. [1] Time before the ext4 driver is initialized. [2] Measured as `iperf -s & iperf -c 127.0.0.1 -t 30`. [3] Measured as `cat /proc/meminfo | grep Slab`. ====== Some notes A few notes: 1. The patchset can be found here: https://github.com/xairy/kasan-prototype/tree/khwasan 2. Building requires a recent Clang version (7.0.0 or later). 3. Stack instrumentation is not supported yet and will be added later. This patch (of 25): Tag-based KASAN changes the value of the top byte of pointers returned from the kernel allocation functions (such as kmalloc). This patch updates KASAN hooks signatures and their usage in SLAB and SLUB code to reflect that. Link: http://lkml.kernel.org/r/aec2b5e3973781ff8a6bb6760f8543643202c451.1544099024.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Reviewed-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Reviewed-by: Dmitry Vyukov <dvyukov@google.com> Cc: Christoph Lameter <cl@linux.com> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 16:29:37 +08:00
return kmalloc_large_node_hook(ptr, size, flags);
}
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
void *__kmalloc_node(size_t size, gfp_t flags, int node)
{
SLUB: direct pass through of page size or higher kmalloc requests This gets rid of all kmalloc caches larger than page size. A kmalloc request larger than PAGE_SIZE > 2 is going to be passed through to the page allocator. This works both inline where we will call __get_free_pages instead of kmem_cache_alloc and in __kmalloc. kfree is modified to check if the object is in a slab page. If not then the page is freed via the page allocator instead. Roughly similar to what SLOB does. Advantages: - Reduces memory overhead for kmalloc array - Large kmalloc operations are faster since they do not need to pass through the slab allocator to get to the page allocator. - Performance increase of 10%-20% on alloc and 50% on free for PAGE_SIZEd allocations. SLUB must call page allocator for each alloc anyways since the higher order pages which that allowed avoiding the page alloc calls are not available in a reliable way anymore. So we are basically removing useless slab allocator overhead. - Large kmallocs yields page aligned object which is what SLAB did. Bad things like using page sized kmalloc allocations to stand in for page allocate allocs can be transparently handled and are not distinguishable from page allocator uses. - Checking for too large objects can be removed since it is done by the page allocator. Drawbacks: - No accounting for large kmalloc slab allocations anymore - No debugging of large kmalloc slab allocations. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 16:24:38 +08:00
struct kmem_cache *s;
void *ret;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
ret = kmalloc_large_node(size, flags, node);
trace_kmalloc_node(_RET_IP_, ret,
size, PAGE_SIZE << get_order(size),
flags, node);
return ret;
}
SLUB: direct pass through of page size or higher kmalloc requests This gets rid of all kmalloc caches larger than page size. A kmalloc request larger than PAGE_SIZE > 2 is going to be passed through to the page allocator. This works both inline where we will call __get_free_pages instead of kmem_cache_alloc and in __kmalloc. kfree is modified to check if the object is in a slab page. If not then the page is freed via the page allocator instead. Roughly similar to what SLOB does. Advantages: - Reduces memory overhead for kmalloc array - Large kmalloc operations are faster since they do not need to pass through the slab allocator to get to the page allocator. - Performance increase of 10%-20% on alloc and 50% on free for PAGE_SIZEd allocations. SLUB must call page allocator for each alloc anyways since the higher order pages which that allowed avoiding the page alloc calls are not available in a reliable way anymore. So we are basically removing useless slab allocator overhead. - Large kmallocs yields page aligned object which is what SLAB did. Bad things like using page sized kmalloc allocations to stand in for page allocate allocs can be transparently handled and are not distinguishable from page allocator uses. - Checking for too large objects can be removed since it is done by the page allocator. Drawbacks: - No accounting for large kmalloc slab allocations anymore - No debugging of large kmalloc slab allocations. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 16:24:38 +08:00
s = kmalloc_slab(size, flags);
SLUB: direct pass through of page size or higher kmalloc requests This gets rid of all kmalloc caches larger than page size. A kmalloc request larger than PAGE_SIZE > 2 is going to be passed through to the page allocator. This works both inline where we will call __get_free_pages instead of kmem_cache_alloc and in __kmalloc. kfree is modified to check if the object is in a slab page. If not then the page is freed via the page allocator instead. Roughly similar to what SLOB does. Advantages: - Reduces memory overhead for kmalloc array - Large kmalloc operations are faster since they do not need to pass through the slab allocator to get to the page allocator. - Performance increase of 10%-20% on alloc and 50% on free for PAGE_SIZEd allocations. SLUB must call page allocator for each alloc anyways since the higher order pages which that allowed avoiding the page alloc calls are not available in a reliable way anymore. So we are basically removing useless slab allocator overhead. - Large kmallocs yields page aligned object which is what SLAB did. Bad things like using page sized kmalloc allocations to stand in for page allocate allocs can be transparently handled and are not distinguishable from page allocator uses. - Checking for too large objects can be removed since it is done by the page allocator. Drawbacks: - No accounting for large kmalloc slab allocations anymore - No debugging of large kmalloc slab allocations. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 16:24:38 +08:00
if (unlikely(ZERO_OR_NULL_PTR(s)))
return s;
mm, kfence: insert KFENCE hooks for SLUB Inserts KFENCE hooks into the SLUB allocator. To pass the originally requested size to KFENCE, add an argument 'orig_size' to slab_alloc*(). The additional argument is required to preserve the requested original size for kmalloc() allocations, which uses size classes (e.g. an allocation of 272 bytes will return an object of size 512). Therefore, kmem_cache::size does not represent the kmalloc-caller's requested size, and we must introduce the argument 'orig_size' to propagate the originally requested size to KFENCE. Without the originally requested size, we would not be able to detect out-of-bounds accesses for objects placed at the end of a KFENCE object page if that object is not equal to the kmalloc-size class it was bucketed into. When KFENCE is disabled, there is no additional overhead, since slab_alloc*() functions are __always_inline. Link: https://lkml.kernel.org/r/20201103175841.3495947-6-elver@google.com Signed-off-by: Marco Elver <elver@google.com> Signed-off-by: Alexander Potapenko <glider@google.com> Reviewed-by: Dmitry Vyukov <dvyukov@google.com> Reviewed-by: Jann Horn <jannh@google.com> Co-developed-by: Marco Elver <elver@google.com> Cc: Andrey Konovalov <andreyknvl@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christopher Lameter <cl@linux.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: David Rientjes <rientjes@google.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Hillf Danton <hdanton@sina.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Joern Engel <joern@purestorage.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kees Cook <keescook@chromium.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Paul E. McKenney <paulmck@kernel.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: SeongJae Park <sjpark@amazon.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-26 09:19:16 +08:00
ret = slab_alloc_node(s, flags, node, _RET_IP_, size);
trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
kasan, mm: change hooks signatures Patch series "kasan: add software tag-based mode for arm64", v13. This patchset adds a new software tag-based mode to KASAN [1]. (Initially this mode was called KHWASAN, but it got renamed, see the naming rationale at the end of this section). The plan is to implement HWASan [2] for the kernel with the incentive, that it's going to have comparable to KASAN performance, but in the same time consume much less memory, trading that off for somewhat imprecise bug detection and being supported only for arm64. The underlying ideas of the approach used by software tag-based KASAN are: 1. By using the Top Byte Ignore (TBI) arm64 CPU feature, we can store pointer tags in the top byte of each kernel pointer. 2. Using shadow memory, we can store memory tags for each chunk of kernel memory. 3. On each memory allocation, we can generate a random tag, embed it into the returned pointer and set the memory tags that correspond to this chunk of memory to the same value. 4. By using compiler instrumentation, before each memory access we can add a check that the pointer tag matches the tag of the memory that is being accessed. 5. On a tag mismatch we report an error. With this patchset the existing KASAN mode gets renamed to generic KASAN, with the word "generic" meaning that the implementation can be supported by any architecture as it is purely software. The new mode this patchset adds is called software tag-based KASAN. The word "tag-based" refers to the fact that this mode uses tags embedded into the top byte of kernel pointers and the TBI arm64 CPU feature that allows to dereference such pointers. The word "software" here means that shadow memory manipulation and tag checking on pointer dereference is done in software. As it is the only tag-based implementation right now, "software tag-based" KASAN is sometimes referred to as simply "tag-based" in this patchset. A potential expansion of this mode is a hardware tag-based mode, which would use hardware memory tagging support (announced by Arm [3]) instead of compiler instrumentation and manual shadow memory manipulation. Same as generic KASAN, software tag-based KASAN is strictly a debugging feature. [1] https://www.kernel.org/doc/html/latest/dev-tools/kasan.html [2] http://clang.llvm.org/docs/HardwareAssistedAddressSanitizerDesign.html [3] https://community.arm.com/processors/b/blog/posts/arm-a-profile-architecture-2018-developments-armv85a ====== Rationale On mobile devices generic KASAN's memory usage is significant problem. One of the main reasons to have tag-based KASAN is to be able to perform a similar set of checks as the generic one does, but with lower memory requirements. Comment from Vishwath Mohan <vishwath@google.com>: I don't have data on-hand, but anecdotally both ASAN and KASAN have proven problematic to enable for environments that don't tolerate the increased memory pressure well. This includes (a) Low-memory form factors - Wear, TV, Things, lower-tier phones like Go, (c) Connected components like Pixel's visual core [1]. These are both places I'd love to have a low(er) memory footprint option at my disposal. Comment from Evgenii Stepanov <eugenis@google.com>: Looking at a live Android device under load, slab (according to /proc/meminfo) + kernel stack take 8-10% available RAM (~350MB). KASAN's overhead of 2x - 3x on top of it is not insignificant. Not having this overhead enables near-production use - ex. running KASAN/KHWASAN kernel on a personal, daily-use device to catch bugs that do not reproduce in test configuration. These are the ones that often cost the most engineering time to track down. CPU overhead is bad, but generally tolerable. RAM is critical, in our experience. Once it gets low enough, OOM-killer makes your life miserable. [1] https://www.blog.google/products/pixel/pixel-visual-core-image-processing-and-machine-learning-pixel-2/ ====== Technical details Software tag-based KASAN mode is implemented in a very similar way to the generic one. This patchset essentially does the following: 1. TCR_TBI1 is set to enable Top Byte Ignore. 2. Shadow memory is used (with a different scale, 1:16, so each shadow byte corresponds to 16 bytes of kernel memory) to store memory tags. 3. All slab objects are aligned to shadow scale, which is 16 bytes. 4. All pointers returned from the slab allocator are tagged with a random tag and the corresponding shadow memory is poisoned with the same value. 5. Compiler instrumentation is used to insert tag checks. Either by calling callbacks or by inlining them (CONFIG_KASAN_OUTLINE and CONFIG_KASAN_INLINE flags are reused). 6. When a tag mismatch is detected in callback instrumentation mode KASAN simply prints a bug report. In case of inline instrumentation, clang inserts a brk instruction, and KASAN has it's own brk handler, which reports the bug. 7. The memory in between slab objects is marked with a reserved tag, and acts as a redzone. 8. When a slab object is freed it's marked with a reserved tag. Bug detection is imprecise for two reasons: 1. We won't catch some small out-of-bounds accesses, that fall into the same shadow cell, as the last byte of a slab object. 2. We only have 1 byte to store tags, which means we have a 1/256 probability of a tag match for an incorrect access (actually even slightly less due to reserved tag values). Despite that there's a particular type of bugs that tag-based KASAN can detect compared to generic KASAN: use-after-free after the object has been allocated by someone else. ====== Testing Some kernel developers voiced a concern that changing the top byte of kernel pointers may lead to subtle bugs that are difficult to discover. To address this concern deliberate testing has been performed. It doesn't seem feasible to do some kind of static checking to find potential issues with pointer tagging, so a dynamic approach was taken. All pointer comparisons/subtractions have been instrumented in an LLVM compiler pass and a kernel module that would print a bug report whenever two pointers with different tags are being compared/subtracted (ignoring comparisons with NULL pointers and with pointers obtained by casting an error code to a pointer type) has been used. Then the kernel has been booted in QEMU and on an Odroid C2 board and syzkaller has been run. This yielded the following results. The two places that look interesting are: is_vmalloc_addr in include/linux/mm.h is_kernel_rodata in mm/util.c Here we compare a pointer with some fixed untagged values to make sure that the pointer lies in a particular part of the kernel address space. Since tag-based KASAN doesn't add tags to pointers that belong to rodata or vmalloc regions, this should work as is. To make sure debug checks to those two functions that check that the result doesn't change whether we operate on pointers with or without untagging has been added. A few other cases that don't look that interesting: Comparing pointers to achieve unique sorting order of pointee objects (e.g. sorting locks addresses before performing a double lock): tty_ldisc_lock_pair_timeout in drivers/tty/tty_ldisc.c pipe_double_lock in fs/pipe.c unix_state_double_lock in net/unix/af_unix.c lock_two_nondirectories in fs/inode.c mutex_lock_double in kernel/events/core.c ep_cmp_ffd in fs/eventpoll.c fsnotify_compare_groups fs/notify/mark.c Nothing needs to be done here, since the tags embedded into pointers don't change, so the sorting order would still be unique. Checks that a pointer belongs to some particular allocation: is_sibling_entry in lib/radix-tree.c object_is_on_stack in include/linux/sched/task_stack.h Nothing needs to be done here either, since two pointers can only belong to the same allocation if they have the same tag. Overall, since the kernel boots and works, there are no critical bugs. As for the rest, the traditional kernel testing way (use until fails) is the only one that looks feasible. Another point here is that tag-based KASAN is available under a separate config option that needs to be deliberately enabled. Even though it might be used in a "near-production" environment to find bugs that are not found during fuzzing or running tests, it is still a debug tool. ====== Benchmarks The following numbers were collected on Odroid C2 board. Both generic and tag-based KASAN were used in inline instrumentation mode. Boot time [1]: * ~1.7 sec for clean kernel * ~5.0 sec for generic KASAN * ~5.0 sec for tag-based KASAN Network performance [2]: * 8.33 Gbits/sec for clean kernel * 3.17 Gbits/sec for generic KASAN * 2.85 Gbits/sec for tag-based KASAN Slab memory usage after boot [3]: * ~40 kb for clean kernel * ~105 kb (~260% overhead) for generic KASAN * ~47 kb (~20% overhead) for tag-based KASAN KASAN memory overhead consists of three main parts: 1. Increased slab memory usage due to redzones. 2. Shadow memory (the whole reserved once during boot). 3. Quaratine (grows gradually until some preset limit; the more the limit, the more the chance to detect a use-after-free). Comparing tag-based vs generic KASAN for each of these points: 1. 20% vs 260% overhead. 2. 1/16th vs 1/8th of physical memory. 3. Tag-based KASAN doesn't require quarantine. [1] Time before the ext4 driver is initialized. [2] Measured as `iperf -s & iperf -c 127.0.0.1 -t 30`. [3] Measured as `cat /proc/meminfo | grep Slab`. ====== Some notes A few notes: 1. The patchset can be found here: https://github.com/xairy/kasan-prototype/tree/khwasan 2. Building requires a recent Clang version (7.0.0 or later). 3. Stack instrumentation is not supported yet and will be added later. This patch (of 25): Tag-based KASAN changes the value of the top byte of pointers returned from the kernel allocation functions (such as kmalloc). This patch updates KASAN hooks signatures and their usage in SLAB and SLUB code to reflect that. Link: http://lkml.kernel.org/r/aec2b5e3973781ff8a6bb6760f8543643202c451.1544099024.git.andreyknvl@google.com Signed-off-by: Andrey Konovalov <andreyknvl@google.com> Reviewed-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Reviewed-by: Dmitry Vyukov <dvyukov@google.com> Cc: Christoph Lameter <cl@linux.com> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-12-28 16:29:37 +08:00
ret = kasan_kmalloc(s, ret, size, flags);
mm: slub: add kernel address sanitizer support for slub allocator With this patch kasan will be able to catch bugs in memory allocated by slub. Initially all objects in newly allocated slab page, marked as redzone. Later, when allocation of slub object happens, requested by caller number of bytes marked as accessible, and the rest of the object (including slub's metadata) marked as redzone (inaccessible). We also mark object as accessible if ksize was called for this object. There is some places in kernel where ksize function is called to inquire size of really allocated area. Such callers could validly access whole allocated memory, so it should be marked as accessible. Code in slub.c and slab_common.c files could validly access to object's metadata, so instrumentation for this files are disabled. Signed-off-by: Andrey Ryabinin <a.ryabinin@samsung.com> Signed-off-by: Dmitry Chernenkov <dmitryc@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Konstantin Serebryany <kcc@google.com> Signed-off-by: Andrey Konovalov <adech.fo@gmail.com> Cc: Yuri Gribov <tetra2005@gmail.com> Cc: Konstantin Khlebnikov <koct9i@gmail.com> Cc: Sasha Levin <sasha.levin@oracle.com> Cc: Christoph Lameter <cl@linux.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Ingo Molnar <mingo@elte.hu> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-14 06:39:42 +08:00
return ret;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
EXPORT_SYMBOL(__kmalloc_node);
#endif /* CONFIG_NUMA */
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
#ifdef CONFIG_HARDENED_USERCOPY
/*
* Rejects incorrectly sized objects and objects that are to be copied
* to/from userspace but do not fall entirely within the containing slab
* cache's usercopy region.
*
* Returns NULL if check passes, otherwise const char * to name of cache
* to indicate an error.
*/
void __check_heap_object(const void *ptr, unsigned long n, struct page *page,
bool to_user)
{
struct kmem_cache *s;
unsigned int offset;
size_t object_size;
mm, kfence: insert KFENCE hooks for SLUB Inserts KFENCE hooks into the SLUB allocator. To pass the originally requested size to KFENCE, add an argument 'orig_size' to slab_alloc*(). The additional argument is required to preserve the requested original size for kmalloc() allocations, which uses size classes (e.g. an allocation of 272 bytes will return an object of size 512). Therefore, kmem_cache::size does not represent the kmalloc-caller's requested size, and we must introduce the argument 'orig_size' to propagate the originally requested size to KFENCE. Without the originally requested size, we would not be able to detect out-of-bounds accesses for objects placed at the end of a KFENCE object page if that object is not equal to the kmalloc-size class it was bucketed into. When KFENCE is disabled, there is no additional overhead, since slab_alloc*() functions are __always_inline. Link: https://lkml.kernel.org/r/20201103175841.3495947-6-elver@google.com Signed-off-by: Marco Elver <elver@google.com> Signed-off-by: Alexander Potapenko <glider@google.com> Reviewed-by: Dmitry Vyukov <dvyukov@google.com> Reviewed-by: Jann Horn <jannh@google.com> Co-developed-by: Marco Elver <elver@google.com> Cc: Andrey Konovalov <andreyknvl@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christopher Lameter <cl@linux.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: David Rientjes <rientjes@google.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Hillf Danton <hdanton@sina.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Joern Engel <joern@purestorage.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kees Cook <keescook@chromium.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Paul E. McKenney <paulmck@kernel.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: SeongJae Park <sjpark@amazon.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-26 09:19:16 +08:00
bool is_kfence = is_kfence_address(ptr);
ptr = kasan_reset_tag(ptr);
/* Find object and usable object size. */
s = page->slab_cache;
/* Reject impossible pointers. */
if (ptr < page_address(page))
usercopy_abort("SLUB object not in SLUB page?!", NULL,
to_user, 0, n);
/* Find offset within object. */
mm, kfence: insert KFENCE hooks for SLUB Inserts KFENCE hooks into the SLUB allocator. To pass the originally requested size to KFENCE, add an argument 'orig_size' to slab_alloc*(). The additional argument is required to preserve the requested original size for kmalloc() allocations, which uses size classes (e.g. an allocation of 272 bytes will return an object of size 512). Therefore, kmem_cache::size does not represent the kmalloc-caller's requested size, and we must introduce the argument 'orig_size' to propagate the originally requested size to KFENCE. Without the originally requested size, we would not be able to detect out-of-bounds accesses for objects placed at the end of a KFENCE object page if that object is not equal to the kmalloc-size class it was bucketed into. When KFENCE is disabled, there is no additional overhead, since slab_alloc*() functions are __always_inline. Link: https://lkml.kernel.org/r/20201103175841.3495947-6-elver@google.com Signed-off-by: Marco Elver <elver@google.com> Signed-off-by: Alexander Potapenko <glider@google.com> Reviewed-by: Dmitry Vyukov <dvyukov@google.com> Reviewed-by: Jann Horn <jannh@google.com> Co-developed-by: Marco Elver <elver@google.com> Cc: Andrey Konovalov <andreyknvl@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christopher Lameter <cl@linux.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: David Rientjes <rientjes@google.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Hillf Danton <hdanton@sina.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Joern Engel <joern@purestorage.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kees Cook <keescook@chromium.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Paul E. McKenney <paulmck@kernel.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: SeongJae Park <sjpark@amazon.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-26 09:19:16 +08:00
if (is_kfence)
offset = ptr - kfence_object_start(ptr);
else
offset = (ptr - page_address(page)) % s->size;
/* Adjust for redzone and reject if within the redzone. */
mm, kfence: insert KFENCE hooks for SLUB Inserts KFENCE hooks into the SLUB allocator. To pass the originally requested size to KFENCE, add an argument 'orig_size' to slab_alloc*(). The additional argument is required to preserve the requested original size for kmalloc() allocations, which uses size classes (e.g. an allocation of 272 bytes will return an object of size 512). Therefore, kmem_cache::size does not represent the kmalloc-caller's requested size, and we must introduce the argument 'orig_size' to propagate the originally requested size to KFENCE. Without the originally requested size, we would not be able to detect out-of-bounds accesses for objects placed at the end of a KFENCE object page if that object is not equal to the kmalloc-size class it was bucketed into. When KFENCE is disabled, there is no additional overhead, since slab_alloc*() functions are __always_inline. Link: https://lkml.kernel.org/r/20201103175841.3495947-6-elver@google.com Signed-off-by: Marco Elver <elver@google.com> Signed-off-by: Alexander Potapenko <glider@google.com> Reviewed-by: Dmitry Vyukov <dvyukov@google.com> Reviewed-by: Jann Horn <jannh@google.com> Co-developed-by: Marco Elver <elver@google.com> Cc: Andrey Konovalov <andreyknvl@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christopher Lameter <cl@linux.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: David Rientjes <rientjes@google.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Hillf Danton <hdanton@sina.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Joern Engel <joern@purestorage.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kees Cook <keescook@chromium.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Paul E. McKenney <paulmck@kernel.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: SeongJae Park <sjpark@amazon.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-26 09:19:16 +08:00
if (!is_kfence && kmem_cache_debug_flags(s, SLAB_RED_ZONE)) {
if (offset < s->red_left_pad)
usercopy_abort("SLUB object in left red zone",
s->name, to_user, offset, n);
offset -= s->red_left_pad;
}
/* Allow address range falling entirely within usercopy region. */
if (offset >= s->useroffset &&
offset - s->useroffset <= s->usersize &&
n <= s->useroffset - offset + s->usersize)
return;
/*
* If the copy is still within the allocated object, produce
* a warning instead of rejecting the copy. This is intended
* to be a temporary method to find any missing usercopy
* whitelists.
*/
object_size = slab_ksize(s);
if (usercopy_fallback &&
offset <= object_size && n <= object_size - offset) {
usercopy_warn("SLUB object", s->name, to_user, offset, n);
return;
}
usercopy_abort("SLUB object", s->name, to_user, offset, n);
}
#endif /* CONFIG_HARDENED_USERCOPY */
size_t __ksize(const void *object)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
{
struct page *page;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (unlikely(object == ZERO_SIZE_PTR))
return 0;
page = virt_to_head_page(object);
if (unlikely(!PageSlab(page))) {
WARN_ON(!PageCompound(page));
return page_size(page);
}
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
return slab_ksize(page->slab_cache);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
EXPORT_SYMBOL(__ksize);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
void kfree(const void *x)
{
struct page *page;
void *object = (void *)x;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
trace_kfree(_RET_IP_, x);
if (unlikely(ZERO_OR_NULL_PTR(x)))
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
return;
page = virt_to_head_page(x);
SLUB: direct pass through of page size or higher kmalloc requests This gets rid of all kmalloc caches larger than page size. A kmalloc request larger than PAGE_SIZE > 2 is going to be passed through to the page allocator. This works both inline where we will call __get_free_pages instead of kmem_cache_alloc and in __kmalloc. kfree is modified to check if the object is in a slab page. If not then the page is freed via the page allocator instead. Roughly similar to what SLOB does. Advantages: - Reduces memory overhead for kmalloc array - Large kmalloc operations are faster since they do not need to pass through the slab allocator to get to the page allocator. - Performance increase of 10%-20% on alloc and 50% on free for PAGE_SIZEd allocations. SLUB must call page allocator for each alloc anyways since the higher order pages which that allowed avoiding the page alloc calls are not available in a reliable way anymore. So we are basically removing useless slab allocator overhead. - Large kmallocs yields page aligned object which is what SLAB did. Bad things like using page sized kmalloc allocations to stand in for page allocate allocs can be transparently handled and are not distinguishable from page allocator uses. - Checking for too large objects can be removed since it is done by the page allocator. Drawbacks: - No accounting for large kmalloc slab allocations anymore - No debugging of large kmalloc slab allocations. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 16:24:38 +08:00
if (unlikely(!PageSlab(page))) {
mm, sl[ou]b: improve memory accounting Patch series "guarantee natural alignment for kmalloc()", v2. This patch (of 2): SLOB currently doesn't account its pages at all, so in /proc/meminfo the Slab field shows zero. Modifying a counter on page allocation and freeing should be acceptable even for the small system scenarios SLOB is intended for. Since reclaimable caches are not separated in SLOB, account everything as unreclaimable. SLUB currently doesn't account kmalloc() and kmalloc_node() allocations larger than order-1 page, that are passed directly to the page allocator. As they also don't appear in /proc/slabinfo, it might look like a memory leak. For consistency, account them as well. (SLAB doesn't actually use page allocator directly, so no change there). Ideally SLOB and SLUB would be handled in separate patches, but due to the shared kmalloc_order() function and different kfree() implementations, it's easier to patch both at once to prevent inconsistencies. Link: http://lkml.kernel.org/r/20190826111627.7505-2-vbabka@suse.cz Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Ming Lei <ming.lei@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: "Darrick J . Wong" <darrick.wong@oracle.com> Cc: Christoph Hellwig <hch@lst.de> Cc: James Bottomley <James.Bottomley@HansenPartnership.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-10-07 08:58:42 +08:00
unsigned int order = compound_order(page);
BUG_ON(!PageCompound(page));
kfree_hook(object);
mod_lruvec_page_state(page, NR_SLAB_UNRECLAIMABLE_B,
-(PAGE_SIZE << order));
mm, sl[ou]b: improve memory accounting Patch series "guarantee natural alignment for kmalloc()", v2. This patch (of 2): SLOB currently doesn't account its pages at all, so in /proc/meminfo the Slab field shows zero. Modifying a counter on page allocation and freeing should be acceptable even for the small system scenarios SLOB is intended for. Since reclaimable caches are not separated in SLOB, account everything as unreclaimable. SLUB currently doesn't account kmalloc() and kmalloc_node() allocations larger than order-1 page, that are passed directly to the page allocator. As they also don't appear in /proc/slabinfo, it might look like a memory leak. For consistency, account them as well. (SLAB doesn't actually use page allocator directly, so no change there). Ideally SLOB and SLUB would be handled in separate patches, but due to the shared kmalloc_order() function and different kfree() implementations, it's easier to patch both at once to prevent inconsistencies. Link: http://lkml.kernel.org/r/20190826111627.7505-2-vbabka@suse.cz Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Ming Lei <ming.lei@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: "Darrick J . Wong" <darrick.wong@oracle.com> Cc: Christoph Hellwig <hch@lst.de> Cc: James Bottomley <James.Bottomley@HansenPartnership.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-10-07 08:58:42 +08:00
__free_pages(page, order);
SLUB: direct pass through of page size or higher kmalloc requests This gets rid of all kmalloc caches larger than page size. A kmalloc request larger than PAGE_SIZE > 2 is going to be passed through to the page allocator. This works both inline where we will call __get_free_pages instead of kmem_cache_alloc and in __kmalloc. kfree is modified to check if the object is in a slab page. If not then the page is freed via the page allocator instead. Roughly similar to what SLOB does. Advantages: - Reduces memory overhead for kmalloc array - Large kmalloc operations are faster since they do not need to pass through the slab allocator to get to the page allocator. - Performance increase of 10%-20% on alloc and 50% on free for PAGE_SIZEd allocations. SLUB must call page allocator for each alloc anyways since the higher order pages which that allowed avoiding the page alloc calls are not available in a reliable way anymore. So we are basically removing useless slab allocator overhead. - Large kmallocs yields page aligned object which is what SLAB did. Bad things like using page sized kmalloc allocations to stand in for page allocate allocs can be transparently handled and are not distinguishable from page allocator uses. - Checking for too large objects can be removed since it is done by the page allocator. Drawbacks: - No accounting for large kmalloc slab allocations anymore - No debugging of large kmalloc slab allocations. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 16:24:38 +08:00
return;
}
slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
EXPORT_SYMBOL(kfree);
slub: never fail to shrink cache SLUB's version of __kmem_cache_shrink() not only removes empty slabs, but also tries to rearrange the partial lists to place slabs filled up most to the head to cope with fragmentation. To achieve that, it allocates a temporary array of lists used to sort slabs by the number of objects in use. If the allocation fails, the whole procedure is aborted. This is unacceptable for the kernel memory accounting extension of the memory cgroup, where we want to make sure that kmem_cache_shrink() successfully discarded empty slabs. Although the allocation failure is utterly unlikely with the current page allocator implementation, which retries GFP_KERNEL allocations of order <= 2 infinitely, it is better not to rely on that. This patch therefore makes __kmem_cache_shrink() allocate the array on stack instead of calling kmalloc, which may fail. The array size is chosen to be equal to 32, because most SLUB caches store not more than 32 objects per slab page. Slab pages with <= 32 free objects are sorted using the array by the number of objects in use and promoted to the head of the partial list, while slab pages with > 32 free objects are left in the end of the list without any ordering imposed on them. Signed-off-by: Vladimir Davydov <vdavydov@parallels.com> Acked-by: Christoph Lameter <cl@linux.com> Acked-by: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Huang Ying <ying.huang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-13 06:59:41 +08:00
#define SHRINK_PROMOTE_MAX 32
/*
slub: never fail to shrink cache SLUB's version of __kmem_cache_shrink() not only removes empty slabs, but also tries to rearrange the partial lists to place slabs filled up most to the head to cope with fragmentation. To achieve that, it allocates a temporary array of lists used to sort slabs by the number of objects in use. If the allocation fails, the whole procedure is aborted. This is unacceptable for the kernel memory accounting extension of the memory cgroup, where we want to make sure that kmem_cache_shrink() successfully discarded empty slabs. Although the allocation failure is utterly unlikely with the current page allocator implementation, which retries GFP_KERNEL allocations of order <= 2 infinitely, it is better not to rely on that. This patch therefore makes __kmem_cache_shrink() allocate the array on stack instead of calling kmalloc, which may fail. The array size is chosen to be equal to 32, because most SLUB caches store not more than 32 objects per slab page. Slab pages with <= 32 free objects are sorted using the array by the number of objects in use and promoted to the head of the partial list, while slab pages with > 32 free objects are left in the end of the list without any ordering imposed on them. Signed-off-by: Vladimir Davydov <vdavydov@parallels.com> Acked-by: Christoph Lameter <cl@linux.com> Acked-by: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Huang Ying <ying.huang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-13 06:59:41 +08:00
* kmem_cache_shrink discards empty slabs and promotes the slabs filled
* up most to the head of the partial lists. New allocations will then
* fill those up and thus they can be removed from the partial lists.
*
* The slabs with the least items are placed last. This results in them
* being allocated from last increasing the chance that the last objects
* are freed in them.
*/
int __kmem_cache_shrink(struct kmem_cache *s)
{
int node;
int i;
struct kmem_cache_node *n;
struct page *page;
struct page *t;
slub: never fail to shrink cache SLUB's version of __kmem_cache_shrink() not only removes empty slabs, but also tries to rearrange the partial lists to place slabs filled up most to the head to cope with fragmentation. To achieve that, it allocates a temporary array of lists used to sort slabs by the number of objects in use. If the allocation fails, the whole procedure is aborted. This is unacceptable for the kernel memory accounting extension of the memory cgroup, where we want to make sure that kmem_cache_shrink() successfully discarded empty slabs. Although the allocation failure is utterly unlikely with the current page allocator implementation, which retries GFP_KERNEL allocations of order <= 2 infinitely, it is better not to rely on that. This patch therefore makes __kmem_cache_shrink() allocate the array on stack instead of calling kmalloc, which may fail. The array size is chosen to be equal to 32, because most SLUB caches store not more than 32 objects per slab page. Slab pages with <= 32 free objects are sorted using the array by the number of objects in use and promoted to the head of the partial list, while slab pages with > 32 free objects are left in the end of the list without any ordering imposed on them. Signed-off-by: Vladimir Davydov <vdavydov@parallels.com> Acked-by: Christoph Lameter <cl@linux.com> Acked-by: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Huang Ying <ying.huang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-13 06:59:41 +08:00
struct list_head discard;
struct list_head promote[SHRINK_PROMOTE_MAX];
unsigned long flags;
int ret = 0;
flush_all(s);
for_each_kmem_cache_node(s, node, n) {
slub: never fail to shrink cache SLUB's version of __kmem_cache_shrink() not only removes empty slabs, but also tries to rearrange the partial lists to place slabs filled up most to the head to cope with fragmentation. To achieve that, it allocates a temporary array of lists used to sort slabs by the number of objects in use. If the allocation fails, the whole procedure is aborted. This is unacceptable for the kernel memory accounting extension of the memory cgroup, where we want to make sure that kmem_cache_shrink() successfully discarded empty slabs. Although the allocation failure is utterly unlikely with the current page allocator implementation, which retries GFP_KERNEL allocations of order <= 2 infinitely, it is better not to rely on that. This patch therefore makes __kmem_cache_shrink() allocate the array on stack instead of calling kmalloc, which may fail. The array size is chosen to be equal to 32, because most SLUB caches store not more than 32 objects per slab page. Slab pages with <= 32 free objects are sorted using the array by the number of objects in use and promoted to the head of the partial list, while slab pages with > 32 free objects are left in the end of the list without any ordering imposed on them. Signed-off-by: Vladimir Davydov <vdavydov@parallels.com> Acked-by: Christoph Lameter <cl@linux.com> Acked-by: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Huang Ying <ying.huang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-13 06:59:41 +08:00
INIT_LIST_HEAD(&discard);
for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
INIT_LIST_HEAD(promote + i);
spin_lock_irqsave(&n->list_lock, flags);
/*
slub: never fail to shrink cache SLUB's version of __kmem_cache_shrink() not only removes empty slabs, but also tries to rearrange the partial lists to place slabs filled up most to the head to cope with fragmentation. To achieve that, it allocates a temporary array of lists used to sort slabs by the number of objects in use. If the allocation fails, the whole procedure is aborted. This is unacceptable for the kernel memory accounting extension of the memory cgroup, where we want to make sure that kmem_cache_shrink() successfully discarded empty slabs. Although the allocation failure is utterly unlikely with the current page allocator implementation, which retries GFP_KERNEL allocations of order <= 2 infinitely, it is better not to rely on that. This patch therefore makes __kmem_cache_shrink() allocate the array on stack instead of calling kmalloc, which may fail. The array size is chosen to be equal to 32, because most SLUB caches store not more than 32 objects per slab page. Slab pages with <= 32 free objects are sorted using the array by the number of objects in use and promoted to the head of the partial list, while slab pages with > 32 free objects are left in the end of the list without any ordering imposed on them. Signed-off-by: Vladimir Davydov <vdavydov@parallels.com> Acked-by: Christoph Lameter <cl@linux.com> Acked-by: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Huang Ying <ying.huang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-13 06:59:41 +08:00
* Build lists of slabs to discard or promote.
*
* Note that concurrent frees may occur while we hold the
* list_lock. page->inuse here is the upper limit.
*/
list_for_each_entry_safe(page, t, &n->partial, slab_list) {
slub: never fail to shrink cache SLUB's version of __kmem_cache_shrink() not only removes empty slabs, but also tries to rearrange the partial lists to place slabs filled up most to the head to cope with fragmentation. To achieve that, it allocates a temporary array of lists used to sort slabs by the number of objects in use. If the allocation fails, the whole procedure is aborted. This is unacceptable for the kernel memory accounting extension of the memory cgroup, where we want to make sure that kmem_cache_shrink() successfully discarded empty slabs. Although the allocation failure is utterly unlikely with the current page allocator implementation, which retries GFP_KERNEL allocations of order <= 2 infinitely, it is better not to rely on that. This patch therefore makes __kmem_cache_shrink() allocate the array on stack instead of calling kmalloc, which may fail. The array size is chosen to be equal to 32, because most SLUB caches store not more than 32 objects per slab page. Slab pages with <= 32 free objects are sorted using the array by the number of objects in use and promoted to the head of the partial list, while slab pages with > 32 free objects are left in the end of the list without any ordering imposed on them. Signed-off-by: Vladimir Davydov <vdavydov@parallels.com> Acked-by: Christoph Lameter <cl@linux.com> Acked-by: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Huang Ying <ying.huang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-13 06:59:41 +08:00
int free = page->objects - page->inuse;
/* Do not reread page->inuse */
barrier();
/* We do not keep full slabs on the list */
BUG_ON(free <= 0);
if (free == page->objects) {
list_move(&page->slab_list, &discard);
n->nr_partial--;
slub: never fail to shrink cache SLUB's version of __kmem_cache_shrink() not only removes empty slabs, but also tries to rearrange the partial lists to place slabs filled up most to the head to cope with fragmentation. To achieve that, it allocates a temporary array of lists used to sort slabs by the number of objects in use. If the allocation fails, the whole procedure is aborted. This is unacceptable for the kernel memory accounting extension of the memory cgroup, where we want to make sure that kmem_cache_shrink() successfully discarded empty slabs. Although the allocation failure is utterly unlikely with the current page allocator implementation, which retries GFP_KERNEL allocations of order <= 2 infinitely, it is better not to rely on that. This patch therefore makes __kmem_cache_shrink() allocate the array on stack instead of calling kmalloc, which may fail. The array size is chosen to be equal to 32, because most SLUB caches store not more than 32 objects per slab page. Slab pages with <= 32 free objects are sorted using the array by the number of objects in use and promoted to the head of the partial list, while slab pages with > 32 free objects are left in the end of the list without any ordering imposed on them. Signed-off-by: Vladimir Davydov <vdavydov@parallels.com> Acked-by: Christoph Lameter <cl@linux.com> Acked-by: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Huang Ying <ying.huang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-13 06:59:41 +08:00
} else if (free <= SHRINK_PROMOTE_MAX)
list_move(&page->slab_list, promote + free - 1);
}
/*
slub: never fail to shrink cache SLUB's version of __kmem_cache_shrink() not only removes empty slabs, but also tries to rearrange the partial lists to place slabs filled up most to the head to cope with fragmentation. To achieve that, it allocates a temporary array of lists used to sort slabs by the number of objects in use. If the allocation fails, the whole procedure is aborted. This is unacceptable for the kernel memory accounting extension of the memory cgroup, where we want to make sure that kmem_cache_shrink() successfully discarded empty slabs. Although the allocation failure is utterly unlikely with the current page allocator implementation, which retries GFP_KERNEL allocations of order <= 2 infinitely, it is better not to rely on that. This patch therefore makes __kmem_cache_shrink() allocate the array on stack instead of calling kmalloc, which may fail. The array size is chosen to be equal to 32, because most SLUB caches store not more than 32 objects per slab page. Slab pages with <= 32 free objects are sorted using the array by the number of objects in use and promoted to the head of the partial list, while slab pages with > 32 free objects are left in the end of the list without any ordering imposed on them. Signed-off-by: Vladimir Davydov <vdavydov@parallels.com> Acked-by: Christoph Lameter <cl@linux.com> Acked-by: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Huang Ying <ying.huang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-13 06:59:41 +08:00
* Promote the slabs filled up most to the head of the
* partial list.
*/
slub: never fail to shrink cache SLUB's version of __kmem_cache_shrink() not only removes empty slabs, but also tries to rearrange the partial lists to place slabs filled up most to the head to cope with fragmentation. To achieve that, it allocates a temporary array of lists used to sort slabs by the number of objects in use. If the allocation fails, the whole procedure is aborted. This is unacceptable for the kernel memory accounting extension of the memory cgroup, where we want to make sure that kmem_cache_shrink() successfully discarded empty slabs. Although the allocation failure is utterly unlikely with the current page allocator implementation, which retries GFP_KERNEL allocations of order <= 2 infinitely, it is better not to rely on that. This patch therefore makes __kmem_cache_shrink() allocate the array on stack instead of calling kmalloc, which may fail. The array size is chosen to be equal to 32, because most SLUB caches store not more than 32 objects per slab page. Slab pages with <= 32 free objects are sorted using the array by the number of objects in use and promoted to the head of the partial list, while slab pages with > 32 free objects are left in the end of the list without any ordering imposed on them. Signed-off-by: Vladimir Davydov <vdavydov@parallels.com> Acked-by: Christoph Lameter <cl@linux.com> Acked-by: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Huang Ying <ying.huang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-13 06:59:41 +08:00
for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
list_splice(promote + i, &n->partial);
spin_unlock_irqrestore(&n->list_lock, flags);
/* Release empty slabs */
list_for_each_entry_safe(page, t, &discard, slab_list)
discard_slab(s, page);
if (slabs_node(s, node))
ret = 1;
}
return ret;
}
static int slab_mem_going_offline_callback(void *arg)
{
struct kmem_cache *s;
mutex_lock(&slab_mutex);
list_for_each_entry(s, &slab_caches, list)
__kmem_cache_shrink(s);
mutex_unlock(&slab_mutex);
return 0;
}
static void slab_mem_offline_callback(void *arg)
{
struct memory_notify *marg = arg;
int offline_node;
slub, hotplug: ignore unrelated node's hot-adding and hot-removing SLUB only focuses on the nodes which have normal memory and it ignores the other node's hot-adding and hot-removing. Aka: if some memory of a node which has no onlined memory is online, but this new memory onlined is not normal memory (for example, highmem), we should not allocate kmem_cache_node for SLUB. And if the last normal memory is offlined, but the node still has memory, we should remove kmem_cache_node for that node. (The current code delays it when all of the memory is offlined) So we only do something when marg->status_change_nid_normal > 0. marg->status_change_nid is not suitable here. The same problem doesn't exist in SLAB, because SLAB allocates kmem_list3 for every node even the node don't have normal memory, SLAB tolerates kmem_list3 on alien nodes. SLUB only focuses on the nodes which have normal memory, it don't tolerate alien kmem_cache_node. The patch makes SLUB become self-compatible and avoids WARNs and BUGs in rare conditions. Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: David Rientjes <rientjes@google.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com> Cc: Rob Landley <rob@landley.net> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Jiang Liu <jiang.liu@huawei.com> Cc: Kay Sievers <kay.sievers@vrfy.org> Cc: Greg Kroah-Hartman <gregkh@suse.de> Cc: Mel Gorman <mgorman@suse.de> Cc: Wen Congyang <wency@cn.fujitsu.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-12-12 08:01:05 +08:00
offline_node = marg->status_change_nid_normal;
/*
* If the node still has available memory. we need kmem_cache_node
* for it yet.
*/
if (offline_node < 0)
return;
mutex_lock(&slab_mutex);
mm, slab, slub: stop taking memory hotplug lock Since commit 03afc0e25f7f ("slab: get_online_mems for kmem_cache_{create,destroy,shrink}") we are taking memory hotplug lock for SLAB and SLUB when creating, destroying or shrinking a cache. It is quite a heavy lock and it's best to avoid it if possible, as we had several issues with lockdep complaining about ordering in the past, see e.g. e4f8e513c3d3 ("mm/slub: fix a deadlock in show_slab_objects()"). The problem scenario in 03afc0e25f7f (solved by the memory hotplug lock) can be summarized as follows: while there's slab_mutex synchronizing new kmem cache creation and SLUB's MEM_GOING_ONLINE callback slab_mem_going_online_callback(), we may miss creation of kmem_cache_node for the hotplugged node in the new kmem cache, because the hotplug callback doesn't yet see the new cache, and cache creation in init_kmem_cache_nodes() only inits kmem_cache_node for nodes in the N_NORMAL_MEMORY nodemask, which however may not yet include the new node, as that happens only later after the MEM_GOING_ONLINE callback. Instead of using get/put_online_mems(), the problem can be solved by SLUB maintaining its own nodemask of nodes for which it has allocated the per-node kmem_cache_node structures. This nodemask would generally mirror the N_NORMAL_MEMORY nodemask, but would be updated only in under SLUB's control in its memory hotplug callbacks under the slab_mutex. This patch adds such nodemask and its handling. Commit 03afc0e25f7f mentiones "issues like [the one above]", but there don't appear to be further issues. All the paths (shared for SLAB and SLUB) taking the memory hotplug locks are also taking the slab_mutex, except kmem_cache_shrink() where 03afc0e25f7f replaced slab_mutex with get/put_online_mems(). We however cannot simply restore slab_mutex in kmem_cache_shrink(), as SLUB can enters the function from a write to sysfs 'shrink' file, thus holding kernfs lock, and in kmem_cache_create() the kernfs lock is nested within slab_mutex. But on closer inspection we don't actually need to protect kmem_cache_shrink() from hotplug callbacks: While SLUB's __kmem_cache_shrink() does for_each_kmem_cache_node(), missing a new node added in parallel hotplug is not fatal, and parallel hotremove does not free kmem_cache_node's anymore after the previous patch, so use-after free cannot happen. The per-node shrinking itself is protected by n->list_lock. Same is true for SLAB, and SLOB is no-op. SLAB also doesn't need the memory hotplug locking, which it only gained by 03afc0e25f7f through the shared paths in slab_common.c. Its memory hotplug callbacks are also protected by slab_mutex against races with these paths. The problem of SLUB relying on N_NORMAL_MEMORY doesn't apply to SLAB, as its setup_kmem_cache_nodes relies on N_ONLINE, and the new node is already set there during the MEM_GOING_ONLINE callback, so no special care is needed for SLAB. As such, this patch removes all get/put_online_mems() usage by the slab subsystem. Link: https://lkml.kernel.org/r/20210113131634.3671-3-vbabka@suse.cz Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Cc: Christoph Lameter <cl@linux.com> Cc: David Hildenbrand <david@redhat.com> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: Qian Cai <cai@redhat.com> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-25 04:01:12 +08:00
node_clear(offline_node, slab_nodes);
mm, slub: stop freeing kmem_cache_node structures on node offline Patch series "mm, slab, slub: remove cpu and memory hotplug locks". Some related work caused me to look at how we use get/put_mems_online() and get/put_online_cpus() during kmem cache creation/descruction/shrinking, and realize that it should be actually safe to remove all of that with rather small effort (as e.g. Michal Hocko suspected in some of the past discussions already). This has the benefit to avoid rather heavy locks that have caused locking order issues already in the past. So this is the result, Patches 2 and 3 remove memory hotplug and cpu hotplug locking, respectively. Patch 1 is due to realization that in fact some races exist despite the locks (even if not removed), but the most sane solution is not to introduce more of them, but rather accept some wasted memory in scenarios that should be rare anyway (full memory hot remove), as we do the same in other contexts already. This patch (of 3): Commit e4f8e513c3d3 ("mm/slub: fix a deadlock in show_slab_objects()") has fixed a problematic locking order by removing the memory hotplug lock get/put_online_mems() from show_slab_objects(). During the discussion, it was argued [1] that this is OK, because existing slabs on the node would prevent a hotremove to proceed. That's true, but per-node kmem_cache_node structures are not necessarily allocated on the same node and may exist even without actual slab pages on the same node. Any path that uses get_node() directly or via for_each_kmem_cache_node() (such as show_slab_objects()) can race with freeing of kmem_cache_node even with the !NULL check, resulting in use-after-free. To that end, commit e4f8e513c3d3 argues in a comment that: * We don't really need mem_hotplug_lock (to hold off * slab_mem_going_offline_callback) here because slab's memory hot * unplug code doesn't destroy the kmem_cache->node[] data. While it's true that slab_mem_going_offline_callback() doesn't free the kmem_cache_node, the later callback slab_mem_offline_callback() actually does, so the race and use-after-free exists. Not just for show_slab_objects() after commit e4f8e513c3d3, but also many other places that are not under slab_mutex. And adding slab_mutex locking or other synchronization to SLUB paths such as get_any_partial() would be bad for performance and error-prone. The easiest solution is therefore to make the abovementioned comment true and stop freeing the kmem_cache_node structures, accepting some wasted memory in the full memory node removal scenario. Analogically we also don't free hotremoved pgdat as mentioned in [1], nor the similar per-node structures in SLAB. Importantly this approach will not block the hotremove, as generally such nodes should be movable in order to succeed hotremove in the first place, and thus the GFP_KERNEL allocated kmem_cache_node will come from elsewhere. [1] https://lore.kernel.org/linux-mm/20190924151147.GB23050@dhcp22.suse.cz/ Link: https://lkml.kernel.org/r/20210113131634.3671-1-vbabka@suse.cz Link: https://lkml.kernel.org/r/20210113131634.3671-2-vbabka@suse.cz Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Cc: Qian Cai <cai@redhat.com> Cc: David Hildenbrand <david@redhat.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Vlastimil Babka <vbabka@suse.cz> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-25 04:01:08 +08:00
/*
* We no longer free kmem_cache_node structures here, as it would be
* racy with all get_node() users, and infeasible to protect them with
* slab_mutex.
*/
mutex_unlock(&slab_mutex);
}
static int slab_mem_going_online_callback(void *arg)
{
struct kmem_cache_node *n;
struct kmem_cache *s;
struct memory_notify *marg = arg;
slub, hotplug: ignore unrelated node's hot-adding and hot-removing SLUB only focuses on the nodes which have normal memory and it ignores the other node's hot-adding and hot-removing. Aka: if some memory of a node which has no onlined memory is online, but this new memory onlined is not normal memory (for example, highmem), we should not allocate kmem_cache_node for SLUB. And if the last normal memory is offlined, but the node still has memory, we should remove kmem_cache_node for that node. (The current code delays it when all of the memory is offlined) So we only do something when marg->status_change_nid_normal > 0. marg->status_change_nid is not suitable here. The same problem doesn't exist in SLAB, because SLAB allocates kmem_list3 for every node even the node don't have normal memory, SLAB tolerates kmem_list3 on alien nodes. SLUB only focuses on the nodes which have normal memory, it don't tolerate alien kmem_cache_node. The patch makes SLUB become self-compatible and avoids WARNs and BUGs in rare conditions. Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: David Rientjes <rientjes@google.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com> Cc: Rob Landley <rob@landley.net> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Jiang Liu <jiang.liu@huawei.com> Cc: Kay Sievers <kay.sievers@vrfy.org> Cc: Greg Kroah-Hartman <gregkh@suse.de> Cc: Mel Gorman <mgorman@suse.de> Cc: Wen Congyang <wency@cn.fujitsu.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-12-12 08:01:05 +08:00
int nid = marg->status_change_nid_normal;
int ret = 0;
/*
* If the node's memory is already available, then kmem_cache_node is
* already created. Nothing to do.
*/
if (nid < 0)
return 0;
/*
* We are bringing a node online. No memory is available yet. We must
* allocate a kmem_cache_node structure in order to bring the node
* online.
*/
mutex_lock(&slab_mutex);
list_for_each_entry(s, &slab_caches, list) {
mm, slub: stop freeing kmem_cache_node structures on node offline Patch series "mm, slab, slub: remove cpu and memory hotplug locks". Some related work caused me to look at how we use get/put_mems_online() and get/put_online_cpus() during kmem cache creation/descruction/shrinking, and realize that it should be actually safe to remove all of that with rather small effort (as e.g. Michal Hocko suspected in some of the past discussions already). This has the benefit to avoid rather heavy locks that have caused locking order issues already in the past. So this is the result, Patches 2 and 3 remove memory hotplug and cpu hotplug locking, respectively. Patch 1 is due to realization that in fact some races exist despite the locks (even if not removed), but the most sane solution is not to introduce more of them, but rather accept some wasted memory in scenarios that should be rare anyway (full memory hot remove), as we do the same in other contexts already. This patch (of 3): Commit e4f8e513c3d3 ("mm/slub: fix a deadlock in show_slab_objects()") has fixed a problematic locking order by removing the memory hotplug lock get/put_online_mems() from show_slab_objects(). During the discussion, it was argued [1] that this is OK, because existing slabs on the node would prevent a hotremove to proceed. That's true, but per-node kmem_cache_node structures are not necessarily allocated on the same node and may exist even without actual slab pages on the same node. Any path that uses get_node() directly or via for_each_kmem_cache_node() (such as show_slab_objects()) can race with freeing of kmem_cache_node even with the !NULL check, resulting in use-after-free. To that end, commit e4f8e513c3d3 argues in a comment that: * We don't really need mem_hotplug_lock (to hold off * slab_mem_going_offline_callback) here because slab's memory hot * unplug code doesn't destroy the kmem_cache->node[] data. While it's true that slab_mem_going_offline_callback() doesn't free the kmem_cache_node, the later callback slab_mem_offline_callback() actually does, so the race and use-after-free exists. Not just for show_slab_objects() after commit e4f8e513c3d3, but also many other places that are not under slab_mutex. And adding slab_mutex locking or other synchronization to SLUB paths such as get_any_partial() would be bad for performance and error-prone. The easiest solution is therefore to make the abovementioned comment true and stop freeing the kmem_cache_node structures, accepting some wasted memory in the full memory node removal scenario. Analogically we also don't free hotremoved pgdat as mentioned in [1], nor the similar per-node structures in SLAB. Importantly this approach will not block the hotremove, as generally such nodes should be movable in order to succeed hotremove in the first place, and thus the GFP_KERNEL allocated kmem_cache_node will come from elsewhere. [1] https://lore.kernel.org/linux-mm/20190924151147.GB23050@dhcp22.suse.cz/ Link: https://lkml.kernel.org/r/20210113131634.3671-1-vbabka@suse.cz Link: https://lkml.kernel.org/r/20210113131634.3671-2-vbabka@suse.cz Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Cc: Qian Cai <cai@redhat.com> Cc: David Hildenbrand <david@redhat.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Vlastimil Babka <vbabka@suse.cz> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-25 04:01:08 +08:00
/*
* The structure may already exist if the node was previously
* onlined and offlined.
*/
if (get_node(s, nid))
continue;
/*
* XXX: kmem_cache_alloc_node will fallback to other nodes
* since memory is not yet available from the node that
* is brought up.
*/
n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
if (!n) {
ret = -ENOMEM;
goto out;
}
init_kmem_cache_node(n);
s->node[nid] = n;
}
mm, slab, slub: stop taking memory hotplug lock Since commit 03afc0e25f7f ("slab: get_online_mems for kmem_cache_{create,destroy,shrink}") we are taking memory hotplug lock for SLAB and SLUB when creating, destroying or shrinking a cache. It is quite a heavy lock and it's best to avoid it if possible, as we had several issues with lockdep complaining about ordering in the past, see e.g. e4f8e513c3d3 ("mm/slub: fix a deadlock in show_slab_objects()"). The problem scenario in 03afc0e25f7f (solved by the memory hotplug lock) can be summarized as follows: while there's slab_mutex synchronizing new kmem cache creation and SLUB's MEM_GOING_ONLINE callback slab_mem_going_online_callback(), we may miss creation of kmem_cache_node for the hotplugged node in the new kmem cache, because the hotplug callback doesn't yet see the new cache, and cache creation in init_kmem_cache_nodes() only inits kmem_cache_node for nodes in the N_NORMAL_MEMORY nodemask, which however may not yet include the new node, as that happens only later after the MEM_GOING_ONLINE callback. Instead of using get/put_online_mems(), the problem can be solved by SLUB maintaining its own nodemask of nodes for which it has allocated the per-node kmem_cache_node structures. This nodemask would generally mirror the N_NORMAL_MEMORY nodemask, but would be updated only in under SLUB's control in its memory hotplug callbacks under the slab_mutex. This patch adds such nodemask and its handling. Commit 03afc0e25f7f mentiones "issues like [the one above]", but there don't appear to be further issues. All the paths (shared for SLAB and SLUB) taking the memory hotplug locks are also taking the slab_mutex, except kmem_cache_shrink() where 03afc0e25f7f replaced slab_mutex with get/put_online_mems(). We however cannot simply restore slab_mutex in kmem_cache_shrink(), as SLUB can enters the function from a write to sysfs 'shrink' file, thus holding kernfs lock, and in kmem_cache_create() the kernfs lock is nested within slab_mutex. But on closer inspection we don't actually need to protect kmem_cache_shrink() from hotplug callbacks: While SLUB's __kmem_cache_shrink() does for_each_kmem_cache_node(), missing a new node added in parallel hotplug is not fatal, and parallel hotremove does not free kmem_cache_node's anymore after the previous patch, so use-after free cannot happen. The per-node shrinking itself is protected by n->list_lock. Same is true for SLAB, and SLOB is no-op. SLAB also doesn't need the memory hotplug locking, which it only gained by 03afc0e25f7f through the shared paths in slab_common.c. Its memory hotplug callbacks are also protected by slab_mutex against races with these paths. The problem of SLUB relying on N_NORMAL_MEMORY doesn't apply to SLAB, as its setup_kmem_cache_nodes relies on N_ONLINE, and the new node is already set there during the MEM_GOING_ONLINE callback, so no special care is needed for SLAB. As such, this patch removes all get/put_online_mems() usage by the slab subsystem. Link: https://lkml.kernel.org/r/20210113131634.3671-3-vbabka@suse.cz Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Cc: Christoph Lameter <cl@linux.com> Cc: David Hildenbrand <david@redhat.com> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: Qian Cai <cai@redhat.com> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-25 04:01:12 +08:00
/*
* Any cache created after this point will also have kmem_cache_node
* initialized for the new node.
*/
node_set(nid, slab_nodes);
out:
mutex_unlock(&slab_mutex);
return ret;
}
static int slab_memory_callback(struct notifier_block *self,
unsigned long action, void *arg)
{
int ret = 0;
switch (action) {
case MEM_GOING_ONLINE:
ret = slab_mem_going_online_callback(arg);
break;
case MEM_GOING_OFFLINE:
ret = slab_mem_going_offline_callback(arg);
break;
case MEM_OFFLINE:
case MEM_CANCEL_ONLINE:
slab_mem_offline_callback(arg);
break;
case MEM_ONLINE:
case MEM_CANCEL_OFFLINE:
break;
}
memcg: memory hotplug fix for notifier callback Fixes for memcg/memory hotplug. While memory hotplug allocate/free memmap, page_cgroup doesn't free page_cgroup at OFFLINE when page_cgroup is allocated via bootomem. (Because freeing bootmem requires special care.) Then, if page_cgroup is allocated by bootmem and memmap is freed/allocated by memory hotplug, page_cgroup->page == page is no longer true. But current MEM_ONLINE handler doesn't check it and update page_cgroup->page if it's not necessary to allocate page_cgroup. (This was not found because memmap is not freed if SPARSEMEM_VMEMMAP is y.) And I noticed that MEM_ONLINE can be called against "part of section". So, freeing page_cgroup at CANCEL_ONLINE will cause trouble. (freeing used page_cgroup) Don't rollback at CANCEL. One more, current memory hotplug notifier is stopped by slub because it sets NOTIFY_STOP_MASK to return vaule. So, page_cgroup's callback never be called. (low priority than slub now.) I think this slub's behavior is not intentional(BUG). and fixes it. Another way to be considered about page_cgroup allocation: - free page_cgroup at OFFLINE even if it's from bootmem and remove specieal handler. But it requires more changes. Addresses http://bugzilla.kernel.org/show_bug.cgi?id=12041 Signed-off-by: KAMEZAWA Hiruyoki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Li Zefan <lizf@cn.fujitsu.com> Cc: Balbir Singh <balbir@in.ibm.com> Cc: Pavel Emelyanov <xemul@openvz.org> Tested-by: Badari Pulavarty <pbadari@us.ibm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-12-02 05:13:48 +08:00
if (ret)
ret = notifier_from_errno(ret);
else
ret = NOTIFY_OK;
return ret;
}
static struct notifier_block slab_memory_callback_nb = {
.notifier_call = slab_memory_callback,
.priority = SLAB_CALLBACK_PRI,
};
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/********************************************************************
* Basic setup of slabs
*******************************************************************/
/*
* Used for early kmem_cache structures that were allocated using
* the page allocator. Allocate them properly then fix up the pointers
* that may be pointing to the wrong kmem_cache structure.
*/
static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
{
int node;
struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
struct kmem_cache_node *n;
memcpy(s, static_cache, kmem_cache->object_size);
/*
* This runs very early, and only the boot processor is supposed to be
* up. Even if it weren't true, IRQs are not up so we couldn't fire
* IPIs around.
*/
__flush_cpu_slab(s, smp_processor_id());
for_each_kmem_cache_node(s, node, n) {
struct page *p;
list_for_each_entry(p, &n->partial, slab_list)
p->slab_cache = s;
#ifdef CONFIG_SLUB_DEBUG
list_for_each_entry(p, &n->full, slab_list)
p->slab_cache = s;
#endif
}
list_add(&s->list, &slab_caches);
return s;
}
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
void __init kmem_cache_init(void)
{
static __initdata struct kmem_cache boot_kmem_cache,
boot_kmem_cache_node;
mm, slab, slub: stop taking memory hotplug lock Since commit 03afc0e25f7f ("slab: get_online_mems for kmem_cache_{create,destroy,shrink}") we are taking memory hotplug lock for SLAB and SLUB when creating, destroying or shrinking a cache. It is quite a heavy lock and it's best to avoid it if possible, as we had several issues with lockdep complaining about ordering in the past, see e.g. e4f8e513c3d3 ("mm/slub: fix a deadlock in show_slab_objects()"). The problem scenario in 03afc0e25f7f (solved by the memory hotplug lock) can be summarized as follows: while there's slab_mutex synchronizing new kmem cache creation and SLUB's MEM_GOING_ONLINE callback slab_mem_going_online_callback(), we may miss creation of kmem_cache_node for the hotplugged node in the new kmem cache, because the hotplug callback doesn't yet see the new cache, and cache creation in init_kmem_cache_nodes() only inits kmem_cache_node for nodes in the N_NORMAL_MEMORY nodemask, which however may not yet include the new node, as that happens only later after the MEM_GOING_ONLINE callback. Instead of using get/put_online_mems(), the problem can be solved by SLUB maintaining its own nodemask of nodes for which it has allocated the per-node kmem_cache_node structures. This nodemask would generally mirror the N_NORMAL_MEMORY nodemask, but would be updated only in under SLUB's control in its memory hotplug callbacks under the slab_mutex. This patch adds such nodemask and its handling. Commit 03afc0e25f7f mentiones "issues like [the one above]", but there don't appear to be further issues. All the paths (shared for SLAB and SLUB) taking the memory hotplug locks are also taking the slab_mutex, except kmem_cache_shrink() where 03afc0e25f7f replaced slab_mutex with get/put_online_mems(). We however cannot simply restore slab_mutex in kmem_cache_shrink(), as SLUB can enters the function from a write to sysfs 'shrink' file, thus holding kernfs lock, and in kmem_cache_create() the kernfs lock is nested within slab_mutex. But on closer inspection we don't actually need to protect kmem_cache_shrink() from hotplug callbacks: While SLUB's __kmem_cache_shrink() does for_each_kmem_cache_node(), missing a new node added in parallel hotplug is not fatal, and parallel hotremove does not free kmem_cache_node's anymore after the previous patch, so use-after free cannot happen. The per-node shrinking itself is protected by n->list_lock. Same is true for SLAB, and SLOB is no-op. SLAB also doesn't need the memory hotplug locking, which it only gained by 03afc0e25f7f through the shared paths in slab_common.c. Its memory hotplug callbacks are also protected by slab_mutex against races with these paths. The problem of SLUB relying on N_NORMAL_MEMORY doesn't apply to SLAB, as its setup_kmem_cache_nodes relies on N_ONLINE, and the new node is already set there during the MEM_GOING_ONLINE callback, so no special care is needed for SLAB. As such, this patch removes all get/put_online_mems() usage by the slab subsystem. Link: https://lkml.kernel.org/r/20210113131634.3671-3-vbabka@suse.cz Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Cc: Christoph Lameter <cl@linux.com> Cc: David Hildenbrand <david@redhat.com> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: Qian Cai <cai@redhat.com> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-25 04:01:12 +08:00
int node;
if (debug_guardpage_minorder())
slub_max_order = 0;
kmem_cache_node = &boot_kmem_cache_node;
kmem_cache = &boot_kmem_cache;
mm, slab, slub: stop taking memory hotplug lock Since commit 03afc0e25f7f ("slab: get_online_mems for kmem_cache_{create,destroy,shrink}") we are taking memory hotplug lock for SLAB and SLUB when creating, destroying or shrinking a cache. It is quite a heavy lock and it's best to avoid it if possible, as we had several issues with lockdep complaining about ordering in the past, see e.g. e4f8e513c3d3 ("mm/slub: fix a deadlock in show_slab_objects()"). The problem scenario in 03afc0e25f7f (solved by the memory hotplug lock) can be summarized as follows: while there's slab_mutex synchronizing new kmem cache creation and SLUB's MEM_GOING_ONLINE callback slab_mem_going_online_callback(), we may miss creation of kmem_cache_node for the hotplugged node in the new kmem cache, because the hotplug callback doesn't yet see the new cache, and cache creation in init_kmem_cache_nodes() only inits kmem_cache_node for nodes in the N_NORMAL_MEMORY nodemask, which however may not yet include the new node, as that happens only later after the MEM_GOING_ONLINE callback. Instead of using get/put_online_mems(), the problem can be solved by SLUB maintaining its own nodemask of nodes for which it has allocated the per-node kmem_cache_node structures. This nodemask would generally mirror the N_NORMAL_MEMORY nodemask, but would be updated only in under SLUB's control in its memory hotplug callbacks under the slab_mutex. This patch adds such nodemask and its handling. Commit 03afc0e25f7f mentiones "issues like [the one above]", but there don't appear to be further issues. All the paths (shared for SLAB and SLUB) taking the memory hotplug locks are also taking the slab_mutex, except kmem_cache_shrink() where 03afc0e25f7f replaced slab_mutex with get/put_online_mems(). We however cannot simply restore slab_mutex in kmem_cache_shrink(), as SLUB can enters the function from a write to sysfs 'shrink' file, thus holding kernfs lock, and in kmem_cache_create() the kernfs lock is nested within slab_mutex. But on closer inspection we don't actually need to protect kmem_cache_shrink() from hotplug callbacks: While SLUB's __kmem_cache_shrink() does for_each_kmem_cache_node(), missing a new node added in parallel hotplug is not fatal, and parallel hotremove does not free kmem_cache_node's anymore after the previous patch, so use-after free cannot happen. The per-node shrinking itself is protected by n->list_lock. Same is true for SLAB, and SLOB is no-op. SLAB also doesn't need the memory hotplug locking, which it only gained by 03afc0e25f7f through the shared paths in slab_common.c. Its memory hotplug callbacks are also protected by slab_mutex against races with these paths. The problem of SLUB relying on N_NORMAL_MEMORY doesn't apply to SLAB, as its setup_kmem_cache_nodes relies on N_ONLINE, and the new node is already set there during the MEM_GOING_ONLINE callback, so no special care is needed for SLAB. As such, this patch removes all get/put_online_mems() usage by the slab subsystem. Link: https://lkml.kernel.org/r/20210113131634.3671-3-vbabka@suse.cz Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Cc: Christoph Lameter <cl@linux.com> Cc: David Hildenbrand <david@redhat.com> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: Qian Cai <cai@redhat.com> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-25 04:01:12 +08:00
/*
* Initialize the nodemask for which we will allocate per node
* structures. Here we don't need taking slab_mutex yet.
*/
for_each_node_state(node, N_NORMAL_MEMORY)
node_set(node, slab_nodes);
create_boot_cache(kmem_cache_node, "kmem_cache_node",
usercopy: Prepare for usercopy whitelisting This patch prepares the slab allocator to handle caches having annotations (useroffset and usersize) defining usercopy regions. This patch is modified from Brad Spengler/PaX Team's PAX_USERCOPY whitelisting code in the last public patch of grsecurity/PaX based on my understanding of the code. Changes or omissions from the original code are mine and don't reflect the original grsecurity/PaX code. Currently, hardened usercopy performs dynamic bounds checking on slab cache objects. This is good, but still leaves a lot of kernel memory available to be copied to/from userspace in the face of bugs. To further restrict what memory is available for copying, this creates a way to whitelist specific areas of a given slab cache object for copying to/from userspace, allowing much finer granularity of access control. Slab caches that are never exposed to userspace can declare no whitelist for their objects, thereby keeping them unavailable to userspace via dynamic copy operations. (Note, an implicit form of whitelisting is the use of constant sizes in usercopy operations and get_user()/put_user(); these bypass hardened usercopy checks since these sizes cannot change at runtime.) To support this whitelist annotation, usercopy region offset and size members are added to struct kmem_cache. The slab allocator receives a new function, kmem_cache_create_usercopy(), that creates a new cache with a usercopy region defined, suitable for declaring spans of fields within the objects that get copied to/from userspace. In this patch, the default kmem_cache_create() marks the entire allocation as whitelisted, leaving it semantically unchanged. Once all fine-grained whitelists have been added (in subsequent patches), this will be changed to a usersize of 0, making caches created with kmem_cache_create() not copyable to/from userspace. After the entire usercopy whitelist series is applied, less than 15% of the slab cache memory remains exposed to potential usercopy bugs after a fresh boot: Total Slab Memory: 48074720 Usercopyable Memory: 6367532 13.2% task_struct 0.2% 4480/1630720 RAW 0.3% 300/96000 RAWv6 2.1% 1408/64768 ext4_inode_cache 3.0% 269760/8740224 dentry 11.1% 585984/5273856 mm_struct 29.1% 54912/188448 kmalloc-8 100.0% 24576/24576 kmalloc-16 100.0% 28672/28672 kmalloc-32 100.0% 81920/81920 kmalloc-192 100.0% 96768/96768 kmalloc-128 100.0% 143360/143360 names_cache 100.0% 163840/163840 kmalloc-64 100.0% 167936/167936 kmalloc-256 100.0% 339968/339968 kmalloc-512 100.0% 350720/350720 kmalloc-96 100.0% 455616/455616 kmalloc-8192 100.0% 655360/655360 kmalloc-1024 100.0% 812032/812032 kmalloc-4096 100.0% 819200/819200 kmalloc-2048 100.0% 1310720/1310720 After some kernel build workloads, the percentage (mainly driven by dentry and inode caches expanding) drops under 10%: Total Slab Memory: 95516184 Usercopyable Memory: 8497452 8.8% task_struct 0.2% 4000/1456000 RAW 0.3% 300/96000 RAWv6 2.1% 1408/64768 ext4_inode_cache 3.0% 1217280/39439872 dentry 11.1% 1623200/14608800 mm_struct 29.1% 73216/251264 kmalloc-8 100.0% 24576/24576 kmalloc-16 100.0% 28672/28672 kmalloc-32 100.0% 94208/94208 kmalloc-192 100.0% 96768/96768 kmalloc-128 100.0% 143360/143360 names_cache 100.0% 163840/163840 kmalloc-64 100.0% 245760/245760 kmalloc-256 100.0% 339968/339968 kmalloc-512 100.0% 350720/350720 kmalloc-96 100.0% 563520/563520 kmalloc-8192 100.0% 655360/655360 kmalloc-1024 100.0% 794624/794624 kmalloc-4096 100.0% 819200/819200 kmalloc-2048 100.0% 1257472/1257472 Signed-off-by: David Windsor <dave@nullcore.net> [kees: adjust commit log, split out a few extra kmalloc hunks] [kees: add field names to function declarations] [kees: convert BUGs to WARNs and fail closed] [kees: add attack surface reduction analysis to commit log] Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: linux-mm@kvack.org Cc: linux-xfs@vger.kernel.org Signed-off-by: Kees Cook <keescook@chromium.org> Acked-by: Christoph Lameter <cl@linux.com>
2017-06-11 10:50:28 +08:00
sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
register_hotmemory_notifier(&slab_memory_callback_nb);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/* Able to allocate the per node structures */
slab_state = PARTIAL;
create_boot_cache(kmem_cache, "kmem_cache",
offsetof(struct kmem_cache, node) +
nr_node_ids * sizeof(struct kmem_cache_node *),
usercopy: Prepare for usercopy whitelisting This patch prepares the slab allocator to handle caches having annotations (useroffset and usersize) defining usercopy regions. This patch is modified from Brad Spengler/PaX Team's PAX_USERCOPY whitelisting code in the last public patch of grsecurity/PaX based on my understanding of the code. Changes or omissions from the original code are mine and don't reflect the original grsecurity/PaX code. Currently, hardened usercopy performs dynamic bounds checking on slab cache objects. This is good, but still leaves a lot of kernel memory available to be copied to/from userspace in the face of bugs. To further restrict what memory is available for copying, this creates a way to whitelist specific areas of a given slab cache object for copying to/from userspace, allowing much finer granularity of access control. Slab caches that are never exposed to userspace can declare no whitelist for their objects, thereby keeping them unavailable to userspace via dynamic copy operations. (Note, an implicit form of whitelisting is the use of constant sizes in usercopy operations and get_user()/put_user(); these bypass hardened usercopy checks since these sizes cannot change at runtime.) To support this whitelist annotation, usercopy region offset and size members are added to struct kmem_cache. The slab allocator receives a new function, kmem_cache_create_usercopy(), that creates a new cache with a usercopy region defined, suitable for declaring spans of fields within the objects that get copied to/from userspace. In this patch, the default kmem_cache_create() marks the entire allocation as whitelisted, leaving it semantically unchanged. Once all fine-grained whitelists have been added (in subsequent patches), this will be changed to a usersize of 0, making caches created with kmem_cache_create() not copyable to/from userspace. After the entire usercopy whitelist series is applied, less than 15% of the slab cache memory remains exposed to potential usercopy bugs after a fresh boot: Total Slab Memory: 48074720 Usercopyable Memory: 6367532 13.2% task_struct 0.2% 4480/1630720 RAW 0.3% 300/96000 RAWv6 2.1% 1408/64768 ext4_inode_cache 3.0% 269760/8740224 dentry 11.1% 585984/5273856 mm_struct 29.1% 54912/188448 kmalloc-8 100.0% 24576/24576 kmalloc-16 100.0% 28672/28672 kmalloc-32 100.0% 81920/81920 kmalloc-192 100.0% 96768/96768 kmalloc-128 100.0% 143360/143360 names_cache 100.0% 163840/163840 kmalloc-64 100.0% 167936/167936 kmalloc-256 100.0% 339968/339968 kmalloc-512 100.0% 350720/350720 kmalloc-96 100.0% 455616/455616 kmalloc-8192 100.0% 655360/655360 kmalloc-1024 100.0% 812032/812032 kmalloc-4096 100.0% 819200/819200 kmalloc-2048 100.0% 1310720/1310720 After some kernel build workloads, the percentage (mainly driven by dentry and inode caches expanding) drops under 10%: Total Slab Memory: 95516184 Usercopyable Memory: 8497452 8.8% task_struct 0.2% 4000/1456000 RAW 0.3% 300/96000 RAWv6 2.1% 1408/64768 ext4_inode_cache 3.0% 1217280/39439872 dentry 11.1% 1623200/14608800 mm_struct 29.1% 73216/251264 kmalloc-8 100.0% 24576/24576 kmalloc-16 100.0% 28672/28672 kmalloc-32 100.0% 94208/94208 kmalloc-192 100.0% 96768/96768 kmalloc-128 100.0% 143360/143360 names_cache 100.0% 163840/163840 kmalloc-64 100.0% 245760/245760 kmalloc-256 100.0% 339968/339968 kmalloc-512 100.0% 350720/350720 kmalloc-96 100.0% 563520/563520 kmalloc-8192 100.0% 655360/655360 kmalloc-1024 100.0% 794624/794624 kmalloc-4096 100.0% 819200/819200 kmalloc-2048 100.0% 1257472/1257472 Signed-off-by: David Windsor <dave@nullcore.net> [kees: adjust commit log, split out a few extra kmalloc hunks] [kees: add field names to function declarations] [kees: convert BUGs to WARNs and fail closed] [kees: add attack surface reduction analysis to commit log] Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: linux-mm@kvack.org Cc: linux-xfs@vger.kernel.org Signed-off-by: Kees Cook <keescook@chromium.org> Acked-by: Christoph Lameter <cl@linux.com>
2017-06-11 10:50:28 +08:00
SLAB_HWCACHE_ALIGN, 0, 0);
kmem_cache = bootstrap(&boot_kmem_cache);
kmem_cache_node = bootstrap(&boot_kmem_cache_node);
/* Now we can use the kmem_cache to allocate kmalloc slabs */
slab: correct size_index table before replacing the bootstrap kmem_cache_node This patch moves the initialization of the size_index table slightly earlier so that the first few kmem_cache_node's can be safely allocated when KMALLOC_MIN_SIZE is large. There are currently two ways to generate indices into kmalloc_caches (via kmalloc_index() and via the size_index table in slab_common.c) and on some arches (possibly only MIPS) they potentially disagree with each other until create_kmalloc_caches() has been called. It seems that the intention is that the size_index table is a fast equivalent to kmalloc_index() and that create_kmalloc_caches() patches the table to return the correct value for the cases where kmalloc_index()'s if-statements apply. The failing sequence was: * kmalloc_caches contains NULL elements * kmem_cache_init initialises the element that 'struct kmem_cache_node' will be allocated to. For 32-bit Mips, this is a 56-byte struct and kmalloc_index returns KMALLOC_SHIFT_LOW (7). * init_list is called which calls kmalloc_node to allocate a 'struct kmem_cache_node'. * kmalloc_slab selects the kmem_caches element using size_index[size_index_elem(size)]. For MIPS, size is 56, and the expression returns 6. * This element of kmalloc_caches is NULL and allocation fails. * If it had not already failed, it would have called create_kmalloc_caches() at this point which would have changed size_index[size_index_elem(size)] to 7. I don't believe the bug to be LLVM specific but GCC doesn't normally encounter the problem. I haven't been able to identify exactly what GCC is doing better (probably inlining) but it seems that GCC is managing to optimize to the point that it eliminates the problematic allocations. This theory is supported by the fact that GCC can be made to fail in the same way by changing inline, __inline, __inline__, and __always_inline in include/linux/compiler-gcc.h such that they don't actually inline things. Signed-off-by: Daniel Sanders <daniel.sanders@imgtec.com> Acked-by: Pekka Enberg <penberg@kernel.org> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-06-25 07:55:57 +08:00
setup_kmalloc_cache_index_table();
create_kmalloc_caches(0);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
mm: SLUB freelist randomization Implements freelist randomization for the SLUB allocator. It was previous implemented for the SLAB allocator. Both use the same configuration option (CONFIG_SLAB_FREELIST_RANDOM). The list is randomized during initialization of a new set of pages. The order on different freelist sizes is pre-computed at boot for performance. Each kmem_cache has its own randomized freelist. This security feature reduces the predictability of the kernel SLUB allocator against heap overflows rendering attacks much less stable. For example these attacks exploit the predictability of the heap: - Linux Kernel CAN SLUB overflow (https://goo.gl/oMNWkU) - Exploiting Linux Kernel Heap corruptions (http://goo.gl/EXLn95) Performance results: slab_test impact is between 3% to 4% on average for 100000 attempts without smp. It is a very focused testing, kernbench show the overall impact on the system is way lower. Before: Single thread testing ===================== 1. Kmalloc: Repeatedly allocate then free test 100000 times kmalloc(8) -> 49 cycles kfree -> 77 cycles 100000 times kmalloc(16) -> 51 cycles kfree -> 79 cycles 100000 times kmalloc(32) -> 53 cycles kfree -> 83 cycles 100000 times kmalloc(64) -> 62 cycles kfree -> 90 cycles 100000 times kmalloc(128) -> 81 cycles kfree -> 97 cycles 100000 times kmalloc(256) -> 98 cycles kfree -> 121 cycles 100000 times kmalloc(512) -> 95 cycles kfree -> 122 cycles 100000 times kmalloc(1024) -> 96 cycles kfree -> 126 cycles 100000 times kmalloc(2048) -> 115 cycles kfree -> 140 cycles 100000 times kmalloc(4096) -> 149 cycles kfree -> 171 cycles 2. Kmalloc: alloc/free test 100000 times kmalloc(8)/kfree -> 70 cycles 100000 times kmalloc(16)/kfree -> 70 cycles 100000 times kmalloc(32)/kfree -> 70 cycles 100000 times kmalloc(64)/kfree -> 70 cycles 100000 times kmalloc(128)/kfree -> 70 cycles 100000 times kmalloc(256)/kfree -> 69 cycles 100000 times kmalloc(512)/kfree -> 70 cycles 100000 times kmalloc(1024)/kfree -> 73 cycles 100000 times kmalloc(2048)/kfree -> 72 cycles 100000 times kmalloc(4096)/kfree -> 71 cycles After: Single thread testing ===================== 1. Kmalloc: Repeatedly allocate then free test 100000 times kmalloc(8) -> 57 cycles kfree -> 78 cycles 100000 times kmalloc(16) -> 61 cycles kfree -> 81 cycles 100000 times kmalloc(32) -> 76 cycles kfree -> 93 cycles 100000 times kmalloc(64) -> 83 cycles kfree -> 94 cycles 100000 times kmalloc(128) -> 106 cycles kfree -> 107 cycles 100000 times kmalloc(256) -> 118 cycles kfree -> 117 cycles 100000 times kmalloc(512) -> 114 cycles kfree -> 116 cycles 100000 times kmalloc(1024) -> 115 cycles kfree -> 118 cycles 100000 times kmalloc(2048) -> 147 cycles kfree -> 131 cycles 100000 times kmalloc(4096) -> 214 cycles kfree -> 161 cycles 2. Kmalloc: alloc/free test 100000 times kmalloc(8)/kfree -> 66 cycles 100000 times kmalloc(16)/kfree -> 66 cycles 100000 times kmalloc(32)/kfree -> 66 cycles 100000 times kmalloc(64)/kfree -> 66 cycles 100000 times kmalloc(128)/kfree -> 65 cycles 100000 times kmalloc(256)/kfree -> 67 cycles 100000 times kmalloc(512)/kfree -> 67 cycles 100000 times kmalloc(1024)/kfree -> 64 cycles 100000 times kmalloc(2048)/kfree -> 67 cycles 100000 times kmalloc(4096)/kfree -> 67 cycles Kernbench, before: Average Optimal load -j 12 Run (std deviation): Elapsed Time 101.873 (1.16069) User Time 1045.22 (1.60447) System Time 88.969 (0.559195) Percent CPU 1112.9 (13.8279) Context Switches 189140 (2282.15) Sleeps 99008.6 (768.091) After: Average Optimal load -j 12 Run (std deviation): Elapsed Time 102.47 (0.562732) User Time 1045.3 (1.34263) System Time 88.311 (0.342554) Percent CPU 1105.8 (6.49444) Context Switches 189081 (2355.78) Sleeps 99231.5 (800.358) Link: http://lkml.kernel.org/r/1464295031-26375-3-git-send-email-thgarnie@google.com Signed-off-by: Thomas Garnier <thgarnie@google.com> Reviewed-by: Kees Cook <keescook@chromium.org> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-07-27 06:21:59 +08:00
/* Setup random freelists for each cache */
init_freelist_randomization();
cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
slub_cpu_dead);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
cache_line_size(),
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
slub_min_order, slub_max_order, slub_min_objects,
nr_cpu_ids, nr_node_ids);
}
void __init kmem_cache_init_late(void)
{
}
struct kmem_cache *
__kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
slab_flags_t flags, void (*ctor)(void *))
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
{
struct kmem_cache *s;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
memcg, slab: never try to merge memcg caches When a kmem cache is created (kmem_cache_create_memcg()), we first try to find a compatible cache that already exists and can handle requests from the new cache, i.e. has the same object size, alignment, ctor, etc. If there is such a cache, we do not create any new caches, instead we simply increment the refcount of the cache found and return it. Currently we do this procedure not only when creating root caches, but also for memcg caches. However, there is no point in that, because, as every memcg cache has exactly the same parameters as its parent and cache merging cannot be turned off in runtime (only on boot by passing "slub_nomerge"), the root caches of any two potentially mergeable memcg caches should be merged already, i.e. it must be the same root cache, and therefore we couldn't even get to the memcg cache creation, because it already exists. The only exception is boot caches - they are explicitly forbidden to be merged by setting their refcount to -1. There are currently only two of them - kmem_cache and kmem_cache_node, which are used in slab internals (I do not count kmalloc caches as their refcount is set to 1 immediately after creation). Since they are prevented from merging preliminary I guess we should avoid to merge their children too. So let's remove the useless code responsible for merging memcg caches. Signed-off-by: Vladimir Davydov <vdavydov@parallels.com> Cc: Michal Hocko <mhocko@suse.cz> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: David Rientjes <rientjes@google.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: Glauber Costa <glommer@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-04-08 06:39:23 +08:00
s = find_mergeable(size, align, flags, name, ctor);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (s) {
s->refcount++;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/*
* Adjust the object sizes so that we clear
* the complete object on kzalloc.
*/
s->object_size = max(s->object_size, size);
s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
if (sysfs_slab_alias(s, name)) {
s->refcount--;
s = NULL;
}
}
return s;
}
int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
{
Revert "mm/sl[aou]b: Move sysfs_slab_add to common" This reverts commit 96d17b7be0a9849d381442030886211dbb2a7061 which caused the following errors at boot: [ 1.114885] kobject (ffff88001a802578): tried to init an initialized object, something is seriously wrong. [ 1.114885] Pid: 1, comm: swapper/0 Tainted: G W 3.6.0-rc1+ #6 [ 1.114885] Call Trace: [ 1.114885] [<ffffffff81273f37>] kobject_init+0x87/0xa0 [ 1.115555] [<ffffffff8127426a>] kobject_init_and_add+0x2a/0x90 [ 1.115555] [<ffffffff8127c870>] ? sprintf+0x40/0x50 [ 1.115555] [<ffffffff81124c60>] sysfs_slab_add+0x80/0x210 [ 1.115555] [<ffffffff81100175>] kmem_cache_create+0xa5/0x250 [ 1.115555] [<ffffffff81cf24cd>] ? md_init+0x144/0x144 [ 1.115555] [<ffffffff81cf25b6>] local_init+0xa4/0x11b [ 1.115555] [<ffffffff81cf24e1>] dm_init+0x14/0x45 [ 1.115836] [<ffffffff810001ba>] do_one_initcall+0x3a/0x160 [ 1.116834] [<ffffffff81cc2c90>] kernel_init+0x133/0x1b7 [ 1.117835] [<ffffffff81cc25c4>] ? do_early_param+0x86/0x86 [ 1.117835] [<ffffffff8171aff4>] kernel_thread_helper+0x4/0x10 [ 1.118401] [<ffffffff81cc2b5d>] ? start_kernel+0x33f/0x33f [ 1.119832] [<ffffffff8171aff0>] ? gs_change+0xb/0xb [ 1.120325] ------------[ cut here ]------------ [ 1.120835] WARNING: at fs/sysfs/dir.c:536 sysfs_add_one+0xc1/0xf0() [ 1.121437] sysfs: cannot create duplicate filename '/kernel/slab/:t-0000016' [ 1.121831] Modules linked in: [ 1.122138] Pid: 1, comm: swapper/0 Tainted: G W 3.6.0-rc1+ #6 [ 1.122831] Call Trace: [ 1.123074] [<ffffffff81195ce1>] ? sysfs_add_one+0xc1/0xf0 [ 1.123833] [<ffffffff8103adfa>] warn_slowpath_common+0x7a/0xb0 [ 1.124405] [<ffffffff8103aed1>] warn_slowpath_fmt+0x41/0x50 [ 1.124832] [<ffffffff81195ce1>] sysfs_add_one+0xc1/0xf0 [ 1.125337] [<ffffffff81195eb3>] create_dir+0x73/0xd0 [ 1.125832] [<ffffffff81196221>] sysfs_create_dir+0x81/0xe0 [ 1.126363] [<ffffffff81273d3d>] kobject_add_internal+0x9d/0x210 [ 1.126832] [<ffffffff812742a3>] kobject_init_and_add+0x63/0x90 [ 1.127406] [<ffffffff81124c60>] sysfs_slab_add+0x80/0x210 [ 1.127832] [<ffffffff81100175>] kmem_cache_create+0xa5/0x250 [ 1.128384] [<ffffffff81cf24cd>] ? md_init+0x144/0x144 [ 1.128833] [<ffffffff81cf25b6>] local_init+0xa4/0x11b [ 1.129831] [<ffffffff81cf24e1>] dm_init+0x14/0x45 [ 1.130305] [<ffffffff810001ba>] do_one_initcall+0x3a/0x160 [ 1.130831] [<ffffffff81cc2c90>] kernel_init+0x133/0x1b7 [ 1.131351] [<ffffffff81cc25c4>] ? do_early_param+0x86/0x86 [ 1.131830] [<ffffffff8171aff4>] kernel_thread_helper+0x4/0x10 [ 1.132392] [<ffffffff81cc2b5d>] ? start_kernel+0x33f/0x33f [ 1.132830] [<ffffffff8171aff0>] ? gs_change+0xb/0xb [ 1.133315] ---[ end trace 2703540871c8fab7 ]--- [ 1.133830] ------------[ cut here ]------------ [ 1.134274] WARNING: at lib/kobject.c:196 kobject_add_internal+0x1f5/0x210() [ 1.134829] kobject_add_internal failed for :t-0000016 with -EEXIST, don't try to register things with the same name in the same directory. [ 1.135829] Modules linked in: [ 1.136135] Pid: 1, comm: swapper/0 Tainted: G W 3.6.0-rc1+ #6 [ 1.136828] Call Trace: [ 1.137071] [<ffffffff81273e95>] ? kobject_add_internal+0x1f5/0x210 [ 1.137830] [<ffffffff8103adfa>] warn_slowpath_common+0x7a/0xb0 [ 1.138402] [<ffffffff8103aed1>] warn_slowpath_fmt+0x41/0x50 [ 1.138830] [<ffffffff811955a3>] ? release_sysfs_dirent+0x73/0xf0 [ 1.139419] [<ffffffff81273e95>] kobject_add_internal+0x1f5/0x210 [ 1.139830] [<ffffffff812742a3>] kobject_init_and_add+0x63/0x90 [ 1.140429] [<ffffffff81124c60>] sysfs_slab_add+0x80/0x210 [ 1.140830] [<ffffffff81100175>] kmem_cache_create+0xa5/0x250 [ 1.141829] [<ffffffff81cf24cd>] ? md_init+0x144/0x144 [ 1.142307] [<ffffffff81cf25b6>] local_init+0xa4/0x11b [ 1.142829] [<ffffffff81cf24e1>] dm_init+0x14/0x45 [ 1.143307] [<ffffffff810001ba>] do_one_initcall+0x3a/0x160 [ 1.143829] [<ffffffff81cc2c90>] kernel_init+0x133/0x1b7 [ 1.144352] [<ffffffff81cc25c4>] ? do_early_param+0x86/0x86 [ 1.144829] [<ffffffff8171aff4>] kernel_thread_helper+0x4/0x10 [ 1.145405] [<ffffffff81cc2b5d>] ? start_kernel+0x33f/0x33f [ 1.145828] [<ffffffff8171aff0>] ? gs_change+0xb/0xb [ 1.146313] ---[ end trace 2703540871c8fab8 ]--- Conflicts: mm/slub.c Signed-off-by: Pekka Enberg <penberg@kernel.org>
2012-09-05 17:07:44 +08:00
int err;
err = kmem_cache_open(s, flags);
if (err)
return err;
/* Mutex is not taken during early boot */
if (slab_state <= UP)
return 0;
Revert "mm/sl[aou]b: Move sysfs_slab_add to common" This reverts commit 96d17b7be0a9849d381442030886211dbb2a7061 which caused the following errors at boot: [ 1.114885] kobject (ffff88001a802578): tried to init an initialized object, something is seriously wrong. [ 1.114885] Pid: 1, comm: swapper/0 Tainted: G W 3.6.0-rc1+ #6 [ 1.114885] Call Trace: [ 1.114885] [<ffffffff81273f37>] kobject_init+0x87/0xa0 [ 1.115555] [<ffffffff8127426a>] kobject_init_and_add+0x2a/0x90 [ 1.115555] [<ffffffff8127c870>] ? sprintf+0x40/0x50 [ 1.115555] [<ffffffff81124c60>] sysfs_slab_add+0x80/0x210 [ 1.115555] [<ffffffff81100175>] kmem_cache_create+0xa5/0x250 [ 1.115555] [<ffffffff81cf24cd>] ? md_init+0x144/0x144 [ 1.115555] [<ffffffff81cf25b6>] local_init+0xa4/0x11b [ 1.115555] [<ffffffff81cf24e1>] dm_init+0x14/0x45 [ 1.115836] [<ffffffff810001ba>] do_one_initcall+0x3a/0x160 [ 1.116834] [<ffffffff81cc2c90>] kernel_init+0x133/0x1b7 [ 1.117835] [<ffffffff81cc25c4>] ? do_early_param+0x86/0x86 [ 1.117835] [<ffffffff8171aff4>] kernel_thread_helper+0x4/0x10 [ 1.118401] [<ffffffff81cc2b5d>] ? start_kernel+0x33f/0x33f [ 1.119832] [<ffffffff8171aff0>] ? gs_change+0xb/0xb [ 1.120325] ------------[ cut here ]------------ [ 1.120835] WARNING: at fs/sysfs/dir.c:536 sysfs_add_one+0xc1/0xf0() [ 1.121437] sysfs: cannot create duplicate filename '/kernel/slab/:t-0000016' [ 1.121831] Modules linked in: [ 1.122138] Pid: 1, comm: swapper/0 Tainted: G W 3.6.0-rc1+ #6 [ 1.122831] Call Trace: [ 1.123074] [<ffffffff81195ce1>] ? sysfs_add_one+0xc1/0xf0 [ 1.123833] [<ffffffff8103adfa>] warn_slowpath_common+0x7a/0xb0 [ 1.124405] [<ffffffff8103aed1>] warn_slowpath_fmt+0x41/0x50 [ 1.124832] [<ffffffff81195ce1>] sysfs_add_one+0xc1/0xf0 [ 1.125337] [<ffffffff81195eb3>] create_dir+0x73/0xd0 [ 1.125832] [<ffffffff81196221>] sysfs_create_dir+0x81/0xe0 [ 1.126363] [<ffffffff81273d3d>] kobject_add_internal+0x9d/0x210 [ 1.126832] [<ffffffff812742a3>] kobject_init_and_add+0x63/0x90 [ 1.127406] [<ffffffff81124c60>] sysfs_slab_add+0x80/0x210 [ 1.127832] [<ffffffff81100175>] kmem_cache_create+0xa5/0x250 [ 1.128384] [<ffffffff81cf24cd>] ? md_init+0x144/0x144 [ 1.128833] [<ffffffff81cf25b6>] local_init+0xa4/0x11b [ 1.129831] [<ffffffff81cf24e1>] dm_init+0x14/0x45 [ 1.130305] [<ffffffff810001ba>] do_one_initcall+0x3a/0x160 [ 1.130831] [<ffffffff81cc2c90>] kernel_init+0x133/0x1b7 [ 1.131351] [<ffffffff81cc25c4>] ? do_early_param+0x86/0x86 [ 1.131830] [<ffffffff8171aff4>] kernel_thread_helper+0x4/0x10 [ 1.132392] [<ffffffff81cc2b5d>] ? start_kernel+0x33f/0x33f [ 1.132830] [<ffffffff8171aff0>] ? gs_change+0xb/0xb [ 1.133315] ---[ end trace 2703540871c8fab7 ]--- [ 1.133830] ------------[ cut here ]------------ [ 1.134274] WARNING: at lib/kobject.c:196 kobject_add_internal+0x1f5/0x210() [ 1.134829] kobject_add_internal failed for :t-0000016 with -EEXIST, don't try to register things with the same name in the same directory. [ 1.135829] Modules linked in: [ 1.136135] Pid: 1, comm: swapper/0 Tainted: G W 3.6.0-rc1+ #6 [ 1.136828] Call Trace: [ 1.137071] [<ffffffff81273e95>] ? kobject_add_internal+0x1f5/0x210 [ 1.137830] [<ffffffff8103adfa>] warn_slowpath_common+0x7a/0xb0 [ 1.138402] [<ffffffff8103aed1>] warn_slowpath_fmt+0x41/0x50 [ 1.138830] [<ffffffff811955a3>] ? release_sysfs_dirent+0x73/0xf0 [ 1.139419] [<ffffffff81273e95>] kobject_add_internal+0x1f5/0x210 [ 1.139830] [<ffffffff812742a3>] kobject_init_and_add+0x63/0x90 [ 1.140429] [<ffffffff81124c60>] sysfs_slab_add+0x80/0x210 [ 1.140830] [<ffffffff81100175>] kmem_cache_create+0xa5/0x250 [ 1.141829] [<ffffffff81cf24cd>] ? md_init+0x144/0x144 [ 1.142307] [<ffffffff81cf25b6>] local_init+0xa4/0x11b [ 1.142829] [<ffffffff81cf24e1>] dm_init+0x14/0x45 [ 1.143307] [<ffffffff810001ba>] do_one_initcall+0x3a/0x160 [ 1.143829] [<ffffffff81cc2c90>] kernel_init+0x133/0x1b7 [ 1.144352] [<ffffffff81cc25c4>] ? do_early_param+0x86/0x86 [ 1.144829] [<ffffffff8171aff4>] kernel_thread_helper+0x4/0x10 [ 1.145405] [<ffffffff81cc2b5d>] ? start_kernel+0x33f/0x33f [ 1.145828] [<ffffffff8171aff0>] ? gs_change+0xb/0xb [ 1.146313] ---[ end trace 2703540871c8fab8 ]--- Conflicts: mm/slub.c Signed-off-by: Pekka Enberg <penberg@kernel.org>
2012-09-05 17:07:44 +08:00
err = sysfs_slab_add(s);
if (err)
mm: slab: free kmem_cache_node after destroy sysfs file When slub_debug alloc_calls_show is enabled we will try to track location and user of slab object on each online node, kmem_cache_node structure and cpu_cache/cpu_slub shouldn't be freed till there is the last reference to sysfs file. This fixes the following panic: BUG: unable to handle kernel NULL pointer dereference at 0000000000000020 IP: list_locations+0x169/0x4e0 PGD 257304067 PUD 438456067 PMD 0 Oops: 0000 [#1] SMP CPU: 3 PID: 973074 Comm: cat ve: 0 Not tainted 3.10.0-229.7.2.ovz.9.30-00007-japdoll-dirty #2 9.30 Hardware name: DEPO Computers To Be Filled By O.E.M./H67DE3, BIOS L1.60c 07/14/2011 task: ffff88042a5dc5b0 ti: ffff88037f8d8000 task.ti: ffff88037f8d8000 RIP: list_locations+0x169/0x4e0 Call Trace: alloc_calls_show+0x1d/0x30 slab_attr_show+0x1b/0x30 sysfs_read_file+0x9a/0x1a0 vfs_read+0x9c/0x170 SyS_read+0x58/0xb0 system_call_fastpath+0x16/0x1b Code: 5e 07 12 00 b9 00 04 00 00 3d 00 04 00 00 0f 4f c1 3d 00 04 00 00 89 45 b0 0f 84 c3 00 00 00 48 63 45 b0 49 8b 9c c4 f8 00 00 00 <48> 8b 43 20 48 85 c0 74 b6 48 89 df e8 46 37 44 00 48 8b 53 10 CR2: 0000000000000020 Separated __kmem_cache_release from __kmem_cache_shutdown which now called on slab_kmem_cache_release (after the last reference to sysfs file object has dropped). Reintroduced locking in free_partial as sysfs file might access cache's partial list after shutdowning - partial revert of the commit 69cb8e6b7c29 ("slub: free slabs without holding locks"). Zap __remove_partial and use remove_partial (w/o underscores) as free_partial now takes list_lock which s partial revert for commit 1e4dd9461fab ("slub: do not assert not having lock in removing freed partial") Signed-off-by: Dmitry Safonov <dsafonov@virtuozzo.com> Suggested-by: Vladimir Davydov <vdavydov@virtuozzo.com> Acked-by: Vladimir Davydov <vdavydov@virtuozzo.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-02-18 05:11:37 +08:00
__kmem_cache_release(s);
Revert "mm/sl[aou]b: Move sysfs_slab_add to common" This reverts commit 96d17b7be0a9849d381442030886211dbb2a7061 which caused the following errors at boot: [ 1.114885] kobject (ffff88001a802578): tried to init an initialized object, something is seriously wrong. [ 1.114885] Pid: 1, comm: swapper/0 Tainted: G W 3.6.0-rc1+ #6 [ 1.114885] Call Trace: [ 1.114885] [<ffffffff81273f37>] kobject_init+0x87/0xa0 [ 1.115555] [<ffffffff8127426a>] kobject_init_and_add+0x2a/0x90 [ 1.115555] [<ffffffff8127c870>] ? sprintf+0x40/0x50 [ 1.115555] [<ffffffff81124c60>] sysfs_slab_add+0x80/0x210 [ 1.115555] [<ffffffff81100175>] kmem_cache_create+0xa5/0x250 [ 1.115555] [<ffffffff81cf24cd>] ? md_init+0x144/0x144 [ 1.115555] [<ffffffff81cf25b6>] local_init+0xa4/0x11b [ 1.115555] [<ffffffff81cf24e1>] dm_init+0x14/0x45 [ 1.115836] [<ffffffff810001ba>] do_one_initcall+0x3a/0x160 [ 1.116834] [<ffffffff81cc2c90>] kernel_init+0x133/0x1b7 [ 1.117835] [<ffffffff81cc25c4>] ? do_early_param+0x86/0x86 [ 1.117835] [<ffffffff8171aff4>] kernel_thread_helper+0x4/0x10 [ 1.118401] [<ffffffff81cc2b5d>] ? start_kernel+0x33f/0x33f [ 1.119832] [<ffffffff8171aff0>] ? gs_change+0xb/0xb [ 1.120325] ------------[ cut here ]------------ [ 1.120835] WARNING: at fs/sysfs/dir.c:536 sysfs_add_one+0xc1/0xf0() [ 1.121437] sysfs: cannot create duplicate filename '/kernel/slab/:t-0000016' [ 1.121831] Modules linked in: [ 1.122138] Pid: 1, comm: swapper/0 Tainted: G W 3.6.0-rc1+ #6 [ 1.122831] Call Trace: [ 1.123074] [<ffffffff81195ce1>] ? sysfs_add_one+0xc1/0xf0 [ 1.123833] [<ffffffff8103adfa>] warn_slowpath_common+0x7a/0xb0 [ 1.124405] [<ffffffff8103aed1>] warn_slowpath_fmt+0x41/0x50 [ 1.124832] [<ffffffff81195ce1>] sysfs_add_one+0xc1/0xf0 [ 1.125337] [<ffffffff81195eb3>] create_dir+0x73/0xd0 [ 1.125832] [<ffffffff81196221>] sysfs_create_dir+0x81/0xe0 [ 1.126363] [<ffffffff81273d3d>] kobject_add_internal+0x9d/0x210 [ 1.126832] [<ffffffff812742a3>] kobject_init_and_add+0x63/0x90 [ 1.127406] [<ffffffff81124c60>] sysfs_slab_add+0x80/0x210 [ 1.127832] [<ffffffff81100175>] kmem_cache_create+0xa5/0x250 [ 1.128384] [<ffffffff81cf24cd>] ? md_init+0x144/0x144 [ 1.128833] [<ffffffff81cf25b6>] local_init+0xa4/0x11b [ 1.129831] [<ffffffff81cf24e1>] dm_init+0x14/0x45 [ 1.130305] [<ffffffff810001ba>] do_one_initcall+0x3a/0x160 [ 1.130831] [<ffffffff81cc2c90>] kernel_init+0x133/0x1b7 [ 1.131351] [<ffffffff81cc25c4>] ? do_early_param+0x86/0x86 [ 1.131830] [<ffffffff8171aff4>] kernel_thread_helper+0x4/0x10 [ 1.132392] [<ffffffff81cc2b5d>] ? start_kernel+0x33f/0x33f [ 1.132830] [<ffffffff8171aff0>] ? gs_change+0xb/0xb [ 1.133315] ---[ end trace 2703540871c8fab7 ]--- [ 1.133830] ------------[ cut here ]------------ [ 1.134274] WARNING: at lib/kobject.c:196 kobject_add_internal+0x1f5/0x210() [ 1.134829] kobject_add_internal failed for :t-0000016 with -EEXIST, don't try to register things with the same name in the same directory. [ 1.135829] Modules linked in: [ 1.136135] Pid: 1, comm: swapper/0 Tainted: G W 3.6.0-rc1+ #6 [ 1.136828] Call Trace: [ 1.137071] [<ffffffff81273e95>] ? kobject_add_internal+0x1f5/0x210 [ 1.137830] [<ffffffff8103adfa>] warn_slowpath_common+0x7a/0xb0 [ 1.138402] [<ffffffff8103aed1>] warn_slowpath_fmt+0x41/0x50 [ 1.138830] [<ffffffff811955a3>] ? release_sysfs_dirent+0x73/0xf0 [ 1.139419] [<ffffffff81273e95>] kobject_add_internal+0x1f5/0x210 [ 1.139830] [<ffffffff812742a3>] kobject_init_and_add+0x63/0x90 [ 1.140429] [<ffffffff81124c60>] sysfs_slab_add+0x80/0x210 [ 1.140830] [<ffffffff81100175>] kmem_cache_create+0xa5/0x250 [ 1.141829] [<ffffffff81cf24cd>] ? md_init+0x144/0x144 [ 1.142307] [<ffffffff81cf25b6>] local_init+0xa4/0x11b [ 1.142829] [<ffffffff81cf24e1>] dm_init+0x14/0x45 [ 1.143307] [<ffffffff810001ba>] do_one_initcall+0x3a/0x160 [ 1.143829] [<ffffffff81cc2c90>] kernel_init+0x133/0x1b7 [ 1.144352] [<ffffffff81cc25c4>] ? do_early_param+0x86/0x86 [ 1.144829] [<ffffffff8171aff4>] kernel_thread_helper+0x4/0x10 [ 1.145405] [<ffffffff81cc2b5d>] ? start_kernel+0x33f/0x33f [ 1.145828] [<ffffffff8171aff0>] ? gs_change+0xb/0xb [ 1.146313] ---[ end trace 2703540871c8fab8 ]--- Conflicts: mm/slub.c Signed-off-by: Pekka Enberg <penberg@kernel.org>
2012-09-05 17:07:44 +08:00
return err;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
{
SLUB: direct pass through of page size or higher kmalloc requests This gets rid of all kmalloc caches larger than page size. A kmalloc request larger than PAGE_SIZE > 2 is going to be passed through to the page allocator. This works both inline where we will call __get_free_pages instead of kmem_cache_alloc and in __kmalloc. kfree is modified to check if the object is in a slab page. If not then the page is freed via the page allocator instead. Roughly similar to what SLOB does. Advantages: - Reduces memory overhead for kmalloc array - Large kmalloc operations are faster since they do not need to pass through the slab allocator to get to the page allocator. - Performance increase of 10%-20% on alloc and 50% on free for PAGE_SIZEd allocations. SLUB must call page allocator for each alloc anyways since the higher order pages which that allowed avoiding the page alloc calls are not available in a reliable way anymore. So we are basically removing useless slab allocator overhead. - Large kmallocs yields page aligned object which is what SLAB did. Bad things like using page sized kmalloc allocations to stand in for page allocate allocs can be transparently handled and are not distinguishable from page allocator uses. - Checking for too large objects can be removed since it is done by the page allocator. Drawbacks: - No accounting for large kmalloc slab allocations anymore - No debugging of large kmalloc slab allocations. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 16:24:38 +08:00
struct kmem_cache *s;
void *ret;
SLUB: direct pass through of page size or higher kmalloc requests This gets rid of all kmalloc caches larger than page size. A kmalloc request larger than PAGE_SIZE > 2 is going to be passed through to the page allocator. This works both inline where we will call __get_free_pages instead of kmem_cache_alloc and in __kmalloc. kfree is modified to check if the object is in a slab page. If not then the page is freed via the page allocator instead. Roughly similar to what SLOB does. Advantages: - Reduces memory overhead for kmalloc array - Large kmalloc operations are faster since they do not need to pass through the slab allocator to get to the page allocator. - Performance increase of 10%-20% on alloc and 50% on free for PAGE_SIZEd allocations. SLUB must call page allocator for each alloc anyways since the higher order pages which that allowed avoiding the page alloc calls are not available in a reliable way anymore. So we are basically removing useless slab allocator overhead. - Large kmallocs yields page aligned object which is what SLAB did. Bad things like using page sized kmalloc allocations to stand in for page allocate allocs can be transparently handled and are not distinguishable from page allocator uses. - Checking for too large objects can be removed since it is done by the page allocator. Drawbacks: - No accounting for large kmalloc slab allocations anymore - No debugging of large kmalloc slab allocations. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 16:24:38 +08:00
if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
return kmalloc_large(size, gfpflags);
s = kmalloc_slab(size, gfpflags);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (unlikely(ZERO_OR_NULL_PTR(s)))
return s;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
mm, kfence: insert KFENCE hooks for SLUB Inserts KFENCE hooks into the SLUB allocator. To pass the originally requested size to KFENCE, add an argument 'orig_size' to slab_alloc*(). The additional argument is required to preserve the requested original size for kmalloc() allocations, which uses size classes (e.g. an allocation of 272 bytes will return an object of size 512). Therefore, kmem_cache::size does not represent the kmalloc-caller's requested size, and we must introduce the argument 'orig_size' to propagate the originally requested size to KFENCE. Without the originally requested size, we would not be able to detect out-of-bounds accesses for objects placed at the end of a KFENCE object page if that object is not equal to the kmalloc-size class it was bucketed into. When KFENCE is disabled, there is no additional overhead, since slab_alloc*() functions are __always_inline. Link: https://lkml.kernel.org/r/20201103175841.3495947-6-elver@google.com Signed-off-by: Marco Elver <elver@google.com> Signed-off-by: Alexander Potapenko <glider@google.com> Reviewed-by: Dmitry Vyukov <dvyukov@google.com> Reviewed-by: Jann Horn <jannh@google.com> Co-developed-by: Marco Elver <elver@google.com> Cc: Andrey Konovalov <andreyknvl@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christopher Lameter <cl@linux.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: David Rientjes <rientjes@google.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Hillf Danton <hdanton@sina.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Joern Engel <joern@purestorage.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kees Cook <keescook@chromium.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Paul E. McKenney <paulmck@kernel.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: SeongJae Park <sjpark@amazon.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-26 09:19:16 +08:00
ret = slab_alloc(s, gfpflags, caller, size);
/* Honor the call site pointer we received. */
trace_kmalloc(caller, ret, size, s->size, gfpflags);
return ret;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
EXPORT_SYMBOL(__kmalloc_track_caller);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
#ifdef CONFIG_NUMA
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
int node, unsigned long caller)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
{
SLUB: direct pass through of page size or higher kmalloc requests This gets rid of all kmalloc caches larger than page size. A kmalloc request larger than PAGE_SIZE > 2 is going to be passed through to the page allocator. This works both inline where we will call __get_free_pages instead of kmem_cache_alloc and in __kmalloc. kfree is modified to check if the object is in a slab page. If not then the page is freed via the page allocator instead. Roughly similar to what SLOB does. Advantages: - Reduces memory overhead for kmalloc array - Large kmalloc operations are faster since they do not need to pass through the slab allocator to get to the page allocator. - Performance increase of 10%-20% on alloc and 50% on free for PAGE_SIZEd allocations. SLUB must call page allocator for each alloc anyways since the higher order pages which that allowed avoiding the page alloc calls are not available in a reliable way anymore. So we are basically removing useless slab allocator overhead. - Large kmallocs yields page aligned object which is what SLAB did. Bad things like using page sized kmalloc allocations to stand in for page allocate allocs can be transparently handled and are not distinguishable from page allocator uses. - Checking for too large objects can be removed since it is done by the page allocator. Drawbacks: - No accounting for large kmalloc slab allocations anymore - No debugging of large kmalloc slab allocations. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 16:24:38 +08:00
struct kmem_cache *s;
void *ret;
SLUB: direct pass through of page size or higher kmalloc requests This gets rid of all kmalloc caches larger than page size. A kmalloc request larger than PAGE_SIZE > 2 is going to be passed through to the page allocator. This works both inline where we will call __get_free_pages instead of kmem_cache_alloc and in __kmalloc. kfree is modified to check if the object is in a slab page. If not then the page is freed via the page allocator instead. Roughly similar to what SLOB does. Advantages: - Reduces memory overhead for kmalloc array - Large kmalloc operations are faster since they do not need to pass through the slab allocator to get to the page allocator. - Performance increase of 10%-20% on alloc and 50% on free for PAGE_SIZEd allocations. SLUB must call page allocator for each alloc anyways since the higher order pages which that allowed avoiding the page alloc calls are not available in a reliable way anymore. So we are basically removing useless slab allocator overhead. - Large kmallocs yields page aligned object which is what SLAB did. Bad things like using page sized kmalloc allocations to stand in for page allocate allocs can be transparently handled and are not distinguishable from page allocator uses. - Checking for too large objects can be removed since it is done by the page allocator. Drawbacks: - No accounting for large kmalloc slab allocations anymore - No debugging of large kmalloc slab allocations. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 16:24:38 +08:00
if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
ret = kmalloc_large_node(size, gfpflags, node);
trace_kmalloc_node(caller, ret,
size, PAGE_SIZE << get_order(size),
gfpflags, node);
return ret;
}
s = kmalloc_slab(size, gfpflags);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (unlikely(ZERO_OR_NULL_PTR(s)))
return s;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
mm, kfence: insert KFENCE hooks for SLUB Inserts KFENCE hooks into the SLUB allocator. To pass the originally requested size to KFENCE, add an argument 'orig_size' to slab_alloc*(). The additional argument is required to preserve the requested original size for kmalloc() allocations, which uses size classes (e.g. an allocation of 272 bytes will return an object of size 512). Therefore, kmem_cache::size does not represent the kmalloc-caller's requested size, and we must introduce the argument 'orig_size' to propagate the originally requested size to KFENCE. Without the originally requested size, we would not be able to detect out-of-bounds accesses for objects placed at the end of a KFENCE object page if that object is not equal to the kmalloc-size class it was bucketed into. When KFENCE is disabled, there is no additional overhead, since slab_alloc*() functions are __always_inline. Link: https://lkml.kernel.org/r/20201103175841.3495947-6-elver@google.com Signed-off-by: Marco Elver <elver@google.com> Signed-off-by: Alexander Potapenko <glider@google.com> Reviewed-by: Dmitry Vyukov <dvyukov@google.com> Reviewed-by: Jann Horn <jannh@google.com> Co-developed-by: Marco Elver <elver@google.com> Cc: Andrey Konovalov <andreyknvl@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Christopher Lameter <cl@linux.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: David Rientjes <rientjes@google.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Hillf Danton <hdanton@sina.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Joern Engel <joern@purestorage.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kees Cook <keescook@chromium.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Paul E. McKenney <paulmck@kernel.org> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: SeongJae Park <sjpark@amazon.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Will Deacon <will@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-02-26 09:19:16 +08:00
ret = slab_alloc_node(s, gfpflags, node, caller, size);
/* Honor the call site pointer we received. */
trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
return ret;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
EXPORT_SYMBOL(__kmalloc_node_track_caller);
#endif
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
#ifdef CONFIG_SYSFS
static int count_inuse(struct page *page)
{
return page->inuse;
}
static int count_total(struct page *page)
{
return page->objects;
}
#endif
#ifdef CONFIG_SLUB_DEBUG
static void validate_slab(struct kmem_cache *s, struct page *page)
slub: validation of slabs (metadata and guard zones) This enables validation of slab. Validation means that all objects are checked to see if there are redzone violations, if padding has been overwritten or any pointers have been corrupted. Also checks the consistency of slab counters. Validation enables the detection of metadata corruption without the kernel having to execute code that actually uses (allocs/frees) and object. It allows one to make sure that the slab metainformation and the guard values around an object have not been compromised. A single slabcache can be checked by writing a 1 to the "validate" file. i.e. echo 1 >/sys/slab/kmalloc-128/validate or use the slabinfo tool to check all slabs slabinfo -v Error messages will show up in the syslog. Note that validation can only reach slabs that are on a list. This means that we are usually restricted to partial slabs and active slabs unless SLAB_STORE_USER is active which will build a full slab list and allows validation of slabs that are fully in use. Booting with "slub_debug" set will enable SLAB_STORE_USER and then full diagnostic are available. Note that we attempt to push cpu slabs back to the lists when we start the check. If the cpu slab is reactivated before we get to it (another processor grabs it before we get to it) then it cannot be checked. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:43 +08:00
{
void *p;
void *addr = page_address(page);
unsigned long *map;
slab_lock(page);
slub: validation of slabs (metadata and guard zones) This enables validation of slab. Validation means that all objects are checked to see if there are redzone violations, if padding has been overwritten or any pointers have been corrupted. Also checks the consistency of slab counters. Validation enables the detection of metadata corruption without the kernel having to execute code that actually uses (allocs/frees) and object. It allows one to make sure that the slab metainformation and the guard values around an object have not been compromised. A single slabcache can be checked by writing a 1 to the "validate" file. i.e. echo 1 >/sys/slab/kmalloc-128/validate or use the slabinfo tool to check all slabs slabinfo -v Error messages will show up in the syslog. Note that validation can only reach slabs that are on a list. This means that we are usually restricted to partial slabs and active slabs unless SLAB_STORE_USER is active which will build a full slab list and allows validation of slabs that are fully in use. Booting with "slub_debug" set will enable SLAB_STORE_USER and then full diagnostic are available. Note that we attempt to push cpu slabs back to the lists when we start the check. If the cpu slab is reactivated before we get to it (another processor grabs it before we get to it) then it cannot be checked. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:43 +08:00
if (!check_slab(s, page) || !on_freelist(s, page, NULL))
goto unlock;
slub: validation of slabs (metadata and guard zones) This enables validation of slab. Validation means that all objects are checked to see if there are redzone violations, if padding has been overwritten or any pointers have been corrupted. Also checks the consistency of slab counters. Validation enables the detection of metadata corruption without the kernel having to execute code that actually uses (allocs/frees) and object. It allows one to make sure that the slab metainformation and the guard values around an object have not been compromised. A single slabcache can be checked by writing a 1 to the "validate" file. i.e. echo 1 >/sys/slab/kmalloc-128/validate or use the slabinfo tool to check all slabs slabinfo -v Error messages will show up in the syslog. Note that validation can only reach slabs that are on a list. This means that we are usually restricted to partial slabs and active slabs unless SLAB_STORE_USER is active which will build a full slab list and allows validation of slabs that are fully in use. Booting with "slub_debug" set will enable SLAB_STORE_USER and then full diagnostic are available. Note that we attempt to push cpu slabs back to the lists when we start the check. If the cpu slab is reactivated before we get to it (another processor grabs it before we get to it) then it cannot be checked. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:43 +08:00
/* Now we know that a valid freelist exists */
map = get_map(s, page);
for_each_object(p, s, addr, page->objects) {
u8 val = test_bit(__obj_to_index(s, addr, p), map) ?
SLUB_RED_INACTIVE : SLUB_RED_ACTIVE;
slub: validation of slabs (metadata and guard zones) This enables validation of slab. Validation means that all objects are checked to see if there are redzone violations, if padding has been overwritten or any pointers have been corrupted. Also checks the consistency of slab counters. Validation enables the detection of metadata corruption without the kernel having to execute code that actually uses (allocs/frees) and object. It allows one to make sure that the slab metainformation and the guard values around an object have not been compromised. A single slabcache can be checked by writing a 1 to the "validate" file. i.e. echo 1 >/sys/slab/kmalloc-128/validate or use the slabinfo tool to check all slabs slabinfo -v Error messages will show up in the syslog. Note that validation can only reach slabs that are on a list. This means that we are usually restricted to partial slabs and active slabs unless SLAB_STORE_USER is active which will build a full slab list and allows validation of slabs that are fully in use. Booting with "slub_debug" set will enable SLAB_STORE_USER and then full diagnostic are available. Note that we attempt to push cpu slabs back to the lists when we start the check. If the cpu slab is reactivated before we get to it (another processor grabs it before we get to it) then it cannot be checked. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:43 +08:00
if (!check_object(s, page, p, val))
break;
}
put_map(map);
unlock:
slab_unlock(page);
slub: validation of slabs (metadata and guard zones) This enables validation of slab. Validation means that all objects are checked to see if there are redzone violations, if padding has been overwritten or any pointers have been corrupted. Also checks the consistency of slab counters. Validation enables the detection of metadata corruption without the kernel having to execute code that actually uses (allocs/frees) and object. It allows one to make sure that the slab metainformation and the guard values around an object have not been compromised. A single slabcache can be checked by writing a 1 to the "validate" file. i.e. echo 1 >/sys/slab/kmalloc-128/validate or use the slabinfo tool to check all slabs slabinfo -v Error messages will show up in the syslog. Note that validation can only reach slabs that are on a list. This means that we are usually restricted to partial slabs and active slabs unless SLAB_STORE_USER is active which will build a full slab list and allows validation of slabs that are fully in use. Booting with "slub_debug" set will enable SLAB_STORE_USER and then full diagnostic are available. Note that we attempt to push cpu slabs back to the lists when we start the check. If the cpu slab is reactivated before we get to it (another processor grabs it before we get to it) then it cannot be checked. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:43 +08:00
}
static int validate_slab_node(struct kmem_cache *s,
struct kmem_cache_node *n)
slub: validation of slabs (metadata and guard zones) This enables validation of slab. Validation means that all objects are checked to see if there are redzone violations, if padding has been overwritten or any pointers have been corrupted. Also checks the consistency of slab counters. Validation enables the detection of metadata corruption without the kernel having to execute code that actually uses (allocs/frees) and object. It allows one to make sure that the slab metainformation and the guard values around an object have not been compromised. A single slabcache can be checked by writing a 1 to the "validate" file. i.e. echo 1 >/sys/slab/kmalloc-128/validate or use the slabinfo tool to check all slabs slabinfo -v Error messages will show up in the syslog. Note that validation can only reach slabs that are on a list. This means that we are usually restricted to partial slabs and active slabs unless SLAB_STORE_USER is active which will build a full slab list and allows validation of slabs that are fully in use. Booting with "slub_debug" set will enable SLAB_STORE_USER and then full diagnostic are available. Note that we attempt to push cpu slabs back to the lists when we start the check. If the cpu slab is reactivated before we get to it (another processor grabs it before we get to it) then it cannot be checked. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:43 +08:00
{
unsigned long count = 0;
struct page *page;
unsigned long flags;
spin_lock_irqsave(&n->list_lock, flags);
list_for_each_entry(page, &n->partial, slab_list) {
validate_slab(s, page);
slub: validation of slabs (metadata and guard zones) This enables validation of slab. Validation means that all objects are checked to see if there are redzone violations, if padding has been overwritten or any pointers have been corrupted. Also checks the consistency of slab counters. Validation enables the detection of metadata corruption without the kernel having to execute code that actually uses (allocs/frees) and object. It allows one to make sure that the slab metainformation and the guard values around an object have not been compromised. A single slabcache can be checked by writing a 1 to the "validate" file. i.e. echo 1 >/sys/slab/kmalloc-128/validate or use the slabinfo tool to check all slabs slabinfo -v Error messages will show up in the syslog. Note that validation can only reach slabs that are on a list. This means that we are usually restricted to partial slabs and active slabs unless SLAB_STORE_USER is active which will build a full slab list and allows validation of slabs that are fully in use. Booting with "slub_debug" set will enable SLAB_STORE_USER and then full diagnostic are available. Note that we attempt to push cpu slabs back to the lists when we start the check. If the cpu slab is reactivated before we get to it (another processor grabs it before we get to it) then it cannot be checked. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:43 +08:00
count++;
}
if (count != n->nr_partial)
pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
s->name, count, n->nr_partial);
slub: validation of slabs (metadata and guard zones) This enables validation of slab. Validation means that all objects are checked to see if there are redzone violations, if padding has been overwritten or any pointers have been corrupted. Also checks the consistency of slab counters. Validation enables the detection of metadata corruption without the kernel having to execute code that actually uses (allocs/frees) and object. It allows one to make sure that the slab metainformation and the guard values around an object have not been compromised. A single slabcache can be checked by writing a 1 to the "validate" file. i.e. echo 1 >/sys/slab/kmalloc-128/validate or use the slabinfo tool to check all slabs slabinfo -v Error messages will show up in the syslog. Note that validation can only reach slabs that are on a list. This means that we are usually restricted to partial slabs and active slabs unless SLAB_STORE_USER is active which will build a full slab list and allows validation of slabs that are fully in use. Booting with "slub_debug" set will enable SLAB_STORE_USER and then full diagnostic are available. Note that we attempt to push cpu slabs back to the lists when we start the check. If the cpu slab is reactivated before we get to it (another processor grabs it before we get to it) then it cannot be checked. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:43 +08:00
if (!(s->flags & SLAB_STORE_USER))
goto out;
list_for_each_entry(page, &n->full, slab_list) {
validate_slab(s, page);
slub: validation of slabs (metadata and guard zones) This enables validation of slab. Validation means that all objects are checked to see if there are redzone violations, if padding has been overwritten or any pointers have been corrupted. Also checks the consistency of slab counters. Validation enables the detection of metadata corruption without the kernel having to execute code that actually uses (allocs/frees) and object. It allows one to make sure that the slab metainformation and the guard values around an object have not been compromised. A single slabcache can be checked by writing a 1 to the "validate" file. i.e. echo 1 >/sys/slab/kmalloc-128/validate or use the slabinfo tool to check all slabs slabinfo -v Error messages will show up in the syslog. Note that validation can only reach slabs that are on a list. This means that we are usually restricted to partial slabs and active slabs unless SLAB_STORE_USER is active which will build a full slab list and allows validation of slabs that are fully in use. Booting with "slub_debug" set will enable SLAB_STORE_USER and then full diagnostic are available. Note that we attempt to push cpu slabs back to the lists when we start the check. If the cpu slab is reactivated before we get to it (another processor grabs it before we get to it) then it cannot be checked. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:43 +08:00
count++;
}
if (count != atomic_long_read(&n->nr_slabs))
pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
s->name, count, atomic_long_read(&n->nr_slabs));
slub: validation of slabs (metadata and guard zones) This enables validation of slab. Validation means that all objects are checked to see if there are redzone violations, if padding has been overwritten or any pointers have been corrupted. Also checks the consistency of slab counters. Validation enables the detection of metadata corruption without the kernel having to execute code that actually uses (allocs/frees) and object. It allows one to make sure that the slab metainformation and the guard values around an object have not been compromised. A single slabcache can be checked by writing a 1 to the "validate" file. i.e. echo 1 >/sys/slab/kmalloc-128/validate or use the slabinfo tool to check all slabs slabinfo -v Error messages will show up in the syslog. Note that validation can only reach slabs that are on a list. This means that we are usually restricted to partial slabs and active slabs unless SLAB_STORE_USER is active which will build a full slab list and allows validation of slabs that are fully in use. Booting with "slub_debug" set will enable SLAB_STORE_USER and then full diagnostic are available. Note that we attempt to push cpu slabs back to the lists when we start the check. If the cpu slab is reactivated before we get to it (another processor grabs it before we get to it) then it cannot be checked. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:43 +08:00
out:
spin_unlock_irqrestore(&n->list_lock, flags);
return count;
}
static long validate_slab_cache(struct kmem_cache *s)
slub: validation of slabs (metadata and guard zones) This enables validation of slab. Validation means that all objects are checked to see if there are redzone violations, if padding has been overwritten or any pointers have been corrupted. Also checks the consistency of slab counters. Validation enables the detection of metadata corruption without the kernel having to execute code that actually uses (allocs/frees) and object. It allows one to make sure that the slab metainformation and the guard values around an object have not been compromised. A single slabcache can be checked by writing a 1 to the "validate" file. i.e. echo 1 >/sys/slab/kmalloc-128/validate or use the slabinfo tool to check all slabs slabinfo -v Error messages will show up in the syslog. Note that validation can only reach slabs that are on a list. This means that we are usually restricted to partial slabs and active slabs unless SLAB_STORE_USER is active which will build a full slab list and allows validation of slabs that are fully in use. Booting with "slub_debug" set will enable SLAB_STORE_USER and then full diagnostic are available. Note that we attempt to push cpu slabs back to the lists when we start the check. If the cpu slab is reactivated before we get to it (another processor grabs it before we get to it) then it cannot be checked. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:43 +08:00
{
int node;
unsigned long count = 0;
struct kmem_cache_node *n;
slub: validation of slabs (metadata and guard zones) This enables validation of slab. Validation means that all objects are checked to see if there are redzone violations, if padding has been overwritten or any pointers have been corrupted. Also checks the consistency of slab counters. Validation enables the detection of metadata corruption without the kernel having to execute code that actually uses (allocs/frees) and object. It allows one to make sure that the slab metainformation and the guard values around an object have not been compromised. A single slabcache can be checked by writing a 1 to the "validate" file. i.e. echo 1 >/sys/slab/kmalloc-128/validate or use the slabinfo tool to check all slabs slabinfo -v Error messages will show up in the syslog. Note that validation can only reach slabs that are on a list. This means that we are usually restricted to partial slabs and active slabs unless SLAB_STORE_USER is active which will build a full slab list and allows validation of slabs that are fully in use. Booting with "slub_debug" set will enable SLAB_STORE_USER and then full diagnostic are available. Note that we attempt to push cpu slabs back to the lists when we start the check. If the cpu slab is reactivated before we get to it (another processor grabs it before we get to it) then it cannot be checked. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:43 +08:00
flush_all(s);
for_each_kmem_cache_node(s, node, n)
count += validate_slab_node(s, n);
slub: validation of slabs (metadata and guard zones) This enables validation of slab. Validation means that all objects are checked to see if there are redzone violations, if padding has been overwritten or any pointers have been corrupted. Also checks the consistency of slab counters. Validation enables the detection of metadata corruption without the kernel having to execute code that actually uses (allocs/frees) and object. It allows one to make sure that the slab metainformation and the guard values around an object have not been compromised. A single slabcache can be checked by writing a 1 to the "validate" file. i.e. echo 1 >/sys/slab/kmalloc-128/validate or use the slabinfo tool to check all slabs slabinfo -v Error messages will show up in the syslog. Note that validation can only reach slabs that are on a list. This means that we are usually restricted to partial slabs and active slabs unless SLAB_STORE_USER is active which will build a full slab list and allows validation of slabs that are fully in use. Booting with "slub_debug" set will enable SLAB_STORE_USER and then full diagnostic are available. Note that we attempt to push cpu slabs back to the lists when we start the check. If the cpu slab is reactivated before we get to it (another processor grabs it before we get to it) then it cannot be checked. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:43 +08:00
return count;
}
slub: add ability to list alloc / free callers per slab This patch enables listing the callers who allocated or freed objects in a cache. For example to list the allocators for kmalloc-128 do cat /sys/slab/kmalloc-128/alloc_calls 7 sn_io_slot_fixup+0x40/0x700 7 sn_io_slot_fixup+0x80/0x700 9 sn_bus_fixup+0xe0/0x380 6 param_sysfs_setup+0xf0/0x280 276 percpu_populate+0xf0/0x1a0 19 __register_chrdev_region+0x30/0x360 8 expand_files+0x2e0/0x6e0 1 sys_epoll_create+0x60/0x200 1 __mounts_open+0x140/0x2c0 65 kmem_alloc+0x110/0x280 3 alloc_disk_node+0xe0/0x200 33 as_get_io_context+0x90/0x280 74 kobject_kset_add_dir+0x40/0x140 12 pci_create_bus+0x2a0/0x5c0 1 acpi_ev_create_gpe_block+0x120/0x9e0 41 con_insert_unipair+0x100/0x1c0 1 uart_open+0x1c0/0xba0 1 dma_pool_create+0xe0/0x340 2 neigh_table_init_no_netlink+0x260/0x4c0 6 neigh_parms_alloc+0x30/0x200 1 netlink_kernel_create+0x130/0x320 5 fz_hash_alloc+0x50/0xe0 2 sn_common_hubdev_init+0xd0/0x6e0 28 kernel_param_sysfs_setup+0x30/0x180 72 process_zones+0x70/0x2e0 cat /sys/slab/kmalloc-128/free_calls 558 <not-available> 3 sn_io_slot_fixup+0x600/0x700 84 free_fdtable_rcu+0x120/0x260 2 seq_release+0x40/0x60 6 kmem_free+0x70/0xc0 24 free_as_io_context+0x20/0x200 1 acpi_get_object_info+0x3a0/0x3e0 1 acpi_add_single_object+0xcf0/0x1e40 2 con_release_unimap+0x80/0x140 1 free+0x20/0x40 SLAB_STORE_USER must be enabled for a slab cache by either booting with "slab_debug" or enabling user tracking specifically for the slab of interest. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:45 +08:00
/*
* Generate lists of code addresses where slabcache objects are allocated
slub: add ability to list alloc / free callers per slab This patch enables listing the callers who allocated or freed objects in a cache. For example to list the allocators for kmalloc-128 do cat /sys/slab/kmalloc-128/alloc_calls 7 sn_io_slot_fixup+0x40/0x700 7 sn_io_slot_fixup+0x80/0x700 9 sn_bus_fixup+0xe0/0x380 6 param_sysfs_setup+0xf0/0x280 276 percpu_populate+0xf0/0x1a0 19 __register_chrdev_region+0x30/0x360 8 expand_files+0x2e0/0x6e0 1 sys_epoll_create+0x60/0x200 1 __mounts_open+0x140/0x2c0 65 kmem_alloc+0x110/0x280 3 alloc_disk_node+0xe0/0x200 33 as_get_io_context+0x90/0x280 74 kobject_kset_add_dir+0x40/0x140 12 pci_create_bus+0x2a0/0x5c0 1 acpi_ev_create_gpe_block+0x120/0x9e0 41 con_insert_unipair+0x100/0x1c0 1 uart_open+0x1c0/0xba0 1 dma_pool_create+0xe0/0x340 2 neigh_table_init_no_netlink+0x260/0x4c0 6 neigh_parms_alloc+0x30/0x200 1 netlink_kernel_create+0x130/0x320 5 fz_hash_alloc+0x50/0xe0 2 sn_common_hubdev_init+0xd0/0x6e0 28 kernel_param_sysfs_setup+0x30/0x180 72 process_zones+0x70/0x2e0 cat /sys/slab/kmalloc-128/free_calls 558 <not-available> 3 sn_io_slot_fixup+0x600/0x700 84 free_fdtable_rcu+0x120/0x260 2 seq_release+0x40/0x60 6 kmem_free+0x70/0xc0 24 free_as_io_context+0x20/0x200 1 acpi_get_object_info+0x3a0/0x3e0 1 acpi_add_single_object+0xcf0/0x1e40 2 con_release_unimap+0x80/0x140 1 free+0x20/0x40 SLAB_STORE_USER must be enabled for a slab cache by either booting with "slab_debug" or enabling user tracking specifically for the slab of interest. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:45 +08:00
* and freed.
*/
struct location {
unsigned long count;
unsigned long addr;
SLUB: include lifetime stats and sets of cpus / nodes in tracking output We have information about how long an object existed and about the nodes and cpus where the allocations and frees took place. Add that information to the tracking output in /sys/slab/xx/alloc_calls and /sys/slab/free_calls This will then enable slabinfo to output nice reports like this: christoph@qirst:~/slub$ ./slabinfo kmalloc-128 Slabcache: kmalloc-128 Aliases: 0 Order : 0 Sizes (bytes) Slabs Debug Memory ------------------------------------------------------------------------ Object : 128 Total : 12 Sanity Checks : On Total: 49152 SlabObj: 200 Full : 7 Redzoning : On Used : 24832 SlabSiz: 4096 Partial: 4 Poisoning : On Loss : 24320 Loss : 72 CpuSlab: 1 Tracking : On Lalig: 13968 Align : 8 Objects: 20 Tracing : Off Lpadd: 1152 kmalloc-128 has no kmem_cache operations kmalloc-128: Kernel object allocation ----------------------------------------------------------------------- 6 param_sysfs_setup+0x71/0x130 age=284512/284512/284512 pid=1 nodes=0-1,3 11 percpu_populate+0x39/0x80 age=283914/284428/284512 pid=1 nodes=0 21 __register_chrdev_region+0x31/0x170 age=282896/284347/284473 pid=1-1705 nodes=0-2 1 sys_inotify_init+0x76/0x1c0 age=283423 pid=1004 nodes=0 19 as_get_io_context+0x32/0xd0 age=6/247567/283988 pid=1-11782 nodes=0,2 10 ida_pre_get+0x4a/0x80 age=277666/283773/284526 pid=0-2177 nodes=0,2 24 kobject_kset_add_dir+0x37/0xb0 age=282727/283860/284472 pid=1-1723 nodes=0-2 1 acpi_ds_build_internal_buffer_obj+0xd3/0x11d age=284508 pid=1 nodes=0 24 con_insert_unipair+0xd7/0x110 age=284438/284438/284438 pid=1 nodes=0,2 1 uart_open+0x2d2/0x4b0 age=283896 pid=1 nodes=0 26 dma_pool_create+0x73/0x1a0 age=282762/282833/282916 pid=1705-1723 nodes=0 1 neigh_table_init_no_netlink+0xd2/0x210 age=284461 pid=1 nodes=0 2 neigh_parms_alloc+0x2b/0xe0 age=284410/284411/284412 pid=1 nodes=2 2 neigh_resolve_output+0x1e1/0x280 age=276289/276291/276293 pid=0-2443 nodes=0 1 netlink_kernel_create+0x90/0x170 age=284472 pid=1 nodes=0 4 xt_alloc_table_info+0x39/0xf0 age=283958/283958/283959 pid=1 nodes=1 3 fn_hash_insert+0x473/0x720 age=277653/277661/277666 pid=2177-2185 nodes=0 1 get_mtrr_state+0x285/0x2a0 age=284526 pid=0 nodes=0 1 cacheinfo_cpu_callback+0x26d/0x3e0 age=284458 pid=1 nodes=0 29 kernel_param_sysfs_setup+0x25/0x90 age=284511/284511/284512 pid=1 nodes=0-1,3 5 process_zones+0x5e/0x170 age=284546/284546/284546 pid=0 nodes=0 1 drm_core_init+0x48/0x160 age=284421 pid=1 nodes=2 kmalloc-128: Kernel object freeing ------------------------------------------------------------------------ 163 <not-available> age=4295176847 pid=0 nodes=0-3 1 __vunmap+0x6e/0xf0 age=282907 pid=1723 nodes=0 28 free_as_io_context+0x12/0x90 age=9243/262197/283474 pid=42-11754 nodes=0 1 acpi_get_object_info+0x1b7/0x1d4 age=284475 pid=1 nodes=0 1 do_acpi_find_child+0x45/0x4e age=284475 pid=1 nodes=0 NUMA nodes : 0 1 2 3 ------------------------------------------ All slabs 7 2 2 1 Partial slabs 2 2 0 0 Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 17:32:45 +08:00
long long sum_time;
long min_time;
long max_time;
long min_pid;
long max_pid;
DECLARE_BITMAP(cpus, NR_CPUS);
SLUB: include lifetime stats and sets of cpus / nodes in tracking output We have information about how long an object existed and about the nodes and cpus where the allocations and frees took place. Add that information to the tracking output in /sys/slab/xx/alloc_calls and /sys/slab/free_calls This will then enable slabinfo to output nice reports like this: christoph@qirst:~/slub$ ./slabinfo kmalloc-128 Slabcache: kmalloc-128 Aliases: 0 Order : 0 Sizes (bytes) Slabs Debug Memory ------------------------------------------------------------------------ Object : 128 Total : 12 Sanity Checks : On Total: 49152 SlabObj: 200 Full : 7 Redzoning : On Used : 24832 SlabSiz: 4096 Partial: 4 Poisoning : On Loss : 24320 Loss : 72 CpuSlab: 1 Tracking : On Lalig: 13968 Align : 8 Objects: 20 Tracing : Off Lpadd: 1152 kmalloc-128 has no kmem_cache operations kmalloc-128: Kernel object allocation ----------------------------------------------------------------------- 6 param_sysfs_setup+0x71/0x130 age=284512/284512/284512 pid=1 nodes=0-1,3 11 percpu_populate+0x39/0x80 age=283914/284428/284512 pid=1 nodes=0 21 __register_chrdev_region+0x31/0x170 age=282896/284347/284473 pid=1-1705 nodes=0-2 1 sys_inotify_init+0x76/0x1c0 age=283423 pid=1004 nodes=0 19 as_get_io_context+0x32/0xd0 age=6/247567/283988 pid=1-11782 nodes=0,2 10 ida_pre_get+0x4a/0x80 age=277666/283773/284526 pid=0-2177 nodes=0,2 24 kobject_kset_add_dir+0x37/0xb0 age=282727/283860/284472 pid=1-1723 nodes=0-2 1 acpi_ds_build_internal_buffer_obj+0xd3/0x11d age=284508 pid=1 nodes=0 24 con_insert_unipair+0xd7/0x110 age=284438/284438/284438 pid=1 nodes=0,2 1 uart_open+0x2d2/0x4b0 age=283896 pid=1 nodes=0 26 dma_pool_create+0x73/0x1a0 age=282762/282833/282916 pid=1705-1723 nodes=0 1 neigh_table_init_no_netlink+0xd2/0x210 age=284461 pid=1 nodes=0 2 neigh_parms_alloc+0x2b/0xe0 age=284410/284411/284412 pid=1 nodes=2 2 neigh_resolve_output+0x1e1/0x280 age=276289/276291/276293 pid=0-2443 nodes=0 1 netlink_kernel_create+0x90/0x170 age=284472 pid=1 nodes=0 4 xt_alloc_table_info+0x39/0xf0 age=283958/283958/283959 pid=1 nodes=1 3 fn_hash_insert+0x473/0x720 age=277653/277661/277666 pid=2177-2185 nodes=0 1 get_mtrr_state+0x285/0x2a0 age=284526 pid=0 nodes=0 1 cacheinfo_cpu_callback+0x26d/0x3e0 age=284458 pid=1 nodes=0 29 kernel_param_sysfs_setup+0x25/0x90 age=284511/284511/284512 pid=1 nodes=0-1,3 5 process_zones+0x5e/0x170 age=284546/284546/284546 pid=0 nodes=0 1 drm_core_init+0x48/0x160 age=284421 pid=1 nodes=2 kmalloc-128: Kernel object freeing ------------------------------------------------------------------------ 163 <not-available> age=4295176847 pid=0 nodes=0-3 1 __vunmap+0x6e/0xf0 age=282907 pid=1723 nodes=0 28 free_as_io_context+0x12/0x90 age=9243/262197/283474 pid=42-11754 nodes=0 1 acpi_get_object_info+0x1b7/0x1d4 age=284475 pid=1 nodes=0 1 do_acpi_find_child+0x45/0x4e age=284475 pid=1 nodes=0 NUMA nodes : 0 1 2 3 ------------------------------------------ All slabs 7 2 2 1 Partial slabs 2 2 0 0 Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 17:32:45 +08:00
nodemask_t nodes;
slub: add ability to list alloc / free callers per slab This patch enables listing the callers who allocated or freed objects in a cache. For example to list the allocators for kmalloc-128 do cat /sys/slab/kmalloc-128/alloc_calls 7 sn_io_slot_fixup+0x40/0x700 7 sn_io_slot_fixup+0x80/0x700 9 sn_bus_fixup+0xe0/0x380 6 param_sysfs_setup+0xf0/0x280 276 percpu_populate+0xf0/0x1a0 19 __register_chrdev_region+0x30/0x360 8 expand_files+0x2e0/0x6e0 1 sys_epoll_create+0x60/0x200 1 __mounts_open+0x140/0x2c0 65 kmem_alloc+0x110/0x280 3 alloc_disk_node+0xe0/0x200 33 as_get_io_context+0x90/0x280 74 kobject_kset_add_dir+0x40/0x140 12 pci_create_bus+0x2a0/0x5c0 1 acpi_ev_create_gpe_block+0x120/0x9e0 41 con_insert_unipair+0x100/0x1c0 1 uart_open+0x1c0/0xba0 1 dma_pool_create+0xe0/0x340 2 neigh_table_init_no_netlink+0x260/0x4c0 6 neigh_parms_alloc+0x30/0x200 1 netlink_kernel_create+0x130/0x320 5 fz_hash_alloc+0x50/0xe0 2 sn_common_hubdev_init+0xd0/0x6e0 28 kernel_param_sysfs_setup+0x30/0x180 72 process_zones+0x70/0x2e0 cat /sys/slab/kmalloc-128/free_calls 558 <not-available> 3 sn_io_slot_fixup+0x600/0x700 84 free_fdtable_rcu+0x120/0x260 2 seq_release+0x40/0x60 6 kmem_free+0x70/0xc0 24 free_as_io_context+0x20/0x200 1 acpi_get_object_info+0x3a0/0x3e0 1 acpi_add_single_object+0xcf0/0x1e40 2 con_release_unimap+0x80/0x140 1 free+0x20/0x40 SLAB_STORE_USER must be enabled for a slab cache by either booting with "slab_debug" or enabling user tracking specifically for the slab of interest. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:45 +08:00
};
struct loc_track {
unsigned long max;
unsigned long count;
struct location *loc;
};
static void free_loc_track(struct loc_track *t)
{
if (t->max)
free_pages((unsigned long)t->loc,
get_order(sizeof(struct location) * t->max));
}
static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
slub: add ability to list alloc / free callers per slab This patch enables listing the callers who allocated or freed objects in a cache. For example to list the allocators for kmalloc-128 do cat /sys/slab/kmalloc-128/alloc_calls 7 sn_io_slot_fixup+0x40/0x700 7 sn_io_slot_fixup+0x80/0x700 9 sn_bus_fixup+0xe0/0x380 6 param_sysfs_setup+0xf0/0x280 276 percpu_populate+0xf0/0x1a0 19 __register_chrdev_region+0x30/0x360 8 expand_files+0x2e0/0x6e0 1 sys_epoll_create+0x60/0x200 1 __mounts_open+0x140/0x2c0 65 kmem_alloc+0x110/0x280 3 alloc_disk_node+0xe0/0x200 33 as_get_io_context+0x90/0x280 74 kobject_kset_add_dir+0x40/0x140 12 pci_create_bus+0x2a0/0x5c0 1 acpi_ev_create_gpe_block+0x120/0x9e0 41 con_insert_unipair+0x100/0x1c0 1 uart_open+0x1c0/0xba0 1 dma_pool_create+0xe0/0x340 2 neigh_table_init_no_netlink+0x260/0x4c0 6 neigh_parms_alloc+0x30/0x200 1 netlink_kernel_create+0x130/0x320 5 fz_hash_alloc+0x50/0xe0 2 sn_common_hubdev_init+0xd0/0x6e0 28 kernel_param_sysfs_setup+0x30/0x180 72 process_zones+0x70/0x2e0 cat /sys/slab/kmalloc-128/free_calls 558 <not-available> 3 sn_io_slot_fixup+0x600/0x700 84 free_fdtable_rcu+0x120/0x260 2 seq_release+0x40/0x60 6 kmem_free+0x70/0xc0 24 free_as_io_context+0x20/0x200 1 acpi_get_object_info+0x3a0/0x3e0 1 acpi_add_single_object+0xcf0/0x1e40 2 con_release_unimap+0x80/0x140 1 free+0x20/0x40 SLAB_STORE_USER must be enabled for a slab cache by either booting with "slab_debug" or enabling user tracking specifically for the slab of interest. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:45 +08:00
{
struct location *l;
int order;
order = get_order(sizeof(struct location) * max);
l = (void *)__get_free_pages(flags, order);
slub: add ability to list alloc / free callers per slab This patch enables listing the callers who allocated or freed objects in a cache. For example to list the allocators for kmalloc-128 do cat /sys/slab/kmalloc-128/alloc_calls 7 sn_io_slot_fixup+0x40/0x700 7 sn_io_slot_fixup+0x80/0x700 9 sn_bus_fixup+0xe0/0x380 6 param_sysfs_setup+0xf0/0x280 276 percpu_populate+0xf0/0x1a0 19 __register_chrdev_region+0x30/0x360 8 expand_files+0x2e0/0x6e0 1 sys_epoll_create+0x60/0x200 1 __mounts_open+0x140/0x2c0 65 kmem_alloc+0x110/0x280 3 alloc_disk_node+0xe0/0x200 33 as_get_io_context+0x90/0x280 74 kobject_kset_add_dir+0x40/0x140 12 pci_create_bus+0x2a0/0x5c0 1 acpi_ev_create_gpe_block+0x120/0x9e0 41 con_insert_unipair+0x100/0x1c0 1 uart_open+0x1c0/0xba0 1 dma_pool_create+0xe0/0x340 2 neigh_table_init_no_netlink+0x260/0x4c0 6 neigh_parms_alloc+0x30/0x200 1 netlink_kernel_create+0x130/0x320 5 fz_hash_alloc+0x50/0xe0 2 sn_common_hubdev_init+0xd0/0x6e0 28 kernel_param_sysfs_setup+0x30/0x180 72 process_zones+0x70/0x2e0 cat /sys/slab/kmalloc-128/free_calls 558 <not-available> 3 sn_io_slot_fixup+0x600/0x700 84 free_fdtable_rcu+0x120/0x260 2 seq_release+0x40/0x60 6 kmem_free+0x70/0xc0 24 free_as_io_context+0x20/0x200 1 acpi_get_object_info+0x3a0/0x3e0 1 acpi_add_single_object+0xcf0/0x1e40 2 con_release_unimap+0x80/0x140 1 free+0x20/0x40 SLAB_STORE_USER must be enabled for a slab cache by either booting with "slab_debug" or enabling user tracking specifically for the slab of interest. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:45 +08:00
if (!l)
return 0;
if (t->count) {
memcpy(l, t->loc, sizeof(struct location) * t->count);
free_loc_track(t);
}
t->max = max;
t->loc = l;
return 1;
}
static int add_location(struct loc_track *t, struct kmem_cache *s,
SLUB: include lifetime stats and sets of cpus / nodes in tracking output We have information about how long an object existed and about the nodes and cpus where the allocations and frees took place. Add that information to the tracking output in /sys/slab/xx/alloc_calls and /sys/slab/free_calls This will then enable slabinfo to output nice reports like this: christoph@qirst:~/slub$ ./slabinfo kmalloc-128 Slabcache: kmalloc-128 Aliases: 0 Order : 0 Sizes (bytes) Slabs Debug Memory ------------------------------------------------------------------------ Object : 128 Total : 12 Sanity Checks : On Total: 49152 SlabObj: 200 Full : 7 Redzoning : On Used : 24832 SlabSiz: 4096 Partial: 4 Poisoning : On Loss : 24320 Loss : 72 CpuSlab: 1 Tracking : On Lalig: 13968 Align : 8 Objects: 20 Tracing : Off Lpadd: 1152 kmalloc-128 has no kmem_cache operations kmalloc-128: Kernel object allocation ----------------------------------------------------------------------- 6 param_sysfs_setup+0x71/0x130 age=284512/284512/284512 pid=1 nodes=0-1,3 11 percpu_populate+0x39/0x80 age=283914/284428/284512 pid=1 nodes=0 21 __register_chrdev_region+0x31/0x170 age=282896/284347/284473 pid=1-1705 nodes=0-2 1 sys_inotify_init+0x76/0x1c0 age=283423 pid=1004 nodes=0 19 as_get_io_context+0x32/0xd0 age=6/247567/283988 pid=1-11782 nodes=0,2 10 ida_pre_get+0x4a/0x80 age=277666/283773/284526 pid=0-2177 nodes=0,2 24 kobject_kset_add_dir+0x37/0xb0 age=282727/283860/284472 pid=1-1723 nodes=0-2 1 acpi_ds_build_internal_buffer_obj+0xd3/0x11d age=284508 pid=1 nodes=0 24 con_insert_unipair+0xd7/0x110 age=284438/284438/284438 pid=1 nodes=0,2 1 uart_open+0x2d2/0x4b0 age=283896 pid=1 nodes=0 26 dma_pool_create+0x73/0x1a0 age=282762/282833/282916 pid=1705-1723 nodes=0 1 neigh_table_init_no_netlink+0xd2/0x210 age=284461 pid=1 nodes=0 2 neigh_parms_alloc+0x2b/0xe0 age=284410/284411/284412 pid=1 nodes=2 2 neigh_resolve_output+0x1e1/0x280 age=276289/276291/276293 pid=0-2443 nodes=0 1 netlink_kernel_create+0x90/0x170 age=284472 pid=1 nodes=0 4 xt_alloc_table_info+0x39/0xf0 age=283958/283958/283959 pid=1 nodes=1 3 fn_hash_insert+0x473/0x720 age=277653/277661/277666 pid=2177-2185 nodes=0 1 get_mtrr_state+0x285/0x2a0 age=284526 pid=0 nodes=0 1 cacheinfo_cpu_callback+0x26d/0x3e0 age=284458 pid=1 nodes=0 29 kernel_param_sysfs_setup+0x25/0x90 age=284511/284511/284512 pid=1 nodes=0-1,3 5 process_zones+0x5e/0x170 age=284546/284546/284546 pid=0 nodes=0 1 drm_core_init+0x48/0x160 age=284421 pid=1 nodes=2 kmalloc-128: Kernel object freeing ------------------------------------------------------------------------ 163 <not-available> age=4295176847 pid=0 nodes=0-3 1 __vunmap+0x6e/0xf0 age=282907 pid=1723 nodes=0 28 free_as_io_context+0x12/0x90 age=9243/262197/283474 pid=42-11754 nodes=0 1 acpi_get_object_info+0x1b7/0x1d4 age=284475 pid=1 nodes=0 1 do_acpi_find_child+0x45/0x4e age=284475 pid=1 nodes=0 NUMA nodes : 0 1 2 3 ------------------------------------------ All slabs 7 2 2 1 Partial slabs 2 2 0 0 Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 17:32:45 +08:00
const struct track *track)
slub: add ability to list alloc / free callers per slab This patch enables listing the callers who allocated or freed objects in a cache. For example to list the allocators for kmalloc-128 do cat /sys/slab/kmalloc-128/alloc_calls 7 sn_io_slot_fixup+0x40/0x700 7 sn_io_slot_fixup+0x80/0x700 9 sn_bus_fixup+0xe0/0x380 6 param_sysfs_setup+0xf0/0x280 276 percpu_populate+0xf0/0x1a0 19 __register_chrdev_region+0x30/0x360 8 expand_files+0x2e0/0x6e0 1 sys_epoll_create+0x60/0x200 1 __mounts_open+0x140/0x2c0 65 kmem_alloc+0x110/0x280 3 alloc_disk_node+0xe0/0x200 33 as_get_io_context+0x90/0x280 74 kobject_kset_add_dir+0x40/0x140 12 pci_create_bus+0x2a0/0x5c0 1 acpi_ev_create_gpe_block+0x120/0x9e0 41 con_insert_unipair+0x100/0x1c0 1 uart_open+0x1c0/0xba0 1 dma_pool_create+0xe0/0x340 2 neigh_table_init_no_netlink+0x260/0x4c0 6 neigh_parms_alloc+0x30/0x200 1 netlink_kernel_create+0x130/0x320 5 fz_hash_alloc+0x50/0xe0 2 sn_common_hubdev_init+0xd0/0x6e0 28 kernel_param_sysfs_setup+0x30/0x180 72 process_zones+0x70/0x2e0 cat /sys/slab/kmalloc-128/free_calls 558 <not-available> 3 sn_io_slot_fixup+0x600/0x700 84 free_fdtable_rcu+0x120/0x260 2 seq_release+0x40/0x60 6 kmem_free+0x70/0xc0 24 free_as_io_context+0x20/0x200 1 acpi_get_object_info+0x3a0/0x3e0 1 acpi_add_single_object+0xcf0/0x1e40 2 con_release_unimap+0x80/0x140 1 free+0x20/0x40 SLAB_STORE_USER must be enabled for a slab cache by either booting with "slab_debug" or enabling user tracking specifically for the slab of interest. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:45 +08:00
{
long start, end, pos;
struct location *l;
unsigned long caddr;
SLUB: include lifetime stats and sets of cpus / nodes in tracking output We have information about how long an object existed and about the nodes and cpus where the allocations and frees took place. Add that information to the tracking output in /sys/slab/xx/alloc_calls and /sys/slab/free_calls This will then enable slabinfo to output nice reports like this: christoph@qirst:~/slub$ ./slabinfo kmalloc-128 Slabcache: kmalloc-128 Aliases: 0 Order : 0 Sizes (bytes) Slabs Debug Memory ------------------------------------------------------------------------ Object : 128 Total : 12 Sanity Checks : On Total: 49152 SlabObj: 200 Full : 7 Redzoning : On Used : 24832 SlabSiz: 4096 Partial: 4 Poisoning : On Loss : 24320 Loss : 72 CpuSlab: 1 Tracking : On Lalig: 13968 Align : 8 Objects: 20 Tracing : Off Lpadd: 1152 kmalloc-128 has no kmem_cache operations kmalloc-128: Kernel object allocation ----------------------------------------------------------------------- 6 param_sysfs_setup+0x71/0x130 age=284512/284512/284512 pid=1 nodes=0-1,3 11 percpu_populate+0x39/0x80 age=283914/284428/284512 pid=1 nodes=0 21 __register_chrdev_region+0x31/0x170 age=282896/284347/284473 pid=1-1705 nodes=0-2 1 sys_inotify_init+0x76/0x1c0 age=283423 pid=1004 nodes=0 19 as_get_io_context+0x32/0xd0 age=6/247567/283988 pid=1-11782 nodes=0,2 10 ida_pre_get+0x4a/0x80 age=277666/283773/284526 pid=0-2177 nodes=0,2 24 kobject_kset_add_dir+0x37/0xb0 age=282727/283860/284472 pid=1-1723 nodes=0-2 1 acpi_ds_build_internal_buffer_obj+0xd3/0x11d age=284508 pid=1 nodes=0 24 con_insert_unipair+0xd7/0x110 age=284438/284438/284438 pid=1 nodes=0,2 1 uart_open+0x2d2/0x4b0 age=283896 pid=1 nodes=0 26 dma_pool_create+0x73/0x1a0 age=282762/282833/282916 pid=1705-1723 nodes=0 1 neigh_table_init_no_netlink+0xd2/0x210 age=284461 pid=1 nodes=0 2 neigh_parms_alloc+0x2b/0xe0 age=284410/284411/284412 pid=1 nodes=2 2 neigh_resolve_output+0x1e1/0x280 age=276289/276291/276293 pid=0-2443 nodes=0 1 netlink_kernel_create+0x90/0x170 age=284472 pid=1 nodes=0 4 xt_alloc_table_info+0x39/0xf0 age=283958/283958/283959 pid=1 nodes=1 3 fn_hash_insert+0x473/0x720 age=277653/277661/277666 pid=2177-2185 nodes=0 1 get_mtrr_state+0x285/0x2a0 age=284526 pid=0 nodes=0 1 cacheinfo_cpu_callback+0x26d/0x3e0 age=284458 pid=1 nodes=0 29 kernel_param_sysfs_setup+0x25/0x90 age=284511/284511/284512 pid=1 nodes=0-1,3 5 process_zones+0x5e/0x170 age=284546/284546/284546 pid=0 nodes=0 1 drm_core_init+0x48/0x160 age=284421 pid=1 nodes=2 kmalloc-128: Kernel object freeing ------------------------------------------------------------------------ 163 <not-available> age=4295176847 pid=0 nodes=0-3 1 __vunmap+0x6e/0xf0 age=282907 pid=1723 nodes=0 28 free_as_io_context+0x12/0x90 age=9243/262197/283474 pid=42-11754 nodes=0 1 acpi_get_object_info+0x1b7/0x1d4 age=284475 pid=1 nodes=0 1 do_acpi_find_child+0x45/0x4e age=284475 pid=1 nodes=0 NUMA nodes : 0 1 2 3 ------------------------------------------ All slabs 7 2 2 1 Partial slabs 2 2 0 0 Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 17:32:45 +08:00
unsigned long age = jiffies - track->when;
slub: add ability to list alloc / free callers per slab This patch enables listing the callers who allocated or freed objects in a cache. For example to list the allocators for kmalloc-128 do cat /sys/slab/kmalloc-128/alloc_calls 7 sn_io_slot_fixup+0x40/0x700 7 sn_io_slot_fixup+0x80/0x700 9 sn_bus_fixup+0xe0/0x380 6 param_sysfs_setup+0xf0/0x280 276 percpu_populate+0xf0/0x1a0 19 __register_chrdev_region+0x30/0x360 8 expand_files+0x2e0/0x6e0 1 sys_epoll_create+0x60/0x200 1 __mounts_open+0x140/0x2c0 65 kmem_alloc+0x110/0x280 3 alloc_disk_node+0xe0/0x200 33 as_get_io_context+0x90/0x280 74 kobject_kset_add_dir+0x40/0x140 12 pci_create_bus+0x2a0/0x5c0 1 acpi_ev_create_gpe_block+0x120/0x9e0 41 con_insert_unipair+0x100/0x1c0 1 uart_open+0x1c0/0xba0 1 dma_pool_create+0xe0/0x340 2 neigh_table_init_no_netlink+0x260/0x4c0 6 neigh_parms_alloc+0x30/0x200 1 netlink_kernel_create+0x130/0x320 5 fz_hash_alloc+0x50/0xe0 2 sn_common_hubdev_init+0xd0/0x6e0 28 kernel_param_sysfs_setup+0x30/0x180 72 process_zones+0x70/0x2e0 cat /sys/slab/kmalloc-128/free_calls 558 <not-available> 3 sn_io_slot_fixup+0x600/0x700 84 free_fdtable_rcu+0x120/0x260 2 seq_release+0x40/0x60 6 kmem_free+0x70/0xc0 24 free_as_io_context+0x20/0x200 1 acpi_get_object_info+0x3a0/0x3e0 1 acpi_add_single_object+0xcf0/0x1e40 2 con_release_unimap+0x80/0x140 1 free+0x20/0x40 SLAB_STORE_USER must be enabled for a slab cache by either booting with "slab_debug" or enabling user tracking specifically for the slab of interest. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:45 +08:00
start = -1;
end = t->count;
for ( ; ; ) {
pos = start + (end - start + 1) / 2;
/*
* There is nothing at "end". If we end up there
* we need to add something to before end.
*/
if (pos == end)
break;
caddr = t->loc[pos].addr;
SLUB: include lifetime stats and sets of cpus / nodes in tracking output We have information about how long an object existed and about the nodes and cpus where the allocations and frees took place. Add that information to the tracking output in /sys/slab/xx/alloc_calls and /sys/slab/free_calls This will then enable slabinfo to output nice reports like this: christoph@qirst:~/slub$ ./slabinfo kmalloc-128 Slabcache: kmalloc-128 Aliases: 0 Order : 0 Sizes (bytes) Slabs Debug Memory ------------------------------------------------------------------------ Object : 128 Total : 12 Sanity Checks : On Total: 49152 SlabObj: 200 Full : 7 Redzoning : On Used : 24832 SlabSiz: 4096 Partial: 4 Poisoning : On Loss : 24320 Loss : 72 CpuSlab: 1 Tracking : On Lalig: 13968 Align : 8 Objects: 20 Tracing : Off Lpadd: 1152 kmalloc-128 has no kmem_cache operations kmalloc-128: Kernel object allocation ----------------------------------------------------------------------- 6 param_sysfs_setup+0x71/0x130 age=284512/284512/284512 pid=1 nodes=0-1,3 11 percpu_populate+0x39/0x80 age=283914/284428/284512 pid=1 nodes=0 21 __register_chrdev_region+0x31/0x170 age=282896/284347/284473 pid=1-1705 nodes=0-2 1 sys_inotify_init+0x76/0x1c0 age=283423 pid=1004 nodes=0 19 as_get_io_context+0x32/0xd0 age=6/247567/283988 pid=1-11782 nodes=0,2 10 ida_pre_get+0x4a/0x80 age=277666/283773/284526 pid=0-2177 nodes=0,2 24 kobject_kset_add_dir+0x37/0xb0 age=282727/283860/284472 pid=1-1723 nodes=0-2 1 acpi_ds_build_internal_buffer_obj+0xd3/0x11d age=284508 pid=1 nodes=0 24 con_insert_unipair+0xd7/0x110 age=284438/284438/284438 pid=1 nodes=0,2 1 uart_open+0x2d2/0x4b0 age=283896 pid=1 nodes=0 26 dma_pool_create+0x73/0x1a0 age=282762/282833/282916 pid=1705-1723 nodes=0 1 neigh_table_init_no_netlink+0xd2/0x210 age=284461 pid=1 nodes=0 2 neigh_parms_alloc+0x2b/0xe0 age=284410/284411/284412 pid=1 nodes=2 2 neigh_resolve_output+0x1e1/0x280 age=276289/276291/276293 pid=0-2443 nodes=0 1 netlink_kernel_create+0x90/0x170 age=284472 pid=1 nodes=0 4 xt_alloc_table_info+0x39/0xf0 age=283958/283958/283959 pid=1 nodes=1 3 fn_hash_insert+0x473/0x720 age=277653/277661/277666 pid=2177-2185 nodes=0 1 get_mtrr_state+0x285/0x2a0 age=284526 pid=0 nodes=0 1 cacheinfo_cpu_callback+0x26d/0x3e0 age=284458 pid=1 nodes=0 29 kernel_param_sysfs_setup+0x25/0x90 age=284511/284511/284512 pid=1 nodes=0-1,3 5 process_zones+0x5e/0x170 age=284546/284546/284546 pid=0 nodes=0 1 drm_core_init+0x48/0x160 age=284421 pid=1 nodes=2 kmalloc-128: Kernel object freeing ------------------------------------------------------------------------ 163 <not-available> age=4295176847 pid=0 nodes=0-3 1 __vunmap+0x6e/0xf0 age=282907 pid=1723 nodes=0 28 free_as_io_context+0x12/0x90 age=9243/262197/283474 pid=42-11754 nodes=0 1 acpi_get_object_info+0x1b7/0x1d4 age=284475 pid=1 nodes=0 1 do_acpi_find_child+0x45/0x4e age=284475 pid=1 nodes=0 NUMA nodes : 0 1 2 3 ------------------------------------------ All slabs 7 2 2 1 Partial slabs 2 2 0 0 Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 17:32:45 +08:00
if (track->addr == caddr) {
l = &t->loc[pos];
l->count++;
if (track->when) {
l->sum_time += age;
if (age < l->min_time)
l->min_time = age;
if (age > l->max_time)
l->max_time = age;
if (track->pid < l->min_pid)
l->min_pid = track->pid;
if (track->pid > l->max_pid)
l->max_pid = track->pid;
cpumask_set_cpu(track->cpu,
to_cpumask(l->cpus));
SLUB: include lifetime stats and sets of cpus / nodes in tracking output We have information about how long an object existed and about the nodes and cpus where the allocations and frees took place. Add that information to the tracking output in /sys/slab/xx/alloc_calls and /sys/slab/free_calls This will then enable slabinfo to output nice reports like this: christoph@qirst:~/slub$ ./slabinfo kmalloc-128 Slabcache: kmalloc-128 Aliases: 0 Order : 0 Sizes (bytes) Slabs Debug Memory ------------------------------------------------------------------------ Object : 128 Total : 12 Sanity Checks : On Total: 49152 SlabObj: 200 Full : 7 Redzoning : On Used : 24832 SlabSiz: 4096 Partial: 4 Poisoning : On Loss : 24320 Loss : 72 CpuSlab: 1 Tracking : On Lalig: 13968 Align : 8 Objects: 20 Tracing : Off Lpadd: 1152 kmalloc-128 has no kmem_cache operations kmalloc-128: Kernel object allocation ----------------------------------------------------------------------- 6 param_sysfs_setup+0x71/0x130 age=284512/284512/284512 pid=1 nodes=0-1,3 11 percpu_populate+0x39/0x80 age=283914/284428/284512 pid=1 nodes=0 21 __register_chrdev_region+0x31/0x170 age=282896/284347/284473 pid=1-1705 nodes=0-2 1 sys_inotify_init+0x76/0x1c0 age=283423 pid=1004 nodes=0 19 as_get_io_context+0x32/0xd0 age=6/247567/283988 pid=1-11782 nodes=0,2 10 ida_pre_get+0x4a/0x80 age=277666/283773/284526 pid=0-2177 nodes=0,2 24 kobject_kset_add_dir+0x37/0xb0 age=282727/283860/284472 pid=1-1723 nodes=0-2 1 acpi_ds_build_internal_buffer_obj+0xd3/0x11d age=284508 pid=1 nodes=0 24 con_insert_unipair+0xd7/0x110 age=284438/284438/284438 pid=1 nodes=0,2 1 uart_open+0x2d2/0x4b0 age=283896 pid=1 nodes=0 26 dma_pool_create+0x73/0x1a0 age=282762/282833/282916 pid=1705-1723 nodes=0 1 neigh_table_init_no_netlink+0xd2/0x210 age=284461 pid=1 nodes=0 2 neigh_parms_alloc+0x2b/0xe0 age=284410/284411/284412 pid=1 nodes=2 2 neigh_resolve_output+0x1e1/0x280 age=276289/276291/276293 pid=0-2443 nodes=0 1 netlink_kernel_create+0x90/0x170 age=284472 pid=1 nodes=0 4 xt_alloc_table_info+0x39/0xf0 age=283958/283958/283959 pid=1 nodes=1 3 fn_hash_insert+0x473/0x720 age=277653/277661/277666 pid=2177-2185 nodes=0 1 get_mtrr_state+0x285/0x2a0 age=284526 pid=0 nodes=0 1 cacheinfo_cpu_callback+0x26d/0x3e0 age=284458 pid=1 nodes=0 29 kernel_param_sysfs_setup+0x25/0x90 age=284511/284511/284512 pid=1 nodes=0-1,3 5 process_zones+0x5e/0x170 age=284546/284546/284546 pid=0 nodes=0 1 drm_core_init+0x48/0x160 age=284421 pid=1 nodes=2 kmalloc-128: Kernel object freeing ------------------------------------------------------------------------ 163 <not-available> age=4295176847 pid=0 nodes=0-3 1 __vunmap+0x6e/0xf0 age=282907 pid=1723 nodes=0 28 free_as_io_context+0x12/0x90 age=9243/262197/283474 pid=42-11754 nodes=0 1 acpi_get_object_info+0x1b7/0x1d4 age=284475 pid=1 nodes=0 1 do_acpi_find_child+0x45/0x4e age=284475 pid=1 nodes=0 NUMA nodes : 0 1 2 3 ------------------------------------------ All slabs 7 2 2 1 Partial slabs 2 2 0 0 Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 17:32:45 +08:00
}
node_set(page_to_nid(virt_to_page(track)), l->nodes);
slub: add ability to list alloc / free callers per slab This patch enables listing the callers who allocated or freed objects in a cache. For example to list the allocators for kmalloc-128 do cat /sys/slab/kmalloc-128/alloc_calls 7 sn_io_slot_fixup+0x40/0x700 7 sn_io_slot_fixup+0x80/0x700 9 sn_bus_fixup+0xe0/0x380 6 param_sysfs_setup+0xf0/0x280 276 percpu_populate+0xf0/0x1a0 19 __register_chrdev_region+0x30/0x360 8 expand_files+0x2e0/0x6e0 1 sys_epoll_create+0x60/0x200 1 __mounts_open+0x140/0x2c0 65 kmem_alloc+0x110/0x280 3 alloc_disk_node+0xe0/0x200 33 as_get_io_context+0x90/0x280 74 kobject_kset_add_dir+0x40/0x140 12 pci_create_bus+0x2a0/0x5c0 1 acpi_ev_create_gpe_block+0x120/0x9e0 41 con_insert_unipair+0x100/0x1c0 1 uart_open+0x1c0/0xba0 1 dma_pool_create+0xe0/0x340 2 neigh_table_init_no_netlink+0x260/0x4c0 6 neigh_parms_alloc+0x30/0x200 1 netlink_kernel_create+0x130/0x320 5 fz_hash_alloc+0x50/0xe0 2 sn_common_hubdev_init+0xd0/0x6e0 28 kernel_param_sysfs_setup+0x30/0x180 72 process_zones+0x70/0x2e0 cat /sys/slab/kmalloc-128/free_calls 558 <not-available> 3 sn_io_slot_fixup+0x600/0x700 84 free_fdtable_rcu+0x120/0x260 2 seq_release+0x40/0x60 6 kmem_free+0x70/0xc0 24 free_as_io_context+0x20/0x200 1 acpi_get_object_info+0x3a0/0x3e0 1 acpi_add_single_object+0xcf0/0x1e40 2 con_release_unimap+0x80/0x140 1 free+0x20/0x40 SLAB_STORE_USER must be enabled for a slab cache by either booting with "slab_debug" or enabling user tracking specifically for the slab of interest. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:45 +08:00
return 1;
}
SLUB: include lifetime stats and sets of cpus / nodes in tracking output We have information about how long an object existed and about the nodes and cpus where the allocations and frees took place. Add that information to the tracking output in /sys/slab/xx/alloc_calls and /sys/slab/free_calls This will then enable slabinfo to output nice reports like this: christoph@qirst:~/slub$ ./slabinfo kmalloc-128 Slabcache: kmalloc-128 Aliases: 0 Order : 0 Sizes (bytes) Slabs Debug Memory ------------------------------------------------------------------------ Object : 128 Total : 12 Sanity Checks : On Total: 49152 SlabObj: 200 Full : 7 Redzoning : On Used : 24832 SlabSiz: 4096 Partial: 4 Poisoning : On Loss : 24320 Loss : 72 CpuSlab: 1 Tracking : On Lalig: 13968 Align : 8 Objects: 20 Tracing : Off Lpadd: 1152 kmalloc-128 has no kmem_cache operations kmalloc-128: Kernel object allocation ----------------------------------------------------------------------- 6 param_sysfs_setup+0x71/0x130 age=284512/284512/284512 pid=1 nodes=0-1,3 11 percpu_populate+0x39/0x80 age=283914/284428/284512 pid=1 nodes=0 21 __register_chrdev_region+0x31/0x170 age=282896/284347/284473 pid=1-1705 nodes=0-2 1 sys_inotify_init+0x76/0x1c0 age=283423 pid=1004 nodes=0 19 as_get_io_context+0x32/0xd0 age=6/247567/283988 pid=1-11782 nodes=0,2 10 ida_pre_get+0x4a/0x80 age=277666/283773/284526 pid=0-2177 nodes=0,2 24 kobject_kset_add_dir+0x37/0xb0 age=282727/283860/284472 pid=1-1723 nodes=0-2 1 acpi_ds_build_internal_buffer_obj+0xd3/0x11d age=284508 pid=1 nodes=0 24 con_insert_unipair+0xd7/0x110 age=284438/284438/284438 pid=1 nodes=0,2 1 uart_open+0x2d2/0x4b0 age=283896 pid=1 nodes=0 26 dma_pool_create+0x73/0x1a0 age=282762/282833/282916 pid=1705-1723 nodes=0 1 neigh_table_init_no_netlink+0xd2/0x210 age=284461 pid=1 nodes=0 2 neigh_parms_alloc+0x2b/0xe0 age=284410/284411/284412 pid=1 nodes=2 2 neigh_resolve_output+0x1e1/0x280 age=276289/276291/276293 pid=0-2443 nodes=0 1 netlink_kernel_create+0x90/0x170 age=284472 pid=1 nodes=0 4 xt_alloc_table_info+0x39/0xf0 age=283958/283958/283959 pid=1 nodes=1 3 fn_hash_insert+0x473/0x720 age=277653/277661/277666 pid=2177-2185 nodes=0 1 get_mtrr_state+0x285/0x2a0 age=284526 pid=0 nodes=0 1 cacheinfo_cpu_callback+0x26d/0x3e0 age=284458 pid=1 nodes=0 29 kernel_param_sysfs_setup+0x25/0x90 age=284511/284511/284512 pid=1 nodes=0-1,3 5 process_zones+0x5e/0x170 age=284546/284546/284546 pid=0 nodes=0 1 drm_core_init+0x48/0x160 age=284421 pid=1 nodes=2 kmalloc-128: Kernel object freeing ------------------------------------------------------------------------ 163 <not-available> age=4295176847 pid=0 nodes=0-3 1 __vunmap+0x6e/0xf0 age=282907 pid=1723 nodes=0 28 free_as_io_context+0x12/0x90 age=9243/262197/283474 pid=42-11754 nodes=0 1 acpi_get_object_info+0x1b7/0x1d4 age=284475 pid=1 nodes=0 1 do_acpi_find_child+0x45/0x4e age=284475 pid=1 nodes=0 NUMA nodes : 0 1 2 3 ------------------------------------------ All slabs 7 2 2 1 Partial slabs 2 2 0 0 Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 17:32:45 +08:00
if (track->addr < caddr)
slub: add ability to list alloc / free callers per slab This patch enables listing the callers who allocated or freed objects in a cache. For example to list the allocators for kmalloc-128 do cat /sys/slab/kmalloc-128/alloc_calls 7 sn_io_slot_fixup+0x40/0x700 7 sn_io_slot_fixup+0x80/0x700 9 sn_bus_fixup+0xe0/0x380 6 param_sysfs_setup+0xf0/0x280 276 percpu_populate+0xf0/0x1a0 19 __register_chrdev_region+0x30/0x360 8 expand_files+0x2e0/0x6e0 1 sys_epoll_create+0x60/0x200 1 __mounts_open+0x140/0x2c0 65 kmem_alloc+0x110/0x280 3 alloc_disk_node+0xe0/0x200 33 as_get_io_context+0x90/0x280 74 kobject_kset_add_dir+0x40/0x140 12 pci_create_bus+0x2a0/0x5c0 1 acpi_ev_create_gpe_block+0x120/0x9e0 41 con_insert_unipair+0x100/0x1c0 1 uart_open+0x1c0/0xba0 1 dma_pool_create+0xe0/0x340 2 neigh_table_init_no_netlink+0x260/0x4c0 6 neigh_parms_alloc+0x30/0x200 1 netlink_kernel_create+0x130/0x320 5 fz_hash_alloc+0x50/0xe0 2 sn_common_hubdev_init+0xd0/0x6e0 28 kernel_param_sysfs_setup+0x30/0x180 72 process_zones+0x70/0x2e0 cat /sys/slab/kmalloc-128/free_calls 558 <not-available> 3 sn_io_slot_fixup+0x600/0x700 84 free_fdtable_rcu+0x120/0x260 2 seq_release+0x40/0x60 6 kmem_free+0x70/0xc0 24 free_as_io_context+0x20/0x200 1 acpi_get_object_info+0x3a0/0x3e0 1 acpi_add_single_object+0xcf0/0x1e40 2 con_release_unimap+0x80/0x140 1 free+0x20/0x40 SLAB_STORE_USER must be enabled for a slab cache by either booting with "slab_debug" or enabling user tracking specifically for the slab of interest. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:45 +08:00
end = pos;
else
start = pos;
}
/*
* Not found. Insert new tracking element.
slub: add ability to list alloc / free callers per slab This patch enables listing the callers who allocated or freed objects in a cache. For example to list the allocators for kmalloc-128 do cat /sys/slab/kmalloc-128/alloc_calls 7 sn_io_slot_fixup+0x40/0x700 7 sn_io_slot_fixup+0x80/0x700 9 sn_bus_fixup+0xe0/0x380 6 param_sysfs_setup+0xf0/0x280 276 percpu_populate+0xf0/0x1a0 19 __register_chrdev_region+0x30/0x360 8 expand_files+0x2e0/0x6e0 1 sys_epoll_create+0x60/0x200 1 __mounts_open+0x140/0x2c0 65 kmem_alloc+0x110/0x280 3 alloc_disk_node+0xe0/0x200 33 as_get_io_context+0x90/0x280 74 kobject_kset_add_dir+0x40/0x140 12 pci_create_bus+0x2a0/0x5c0 1 acpi_ev_create_gpe_block+0x120/0x9e0 41 con_insert_unipair+0x100/0x1c0 1 uart_open+0x1c0/0xba0 1 dma_pool_create+0xe0/0x340 2 neigh_table_init_no_netlink+0x260/0x4c0 6 neigh_parms_alloc+0x30/0x200 1 netlink_kernel_create+0x130/0x320 5 fz_hash_alloc+0x50/0xe0 2 sn_common_hubdev_init+0xd0/0x6e0 28 kernel_param_sysfs_setup+0x30/0x180 72 process_zones+0x70/0x2e0 cat /sys/slab/kmalloc-128/free_calls 558 <not-available> 3 sn_io_slot_fixup+0x600/0x700 84 free_fdtable_rcu+0x120/0x260 2 seq_release+0x40/0x60 6 kmem_free+0x70/0xc0 24 free_as_io_context+0x20/0x200 1 acpi_get_object_info+0x3a0/0x3e0 1 acpi_add_single_object+0xcf0/0x1e40 2 con_release_unimap+0x80/0x140 1 free+0x20/0x40 SLAB_STORE_USER must be enabled for a slab cache by either booting with "slab_debug" or enabling user tracking specifically for the slab of interest. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:45 +08:00
*/
if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
slub: add ability to list alloc / free callers per slab This patch enables listing the callers who allocated or freed objects in a cache. For example to list the allocators for kmalloc-128 do cat /sys/slab/kmalloc-128/alloc_calls 7 sn_io_slot_fixup+0x40/0x700 7 sn_io_slot_fixup+0x80/0x700 9 sn_bus_fixup+0xe0/0x380 6 param_sysfs_setup+0xf0/0x280 276 percpu_populate+0xf0/0x1a0 19 __register_chrdev_region+0x30/0x360 8 expand_files+0x2e0/0x6e0 1 sys_epoll_create+0x60/0x200 1 __mounts_open+0x140/0x2c0 65 kmem_alloc+0x110/0x280 3 alloc_disk_node+0xe0/0x200 33 as_get_io_context+0x90/0x280 74 kobject_kset_add_dir+0x40/0x140 12 pci_create_bus+0x2a0/0x5c0 1 acpi_ev_create_gpe_block+0x120/0x9e0 41 con_insert_unipair+0x100/0x1c0 1 uart_open+0x1c0/0xba0 1 dma_pool_create+0xe0/0x340 2 neigh_table_init_no_netlink+0x260/0x4c0 6 neigh_parms_alloc+0x30/0x200 1 netlink_kernel_create+0x130/0x320 5 fz_hash_alloc+0x50/0xe0 2 sn_common_hubdev_init+0xd0/0x6e0 28 kernel_param_sysfs_setup+0x30/0x180 72 process_zones+0x70/0x2e0 cat /sys/slab/kmalloc-128/free_calls 558 <not-available> 3 sn_io_slot_fixup+0x600/0x700 84 free_fdtable_rcu+0x120/0x260 2 seq_release+0x40/0x60 6 kmem_free+0x70/0xc0 24 free_as_io_context+0x20/0x200 1 acpi_get_object_info+0x3a0/0x3e0 1 acpi_add_single_object+0xcf0/0x1e40 2 con_release_unimap+0x80/0x140 1 free+0x20/0x40 SLAB_STORE_USER must be enabled for a slab cache by either booting with "slab_debug" or enabling user tracking specifically for the slab of interest. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:45 +08:00
return 0;
l = t->loc + pos;
if (pos < t->count)
memmove(l + 1, l,
(t->count - pos) * sizeof(struct location));
t->count++;
l->count = 1;
SLUB: include lifetime stats and sets of cpus / nodes in tracking output We have information about how long an object existed and about the nodes and cpus where the allocations and frees took place. Add that information to the tracking output in /sys/slab/xx/alloc_calls and /sys/slab/free_calls This will then enable slabinfo to output nice reports like this: christoph@qirst:~/slub$ ./slabinfo kmalloc-128 Slabcache: kmalloc-128 Aliases: 0 Order : 0 Sizes (bytes) Slabs Debug Memory ------------------------------------------------------------------------ Object : 128 Total : 12 Sanity Checks : On Total: 49152 SlabObj: 200 Full : 7 Redzoning : On Used : 24832 SlabSiz: 4096 Partial: 4 Poisoning : On Loss : 24320 Loss : 72 CpuSlab: 1 Tracking : On Lalig: 13968 Align : 8 Objects: 20 Tracing : Off Lpadd: 1152 kmalloc-128 has no kmem_cache operations kmalloc-128: Kernel object allocation ----------------------------------------------------------------------- 6 param_sysfs_setup+0x71/0x130 age=284512/284512/284512 pid=1 nodes=0-1,3 11 percpu_populate+0x39/0x80 age=283914/284428/284512 pid=1 nodes=0 21 __register_chrdev_region+0x31/0x170 age=282896/284347/284473 pid=1-1705 nodes=0-2 1 sys_inotify_init+0x76/0x1c0 age=283423 pid=1004 nodes=0 19 as_get_io_context+0x32/0xd0 age=6/247567/283988 pid=1-11782 nodes=0,2 10 ida_pre_get+0x4a/0x80 age=277666/283773/284526 pid=0-2177 nodes=0,2 24 kobject_kset_add_dir+0x37/0xb0 age=282727/283860/284472 pid=1-1723 nodes=0-2 1 acpi_ds_build_internal_buffer_obj+0xd3/0x11d age=284508 pid=1 nodes=0 24 con_insert_unipair+0xd7/0x110 age=284438/284438/284438 pid=1 nodes=0,2 1 uart_open+0x2d2/0x4b0 age=283896 pid=1 nodes=0 26 dma_pool_create+0x73/0x1a0 age=282762/282833/282916 pid=1705-1723 nodes=0 1 neigh_table_init_no_netlink+0xd2/0x210 age=284461 pid=1 nodes=0 2 neigh_parms_alloc+0x2b/0xe0 age=284410/284411/284412 pid=1 nodes=2 2 neigh_resolve_output+0x1e1/0x280 age=276289/276291/276293 pid=0-2443 nodes=0 1 netlink_kernel_create+0x90/0x170 age=284472 pid=1 nodes=0 4 xt_alloc_table_info+0x39/0xf0 age=283958/283958/283959 pid=1 nodes=1 3 fn_hash_insert+0x473/0x720 age=277653/277661/277666 pid=2177-2185 nodes=0 1 get_mtrr_state+0x285/0x2a0 age=284526 pid=0 nodes=0 1 cacheinfo_cpu_callback+0x26d/0x3e0 age=284458 pid=1 nodes=0 29 kernel_param_sysfs_setup+0x25/0x90 age=284511/284511/284512 pid=1 nodes=0-1,3 5 process_zones+0x5e/0x170 age=284546/284546/284546 pid=0 nodes=0 1 drm_core_init+0x48/0x160 age=284421 pid=1 nodes=2 kmalloc-128: Kernel object freeing ------------------------------------------------------------------------ 163 <not-available> age=4295176847 pid=0 nodes=0-3 1 __vunmap+0x6e/0xf0 age=282907 pid=1723 nodes=0 28 free_as_io_context+0x12/0x90 age=9243/262197/283474 pid=42-11754 nodes=0 1 acpi_get_object_info+0x1b7/0x1d4 age=284475 pid=1 nodes=0 1 do_acpi_find_child+0x45/0x4e age=284475 pid=1 nodes=0 NUMA nodes : 0 1 2 3 ------------------------------------------ All slabs 7 2 2 1 Partial slabs 2 2 0 0 Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 17:32:45 +08:00
l->addr = track->addr;
l->sum_time = age;
l->min_time = age;
l->max_time = age;
l->min_pid = track->pid;
l->max_pid = track->pid;
cpumask_clear(to_cpumask(l->cpus));
cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
SLUB: include lifetime stats and sets of cpus / nodes in tracking output We have information about how long an object existed and about the nodes and cpus where the allocations and frees took place. Add that information to the tracking output in /sys/slab/xx/alloc_calls and /sys/slab/free_calls This will then enable slabinfo to output nice reports like this: christoph@qirst:~/slub$ ./slabinfo kmalloc-128 Slabcache: kmalloc-128 Aliases: 0 Order : 0 Sizes (bytes) Slabs Debug Memory ------------------------------------------------------------------------ Object : 128 Total : 12 Sanity Checks : On Total: 49152 SlabObj: 200 Full : 7 Redzoning : On Used : 24832 SlabSiz: 4096 Partial: 4 Poisoning : On Loss : 24320 Loss : 72 CpuSlab: 1 Tracking : On Lalig: 13968 Align : 8 Objects: 20 Tracing : Off Lpadd: 1152 kmalloc-128 has no kmem_cache operations kmalloc-128: Kernel object allocation ----------------------------------------------------------------------- 6 param_sysfs_setup+0x71/0x130 age=284512/284512/284512 pid=1 nodes=0-1,3 11 percpu_populate+0x39/0x80 age=283914/284428/284512 pid=1 nodes=0 21 __register_chrdev_region+0x31/0x170 age=282896/284347/284473 pid=1-1705 nodes=0-2 1 sys_inotify_init+0x76/0x1c0 age=283423 pid=1004 nodes=0 19 as_get_io_context+0x32/0xd0 age=6/247567/283988 pid=1-11782 nodes=0,2 10 ida_pre_get+0x4a/0x80 age=277666/283773/284526 pid=0-2177 nodes=0,2 24 kobject_kset_add_dir+0x37/0xb0 age=282727/283860/284472 pid=1-1723 nodes=0-2 1 acpi_ds_build_internal_buffer_obj+0xd3/0x11d age=284508 pid=1 nodes=0 24 con_insert_unipair+0xd7/0x110 age=284438/284438/284438 pid=1 nodes=0,2 1 uart_open+0x2d2/0x4b0 age=283896 pid=1 nodes=0 26 dma_pool_create+0x73/0x1a0 age=282762/282833/282916 pid=1705-1723 nodes=0 1 neigh_table_init_no_netlink+0xd2/0x210 age=284461 pid=1 nodes=0 2 neigh_parms_alloc+0x2b/0xe0 age=284410/284411/284412 pid=1 nodes=2 2 neigh_resolve_output+0x1e1/0x280 age=276289/276291/276293 pid=0-2443 nodes=0 1 netlink_kernel_create+0x90/0x170 age=284472 pid=1 nodes=0 4 xt_alloc_table_info+0x39/0xf0 age=283958/283958/283959 pid=1 nodes=1 3 fn_hash_insert+0x473/0x720 age=277653/277661/277666 pid=2177-2185 nodes=0 1 get_mtrr_state+0x285/0x2a0 age=284526 pid=0 nodes=0 1 cacheinfo_cpu_callback+0x26d/0x3e0 age=284458 pid=1 nodes=0 29 kernel_param_sysfs_setup+0x25/0x90 age=284511/284511/284512 pid=1 nodes=0-1,3 5 process_zones+0x5e/0x170 age=284546/284546/284546 pid=0 nodes=0 1 drm_core_init+0x48/0x160 age=284421 pid=1 nodes=2 kmalloc-128: Kernel object freeing ------------------------------------------------------------------------ 163 <not-available> age=4295176847 pid=0 nodes=0-3 1 __vunmap+0x6e/0xf0 age=282907 pid=1723 nodes=0 28 free_as_io_context+0x12/0x90 age=9243/262197/283474 pid=42-11754 nodes=0 1 acpi_get_object_info+0x1b7/0x1d4 age=284475 pid=1 nodes=0 1 do_acpi_find_child+0x45/0x4e age=284475 pid=1 nodes=0 NUMA nodes : 0 1 2 3 ------------------------------------------ All slabs 7 2 2 1 Partial slabs 2 2 0 0 Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 17:32:45 +08:00
nodes_clear(l->nodes);
node_set(page_to_nid(virt_to_page(track)), l->nodes);
slub: add ability to list alloc / free callers per slab This patch enables listing the callers who allocated or freed objects in a cache. For example to list the allocators for kmalloc-128 do cat /sys/slab/kmalloc-128/alloc_calls 7 sn_io_slot_fixup+0x40/0x700 7 sn_io_slot_fixup+0x80/0x700 9 sn_bus_fixup+0xe0/0x380 6 param_sysfs_setup+0xf0/0x280 276 percpu_populate+0xf0/0x1a0 19 __register_chrdev_region+0x30/0x360 8 expand_files+0x2e0/0x6e0 1 sys_epoll_create+0x60/0x200 1 __mounts_open+0x140/0x2c0 65 kmem_alloc+0x110/0x280 3 alloc_disk_node+0xe0/0x200 33 as_get_io_context+0x90/0x280 74 kobject_kset_add_dir+0x40/0x140 12 pci_create_bus+0x2a0/0x5c0 1 acpi_ev_create_gpe_block+0x120/0x9e0 41 con_insert_unipair+0x100/0x1c0 1 uart_open+0x1c0/0xba0 1 dma_pool_create+0xe0/0x340 2 neigh_table_init_no_netlink+0x260/0x4c0 6 neigh_parms_alloc+0x30/0x200 1 netlink_kernel_create+0x130/0x320 5 fz_hash_alloc+0x50/0xe0 2 sn_common_hubdev_init+0xd0/0x6e0 28 kernel_param_sysfs_setup+0x30/0x180 72 process_zones+0x70/0x2e0 cat /sys/slab/kmalloc-128/free_calls 558 <not-available> 3 sn_io_slot_fixup+0x600/0x700 84 free_fdtable_rcu+0x120/0x260 2 seq_release+0x40/0x60 6 kmem_free+0x70/0xc0 24 free_as_io_context+0x20/0x200 1 acpi_get_object_info+0x3a0/0x3e0 1 acpi_add_single_object+0xcf0/0x1e40 2 con_release_unimap+0x80/0x140 1 free+0x20/0x40 SLAB_STORE_USER must be enabled for a slab cache by either booting with "slab_debug" or enabling user tracking specifically for the slab of interest. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:45 +08:00
return 1;
}
static void process_slab(struct loc_track *t, struct kmem_cache *s,
struct page *page, enum track_item alloc)
slub: add ability to list alloc / free callers per slab This patch enables listing the callers who allocated or freed objects in a cache. For example to list the allocators for kmalloc-128 do cat /sys/slab/kmalloc-128/alloc_calls 7 sn_io_slot_fixup+0x40/0x700 7 sn_io_slot_fixup+0x80/0x700 9 sn_bus_fixup+0xe0/0x380 6 param_sysfs_setup+0xf0/0x280 276 percpu_populate+0xf0/0x1a0 19 __register_chrdev_region+0x30/0x360 8 expand_files+0x2e0/0x6e0 1 sys_epoll_create+0x60/0x200 1 __mounts_open+0x140/0x2c0 65 kmem_alloc+0x110/0x280 3 alloc_disk_node+0xe0/0x200 33 as_get_io_context+0x90/0x280 74 kobject_kset_add_dir+0x40/0x140 12 pci_create_bus+0x2a0/0x5c0 1 acpi_ev_create_gpe_block+0x120/0x9e0 41 con_insert_unipair+0x100/0x1c0 1 uart_open+0x1c0/0xba0 1 dma_pool_create+0xe0/0x340 2 neigh_table_init_no_netlink+0x260/0x4c0 6 neigh_parms_alloc+0x30/0x200 1 netlink_kernel_create+0x130/0x320 5 fz_hash_alloc+0x50/0xe0 2 sn_common_hubdev_init+0xd0/0x6e0 28 kernel_param_sysfs_setup+0x30/0x180 72 process_zones+0x70/0x2e0 cat /sys/slab/kmalloc-128/free_calls 558 <not-available> 3 sn_io_slot_fixup+0x600/0x700 84 free_fdtable_rcu+0x120/0x260 2 seq_release+0x40/0x60 6 kmem_free+0x70/0xc0 24 free_as_io_context+0x20/0x200 1 acpi_get_object_info+0x3a0/0x3e0 1 acpi_add_single_object+0xcf0/0x1e40 2 con_release_unimap+0x80/0x140 1 free+0x20/0x40 SLAB_STORE_USER must be enabled for a slab cache by either booting with "slab_debug" or enabling user tracking specifically for the slab of interest. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:45 +08:00
{
void *addr = page_address(page);
slub: add ability to list alloc / free callers per slab This patch enables listing the callers who allocated or freed objects in a cache. For example to list the allocators for kmalloc-128 do cat /sys/slab/kmalloc-128/alloc_calls 7 sn_io_slot_fixup+0x40/0x700 7 sn_io_slot_fixup+0x80/0x700 9 sn_bus_fixup+0xe0/0x380 6 param_sysfs_setup+0xf0/0x280 276 percpu_populate+0xf0/0x1a0 19 __register_chrdev_region+0x30/0x360 8 expand_files+0x2e0/0x6e0 1 sys_epoll_create+0x60/0x200 1 __mounts_open+0x140/0x2c0 65 kmem_alloc+0x110/0x280 3 alloc_disk_node+0xe0/0x200 33 as_get_io_context+0x90/0x280 74 kobject_kset_add_dir+0x40/0x140 12 pci_create_bus+0x2a0/0x5c0 1 acpi_ev_create_gpe_block+0x120/0x9e0 41 con_insert_unipair+0x100/0x1c0 1 uart_open+0x1c0/0xba0 1 dma_pool_create+0xe0/0x340 2 neigh_table_init_no_netlink+0x260/0x4c0 6 neigh_parms_alloc+0x30/0x200 1 netlink_kernel_create+0x130/0x320 5 fz_hash_alloc+0x50/0xe0 2 sn_common_hubdev_init+0xd0/0x6e0 28 kernel_param_sysfs_setup+0x30/0x180 72 process_zones+0x70/0x2e0 cat /sys/slab/kmalloc-128/free_calls 558 <not-available> 3 sn_io_slot_fixup+0x600/0x700 84 free_fdtable_rcu+0x120/0x260 2 seq_release+0x40/0x60 6 kmem_free+0x70/0xc0 24 free_as_io_context+0x20/0x200 1 acpi_get_object_info+0x3a0/0x3e0 1 acpi_add_single_object+0xcf0/0x1e40 2 con_release_unimap+0x80/0x140 1 free+0x20/0x40 SLAB_STORE_USER must be enabled for a slab cache by either booting with "slab_debug" or enabling user tracking specifically for the slab of interest. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:45 +08:00
void *p;
unsigned long *map;
slub: add ability to list alloc / free callers per slab This patch enables listing the callers who allocated or freed objects in a cache. For example to list the allocators for kmalloc-128 do cat /sys/slab/kmalloc-128/alloc_calls 7 sn_io_slot_fixup+0x40/0x700 7 sn_io_slot_fixup+0x80/0x700 9 sn_bus_fixup+0xe0/0x380 6 param_sysfs_setup+0xf0/0x280 276 percpu_populate+0xf0/0x1a0 19 __register_chrdev_region+0x30/0x360 8 expand_files+0x2e0/0x6e0 1 sys_epoll_create+0x60/0x200 1 __mounts_open+0x140/0x2c0 65 kmem_alloc+0x110/0x280 3 alloc_disk_node+0xe0/0x200 33 as_get_io_context+0x90/0x280 74 kobject_kset_add_dir+0x40/0x140 12 pci_create_bus+0x2a0/0x5c0 1 acpi_ev_create_gpe_block+0x120/0x9e0 41 con_insert_unipair+0x100/0x1c0 1 uart_open+0x1c0/0xba0 1 dma_pool_create+0xe0/0x340 2 neigh_table_init_no_netlink+0x260/0x4c0 6 neigh_parms_alloc+0x30/0x200 1 netlink_kernel_create+0x130/0x320 5 fz_hash_alloc+0x50/0xe0 2 sn_common_hubdev_init+0xd0/0x6e0 28 kernel_param_sysfs_setup+0x30/0x180 72 process_zones+0x70/0x2e0 cat /sys/slab/kmalloc-128/free_calls 558 <not-available> 3 sn_io_slot_fixup+0x600/0x700 84 free_fdtable_rcu+0x120/0x260 2 seq_release+0x40/0x60 6 kmem_free+0x70/0xc0 24 free_as_io_context+0x20/0x200 1 acpi_get_object_info+0x3a0/0x3e0 1 acpi_add_single_object+0xcf0/0x1e40 2 con_release_unimap+0x80/0x140 1 free+0x20/0x40 SLAB_STORE_USER must be enabled for a slab cache by either booting with "slab_debug" or enabling user tracking specifically for the slab of interest. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:45 +08:00
map = get_map(s, page);
for_each_object(p, s, addr, page->objects)
if (!test_bit(__obj_to_index(s, addr, p), map))
SLUB: include lifetime stats and sets of cpus / nodes in tracking output We have information about how long an object existed and about the nodes and cpus where the allocations and frees took place. Add that information to the tracking output in /sys/slab/xx/alloc_calls and /sys/slab/free_calls This will then enable slabinfo to output nice reports like this: christoph@qirst:~/slub$ ./slabinfo kmalloc-128 Slabcache: kmalloc-128 Aliases: 0 Order : 0 Sizes (bytes) Slabs Debug Memory ------------------------------------------------------------------------ Object : 128 Total : 12 Sanity Checks : On Total: 49152 SlabObj: 200 Full : 7 Redzoning : On Used : 24832 SlabSiz: 4096 Partial: 4 Poisoning : On Loss : 24320 Loss : 72 CpuSlab: 1 Tracking : On Lalig: 13968 Align : 8 Objects: 20 Tracing : Off Lpadd: 1152 kmalloc-128 has no kmem_cache operations kmalloc-128: Kernel object allocation ----------------------------------------------------------------------- 6 param_sysfs_setup+0x71/0x130 age=284512/284512/284512 pid=1 nodes=0-1,3 11 percpu_populate+0x39/0x80 age=283914/284428/284512 pid=1 nodes=0 21 __register_chrdev_region+0x31/0x170 age=282896/284347/284473 pid=1-1705 nodes=0-2 1 sys_inotify_init+0x76/0x1c0 age=283423 pid=1004 nodes=0 19 as_get_io_context+0x32/0xd0 age=6/247567/283988 pid=1-11782 nodes=0,2 10 ida_pre_get+0x4a/0x80 age=277666/283773/284526 pid=0-2177 nodes=0,2 24 kobject_kset_add_dir+0x37/0xb0 age=282727/283860/284472 pid=1-1723 nodes=0-2 1 acpi_ds_build_internal_buffer_obj+0xd3/0x11d age=284508 pid=1 nodes=0 24 con_insert_unipair+0xd7/0x110 age=284438/284438/284438 pid=1 nodes=0,2 1 uart_open+0x2d2/0x4b0 age=283896 pid=1 nodes=0 26 dma_pool_create+0x73/0x1a0 age=282762/282833/282916 pid=1705-1723 nodes=0 1 neigh_table_init_no_netlink+0xd2/0x210 age=284461 pid=1 nodes=0 2 neigh_parms_alloc+0x2b/0xe0 age=284410/284411/284412 pid=1 nodes=2 2 neigh_resolve_output+0x1e1/0x280 age=276289/276291/276293 pid=0-2443 nodes=0 1 netlink_kernel_create+0x90/0x170 age=284472 pid=1 nodes=0 4 xt_alloc_table_info+0x39/0xf0 age=283958/283958/283959 pid=1 nodes=1 3 fn_hash_insert+0x473/0x720 age=277653/277661/277666 pid=2177-2185 nodes=0 1 get_mtrr_state+0x285/0x2a0 age=284526 pid=0 nodes=0 1 cacheinfo_cpu_callback+0x26d/0x3e0 age=284458 pid=1 nodes=0 29 kernel_param_sysfs_setup+0x25/0x90 age=284511/284511/284512 pid=1 nodes=0-1,3 5 process_zones+0x5e/0x170 age=284546/284546/284546 pid=0 nodes=0 1 drm_core_init+0x48/0x160 age=284421 pid=1 nodes=2 kmalloc-128: Kernel object freeing ------------------------------------------------------------------------ 163 <not-available> age=4295176847 pid=0 nodes=0-3 1 __vunmap+0x6e/0xf0 age=282907 pid=1723 nodes=0 28 free_as_io_context+0x12/0x90 age=9243/262197/283474 pid=42-11754 nodes=0 1 acpi_get_object_info+0x1b7/0x1d4 age=284475 pid=1 nodes=0 1 do_acpi_find_child+0x45/0x4e age=284475 pid=1 nodes=0 NUMA nodes : 0 1 2 3 ------------------------------------------ All slabs 7 2 2 1 Partial slabs 2 2 0 0 Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 17:32:45 +08:00
add_location(t, s, get_track(s, p, alloc));
put_map(map);
slub: add ability to list alloc / free callers per slab This patch enables listing the callers who allocated or freed objects in a cache. For example to list the allocators for kmalloc-128 do cat /sys/slab/kmalloc-128/alloc_calls 7 sn_io_slot_fixup+0x40/0x700 7 sn_io_slot_fixup+0x80/0x700 9 sn_bus_fixup+0xe0/0x380 6 param_sysfs_setup+0xf0/0x280 276 percpu_populate+0xf0/0x1a0 19 __register_chrdev_region+0x30/0x360 8 expand_files+0x2e0/0x6e0 1 sys_epoll_create+0x60/0x200 1 __mounts_open+0x140/0x2c0 65 kmem_alloc+0x110/0x280 3 alloc_disk_node+0xe0/0x200 33 as_get_io_context+0x90/0x280 74 kobject_kset_add_dir+0x40/0x140 12 pci_create_bus+0x2a0/0x5c0 1 acpi_ev_create_gpe_block+0x120/0x9e0 41 con_insert_unipair+0x100/0x1c0 1 uart_open+0x1c0/0xba0 1 dma_pool_create+0xe0/0x340 2 neigh_table_init_no_netlink+0x260/0x4c0 6 neigh_parms_alloc+0x30/0x200 1 netlink_kernel_create+0x130/0x320 5 fz_hash_alloc+0x50/0xe0 2 sn_common_hubdev_init+0xd0/0x6e0 28 kernel_param_sysfs_setup+0x30/0x180 72 process_zones+0x70/0x2e0 cat /sys/slab/kmalloc-128/free_calls 558 <not-available> 3 sn_io_slot_fixup+0x600/0x700 84 free_fdtable_rcu+0x120/0x260 2 seq_release+0x40/0x60 6 kmem_free+0x70/0xc0 24 free_as_io_context+0x20/0x200 1 acpi_get_object_info+0x3a0/0x3e0 1 acpi_add_single_object+0xcf0/0x1e40 2 con_release_unimap+0x80/0x140 1 free+0x20/0x40 SLAB_STORE_USER must be enabled for a slab cache by either booting with "slab_debug" or enabling user tracking specifically for the slab of interest. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:45 +08:00
}
static int list_locations(struct kmem_cache *s, char *buf,
enum track_item alloc)
slub: add ability to list alloc / free callers per slab This patch enables listing the callers who allocated or freed objects in a cache. For example to list the allocators for kmalloc-128 do cat /sys/slab/kmalloc-128/alloc_calls 7 sn_io_slot_fixup+0x40/0x700 7 sn_io_slot_fixup+0x80/0x700 9 sn_bus_fixup+0xe0/0x380 6 param_sysfs_setup+0xf0/0x280 276 percpu_populate+0xf0/0x1a0 19 __register_chrdev_region+0x30/0x360 8 expand_files+0x2e0/0x6e0 1 sys_epoll_create+0x60/0x200 1 __mounts_open+0x140/0x2c0 65 kmem_alloc+0x110/0x280 3 alloc_disk_node+0xe0/0x200 33 as_get_io_context+0x90/0x280 74 kobject_kset_add_dir+0x40/0x140 12 pci_create_bus+0x2a0/0x5c0 1 acpi_ev_create_gpe_block+0x120/0x9e0 41 con_insert_unipair+0x100/0x1c0 1 uart_open+0x1c0/0xba0 1 dma_pool_create+0xe0/0x340 2 neigh_table_init_no_netlink+0x260/0x4c0 6 neigh_parms_alloc+0x30/0x200 1 netlink_kernel_create+0x130/0x320 5 fz_hash_alloc+0x50/0xe0 2 sn_common_hubdev_init+0xd0/0x6e0 28 kernel_param_sysfs_setup+0x30/0x180 72 process_zones+0x70/0x2e0 cat /sys/slab/kmalloc-128/free_calls 558 <not-available> 3 sn_io_slot_fixup+0x600/0x700 84 free_fdtable_rcu+0x120/0x260 2 seq_release+0x40/0x60 6 kmem_free+0x70/0xc0 24 free_as_io_context+0x20/0x200 1 acpi_get_object_info+0x3a0/0x3e0 1 acpi_add_single_object+0xcf0/0x1e40 2 con_release_unimap+0x80/0x140 1 free+0x20/0x40 SLAB_STORE_USER must be enabled for a slab cache by either booting with "slab_debug" or enabling user tracking specifically for the slab of interest. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:45 +08:00
{
int len = 0;
slub: add ability to list alloc / free callers per slab This patch enables listing the callers who allocated or freed objects in a cache. For example to list the allocators for kmalloc-128 do cat /sys/slab/kmalloc-128/alloc_calls 7 sn_io_slot_fixup+0x40/0x700 7 sn_io_slot_fixup+0x80/0x700 9 sn_bus_fixup+0xe0/0x380 6 param_sysfs_setup+0xf0/0x280 276 percpu_populate+0xf0/0x1a0 19 __register_chrdev_region+0x30/0x360 8 expand_files+0x2e0/0x6e0 1 sys_epoll_create+0x60/0x200 1 __mounts_open+0x140/0x2c0 65 kmem_alloc+0x110/0x280 3 alloc_disk_node+0xe0/0x200 33 as_get_io_context+0x90/0x280 74 kobject_kset_add_dir+0x40/0x140 12 pci_create_bus+0x2a0/0x5c0 1 acpi_ev_create_gpe_block+0x120/0x9e0 41 con_insert_unipair+0x100/0x1c0 1 uart_open+0x1c0/0xba0 1 dma_pool_create+0xe0/0x340 2 neigh_table_init_no_netlink+0x260/0x4c0 6 neigh_parms_alloc+0x30/0x200 1 netlink_kernel_create+0x130/0x320 5 fz_hash_alloc+0x50/0xe0 2 sn_common_hubdev_init+0xd0/0x6e0 28 kernel_param_sysfs_setup+0x30/0x180 72 process_zones+0x70/0x2e0 cat /sys/slab/kmalloc-128/free_calls 558 <not-available> 3 sn_io_slot_fixup+0x600/0x700 84 free_fdtable_rcu+0x120/0x260 2 seq_release+0x40/0x60 6 kmem_free+0x70/0xc0 24 free_as_io_context+0x20/0x200 1 acpi_get_object_info+0x3a0/0x3e0 1 acpi_add_single_object+0xcf0/0x1e40 2 con_release_unimap+0x80/0x140 1 free+0x20/0x40 SLAB_STORE_USER must be enabled for a slab cache by either booting with "slab_debug" or enabling user tracking specifically for the slab of interest. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:45 +08:00
unsigned long i;
struct loc_track t = { 0, 0, NULL };
slub: add ability to list alloc / free callers per slab This patch enables listing the callers who allocated or freed objects in a cache. For example to list the allocators for kmalloc-128 do cat /sys/slab/kmalloc-128/alloc_calls 7 sn_io_slot_fixup+0x40/0x700 7 sn_io_slot_fixup+0x80/0x700 9 sn_bus_fixup+0xe0/0x380 6 param_sysfs_setup+0xf0/0x280 276 percpu_populate+0xf0/0x1a0 19 __register_chrdev_region+0x30/0x360 8 expand_files+0x2e0/0x6e0 1 sys_epoll_create+0x60/0x200 1 __mounts_open+0x140/0x2c0 65 kmem_alloc+0x110/0x280 3 alloc_disk_node+0xe0/0x200 33 as_get_io_context+0x90/0x280 74 kobject_kset_add_dir+0x40/0x140 12 pci_create_bus+0x2a0/0x5c0 1 acpi_ev_create_gpe_block+0x120/0x9e0 41 con_insert_unipair+0x100/0x1c0 1 uart_open+0x1c0/0xba0 1 dma_pool_create+0xe0/0x340 2 neigh_table_init_no_netlink+0x260/0x4c0 6 neigh_parms_alloc+0x30/0x200 1 netlink_kernel_create+0x130/0x320 5 fz_hash_alloc+0x50/0xe0 2 sn_common_hubdev_init+0xd0/0x6e0 28 kernel_param_sysfs_setup+0x30/0x180 72 process_zones+0x70/0x2e0 cat /sys/slab/kmalloc-128/free_calls 558 <not-available> 3 sn_io_slot_fixup+0x600/0x700 84 free_fdtable_rcu+0x120/0x260 2 seq_release+0x40/0x60 6 kmem_free+0x70/0xc0 24 free_as_io_context+0x20/0x200 1 acpi_get_object_info+0x3a0/0x3e0 1 acpi_add_single_object+0xcf0/0x1e40 2 con_release_unimap+0x80/0x140 1 free+0x20/0x40 SLAB_STORE_USER must be enabled for a slab cache by either booting with "slab_debug" or enabling user tracking specifically for the slab of interest. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:45 +08:00
int node;
struct kmem_cache_node *n;
slub: add ability to list alloc / free callers per slab This patch enables listing the callers who allocated or freed objects in a cache. For example to list the allocators for kmalloc-128 do cat /sys/slab/kmalloc-128/alloc_calls 7 sn_io_slot_fixup+0x40/0x700 7 sn_io_slot_fixup+0x80/0x700 9 sn_bus_fixup+0xe0/0x380 6 param_sysfs_setup+0xf0/0x280 276 percpu_populate+0xf0/0x1a0 19 __register_chrdev_region+0x30/0x360 8 expand_files+0x2e0/0x6e0 1 sys_epoll_create+0x60/0x200 1 __mounts_open+0x140/0x2c0 65 kmem_alloc+0x110/0x280 3 alloc_disk_node+0xe0/0x200 33 as_get_io_context+0x90/0x280 74 kobject_kset_add_dir+0x40/0x140 12 pci_create_bus+0x2a0/0x5c0 1 acpi_ev_create_gpe_block+0x120/0x9e0 41 con_insert_unipair+0x100/0x1c0 1 uart_open+0x1c0/0xba0 1 dma_pool_create+0xe0/0x340 2 neigh_table_init_no_netlink+0x260/0x4c0 6 neigh_parms_alloc+0x30/0x200 1 netlink_kernel_create+0x130/0x320 5 fz_hash_alloc+0x50/0xe0 2 sn_common_hubdev_init+0xd0/0x6e0 28 kernel_param_sysfs_setup+0x30/0x180 72 process_zones+0x70/0x2e0 cat /sys/slab/kmalloc-128/free_calls 558 <not-available> 3 sn_io_slot_fixup+0x600/0x700 84 free_fdtable_rcu+0x120/0x260 2 seq_release+0x40/0x60 6 kmem_free+0x70/0xc0 24 free_as_io_context+0x20/0x200 1 acpi_get_object_info+0x3a0/0x3e0 1 acpi_add_single_object+0xcf0/0x1e40 2 con_release_unimap+0x80/0x140 1 free+0x20/0x40 SLAB_STORE_USER must be enabled for a slab cache by either booting with "slab_debug" or enabling user tracking specifically for the slab of interest. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:45 +08:00
if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
GFP_KERNEL)) {
return sysfs_emit(buf, "Out of memory\n");
}
slub: add ability to list alloc / free callers per slab This patch enables listing the callers who allocated or freed objects in a cache. For example to list the allocators for kmalloc-128 do cat /sys/slab/kmalloc-128/alloc_calls 7 sn_io_slot_fixup+0x40/0x700 7 sn_io_slot_fixup+0x80/0x700 9 sn_bus_fixup+0xe0/0x380 6 param_sysfs_setup+0xf0/0x280 276 percpu_populate+0xf0/0x1a0 19 __register_chrdev_region+0x30/0x360 8 expand_files+0x2e0/0x6e0 1 sys_epoll_create+0x60/0x200 1 __mounts_open+0x140/0x2c0 65 kmem_alloc+0x110/0x280 3 alloc_disk_node+0xe0/0x200 33 as_get_io_context+0x90/0x280 74 kobject_kset_add_dir+0x40/0x140 12 pci_create_bus+0x2a0/0x5c0 1 acpi_ev_create_gpe_block+0x120/0x9e0 41 con_insert_unipair+0x100/0x1c0 1 uart_open+0x1c0/0xba0 1 dma_pool_create+0xe0/0x340 2 neigh_table_init_no_netlink+0x260/0x4c0 6 neigh_parms_alloc+0x30/0x200 1 netlink_kernel_create+0x130/0x320 5 fz_hash_alloc+0x50/0xe0 2 sn_common_hubdev_init+0xd0/0x6e0 28 kernel_param_sysfs_setup+0x30/0x180 72 process_zones+0x70/0x2e0 cat /sys/slab/kmalloc-128/free_calls 558 <not-available> 3 sn_io_slot_fixup+0x600/0x700 84 free_fdtable_rcu+0x120/0x260 2 seq_release+0x40/0x60 6 kmem_free+0x70/0xc0 24 free_as_io_context+0x20/0x200 1 acpi_get_object_info+0x3a0/0x3e0 1 acpi_add_single_object+0xcf0/0x1e40 2 con_release_unimap+0x80/0x140 1 free+0x20/0x40 SLAB_STORE_USER must be enabled for a slab cache by either booting with "slab_debug" or enabling user tracking specifically for the slab of interest. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:45 +08:00
/* Push back cpu slabs */
flush_all(s);
for_each_kmem_cache_node(s, node, n) {
slub: add ability to list alloc / free callers per slab This patch enables listing the callers who allocated or freed objects in a cache. For example to list the allocators for kmalloc-128 do cat /sys/slab/kmalloc-128/alloc_calls 7 sn_io_slot_fixup+0x40/0x700 7 sn_io_slot_fixup+0x80/0x700 9 sn_bus_fixup+0xe0/0x380 6 param_sysfs_setup+0xf0/0x280 276 percpu_populate+0xf0/0x1a0 19 __register_chrdev_region+0x30/0x360 8 expand_files+0x2e0/0x6e0 1 sys_epoll_create+0x60/0x200 1 __mounts_open+0x140/0x2c0 65 kmem_alloc+0x110/0x280 3 alloc_disk_node+0xe0/0x200 33 as_get_io_context+0x90/0x280 74 kobject_kset_add_dir+0x40/0x140 12 pci_create_bus+0x2a0/0x5c0 1 acpi_ev_create_gpe_block+0x120/0x9e0 41 con_insert_unipair+0x100/0x1c0 1 uart_open+0x1c0/0xba0 1 dma_pool_create+0xe0/0x340 2 neigh_table_init_no_netlink+0x260/0x4c0 6 neigh_parms_alloc+0x30/0x200 1 netlink_kernel_create+0x130/0x320 5 fz_hash_alloc+0x50/0xe0 2 sn_common_hubdev_init+0xd0/0x6e0 28 kernel_param_sysfs_setup+0x30/0x180 72 process_zones+0x70/0x2e0 cat /sys/slab/kmalloc-128/free_calls 558 <not-available> 3 sn_io_slot_fixup+0x600/0x700 84 free_fdtable_rcu+0x120/0x260 2 seq_release+0x40/0x60 6 kmem_free+0x70/0xc0 24 free_as_io_context+0x20/0x200 1 acpi_get_object_info+0x3a0/0x3e0 1 acpi_add_single_object+0xcf0/0x1e40 2 con_release_unimap+0x80/0x140 1 free+0x20/0x40 SLAB_STORE_USER must be enabled for a slab cache by either booting with "slab_debug" or enabling user tracking specifically for the slab of interest. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:45 +08:00
unsigned long flags;
struct page *page;
if (!atomic_long_read(&n->nr_slabs))
slub: add ability to list alloc / free callers per slab This patch enables listing the callers who allocated or freed objects in a cache. For example to list the allocators for kmalloc-128 do cat /sys/slab/kmalloc-128/alloc_calls 7 sn_io_slot_fixup+0x40/0x700 7 sn_io_slot_fixup+0x80/0x700 9 sn_bus_fixup+0xe0/0x380 6 param_sysfs_setup+0xf0/0x280 276 percpu_populate+0xf0/0x1a0 19 __register_chrdev_region+0x30/0x360 8 expand_files+0x2e0/0x6e0 1 sys_epoll_create+0x60/0x200 1 __mounts_open+0x140/0x2c0 65 kmem_alloc+0x110/0x280 3 alloc_disk_node+0xe0/0x200 33 as_get_io_context+0x90/0x280 74 kobject_kset_add_dir+0x40/0x140 12 pci_create_bus+0x2a0/0x5c0 1 acpi_ev_create_gpe_block+0x120/0x9e0 41 con_insert_unipair+0x100/0x1c0 1 uart_open+0x1c0/0xba0 1 dma_pool_create+0xe0/0x340 2 neigh_table_init_no_netlink+0x260/0x4c0 6 neigh_parms_alloc+0x30/0x200 1 netlink_kernel_create+0x130/0x320 5 fz_hash_alloc+0x50/0xe0 2 sn_common_hubdev_init+0xd0/0x6e0 28 kernel_param_sysfs_setup+0x30/0x180 72 process_zones+0x70/0x2e0 cat /sys/slab/kmalloc-128/free_calls 558 <not-available> 3 sn_io_slot_fixup+0x600/0x700 84 free_fdtable_rcu+0x120/0x260 2 seq_release+0x40/0x60 6 kmem_free+0x70/0xc0 24 free_as_io_context+0x20/0x200 1 acpi_get_object_info+0x3a0/0x3e0 1 acpi_add_single_object+0xcf0/0x1e40 2 con_release_unimap+0x80/0x140 1 free+0x20/0x40 SLAB_STORE_USER must be enabled for a slab cache by either booting with "slab_debug" or enabling user tracking specifically for the slab of interest. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:45 +08:00
continue;
spin_lock_irqsave(&n->list_lock, flags);
list_for_each_entry(page, &n->partial, slab_list)
process_slab(&t, s, page, alloc);
list_for_each_entry(page, &n->full, slab_list)
process_slab(&t, s, page, alloc);
slub: add ability to list alloc / free callers per slab This patch enables listing the callers who allocated or freed objects in a cache. For example to list the allocators for kmalloc-128 do cat /sys/slab/kmalloc-128/alloc_calls 7 sn_io_slot_fixup+0x40/0x700 7 sn_io_slot_fixup+0x80/0x700 9 sn_bus_fixup+0xe0/0x380 6 param_sysfs_setup+0xf0/0x280 276 percpu_populate+0xf0/0x1a0 19 __register_chrdev_region+0x30/0x360 8 expand_files+0x2e0/0x6e0 1 sys_epoll_create+0x60/0x200 1 __mounts_open+0x140/0x2c0 65 kmem_alloc+0x110/0x280 3 alloc_disk_node+0xe0/0x200 33 as_get_io_context+0x90/0x280 74 kobject_kset_add_dir+0x40/0x140 12 pci_create_bus+0x2a0/0x5c0 1 acpi_ev_create_gpe_block+0x120/0x9e0 41 con_insert_unipair+0x100/0x1c0 1 uart_open+0x1c0/0xba0 1 dma_pool_create+0xe0/0x340 2 neigh_table_init_no_netlink+0x260/0x4c0 6 neigh_parms_alloc+0x30/0x200 1 netlink_kernel_create+0x130/0x320 5 fz_hash_alloc+0x50/0xe0 2 sn_common_hubdev_init+0xd0/0x6e0 28 kernel_param_sysfs_setup+0x30/0x180 72 process_zones+0x70/0x2e0 cat /sys/slab/kmalloc-128/free_calls 558 <not-available> 3 sn_io_slot_fixup+0x600/0x700 84 free_fdtable_rcu+0x120/0x260 2 seq_release+0x40/0x60 6 kmem_free+0x70/0xc0 24 free_as_io_context+0x20/0x200 1 acpi_get_object_info+0x3a0/0x3e0 1 acpi_add_single_object+0xcf0/0x1e40 2 con_release_unimap+0x80/0x140 1 free+0x20/0x40 SLAB_STORE_USER must be enabled for a slab cache by either booting with "slab_debug" or enabling user tracking specifically for the slab of interest. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:45 +08:00
spin_unlock_irqrestore(&n->list_lock, flags);
}
for (i = 0; i < t.count; i++) {
SLUB: include lifetime stats and sets of cpus / nodes in tracking output We have information about how long an object existed and about the nodes and cpus where the allocations and frees took place. Add that information to the tracking output in /sys/slab/xx/alloc_calls and /sys/slab/free_calls This will then enable slabinfo to output nice reports like this: christoph@qirst:~/slub$ ./slabinfo kmalloc-128 Slabcache: kmalloc-128 Aliases: 0 Order : 0 Sizes (bytes) Slabs Debug Memory ------------------------------------------------------------------------ Object : 128 Total : 12 Sanity Checks : On Total: 49152 SlabObj: 200 Full : 7 Redzoning : On Used : 24832 SlabSiz: 4096 Partial: 4 Poisoning : On Loss : 24320 Loss : 72 CpuSlab: 1 Tracking : On Lalig: 13968 Align : 8 Objects: 20 Tracing : Off Lpadd: 1152 kmalloc-128 has no kmem_cache operations kmalloc-128: Kernel object allocation ----------------------------------------------------------------------- 6 param_sysfs_setup+0x71/0x130 age=284512/284512/284512 pid=1 nodes=0-1,3 11 percpu_populate+0x39/0x80 age=283914/284428/284512 pid=1 nodes=0 21 __register_chrdev_region+0x31/0x170 age=282896/284347/284473 pid=1-1705 nodes=0-2 1 sys_inotify_init+0x76/0x1c0 age=283423 pid=1004 nodes=0 19 as_get_io_context+0x32/0xd0 age=6/247567/283988 pid=1-11782 nodes=0,2 10 ida_pre_get+0x4a/0x80 age=277666/283773/284526 pid=0-2177 nodes=0,2 24 kobject_kset_add_dir+0x37/0xb0 age=282727/283860/284472 pid=1-1723 nodes=0-2 1 acpi_ds_build_internal_buffer_obj+0xd3/0x11d age=284508 pid=1 nodes=0 24 con_insert_unipair+0xd7/0x110 age=284438/284438/284438 pid=1 nodes=0,2 1 uart_open+0x2d2/0x4b0 age=283896 pid=1 nodes=0 26 dma_pool_create+0x73/0x1a0 age=282762/282833/282916 pid=1705-1723 nodes=0 1 neigh_table_init_no_netlink+0xd2/0x210 age=284461 pid=1 nodes=0 2 neigh_parms_alloc+0x2b/0xe0 age=284410/284411/284412 pid=1 nodes=2 2 neigh_resolve_output+0x1e1/0x280 age=276289/276291/276293 pid=0-2443 nodes=0 1 netlink_kernel_create+0x90/0x170 age=284472 pid=1 nodes=0 4 xt_alloc_table_info+0x39/0xf0 age=283958/283958/283959 pid=1 nodes=1 3 fn_hash_insert+0x473/0x720 age=277653/277661/277666 pid=2177-2185 nodes=0 1 get_mtrr_state+0x285/0x2a0 age=284526 pid=0 nodes=0 1 cacheinfo_cpu_callback+0x26d/0x3e0 age=284458 pid=1 nodes=0 29 kernel_param_sysfs_setup+0x25/0x90 age=284511/284511/284512 pid=1 nodes=0-1,3 5 process_zones+0x5e/0x170 age=284546/284546/284546 pid=0 nodes=0 1 drm_core_init+0x48/0x160 age=284421 pid=1 nodes=2 kmalloc-128: Kernel object freeing ------------------------------------------------------------------------ 163 <not-available> age=4295176847 pid=0 nodes=0-3 1 __vunmap+0x6e/0xf0 age=282907 pid=1723 nodes=0 28 free_as_io_context+0x12/0x90 age=9243/262197/283474 pid=42-11754 nodes=0 1 acpi_get_object_info+0x1b7/0x1d4 age=284475 pid=1 nodes=0 1 do_acpi_find_child+0x45/0x4e age=284475 pid=1 nodes=0 NUMA nodes : 0 1 2 3 ------------------------------------------ All slabs 7 2 2 1 Partial slabs 2 2 0 0 Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 17:32:45 +08:00
struct location *l = &t.loc[i];
slub: add ability to list alloc / free callers per slab This patch enables listing the callers who allocated or freed objects in a cache. For example to list the allocators for kmalloc-128 do cat /sys/slab/kmalloc-128/alloc_calls 7 sn_io_slot_fixup+0x40/0x700 7 sn_io_slot_fixup+0x80/0x700 9 sn_bus_fixup+0xe0/0x380 6 param_sysfs_setup+0xf0/0x280 276 percpu_populate+0xf0/0x1a0 19 __register_chrdev_region+0x30/0x360 8 expand_files+0x2e0/0x6e0 1 sys_epoll_create+0x60/0x200 1 __mounts_open+0x140/0x2c0 65 kmem_alloc+0x110/0x280 3 alloc_disk_node+0xe0/0x200 33 as_get_io_context+0x90/0x280 74 kobject_kset_add_dir+0x40/0x140 12 pci_create_bus+0x2a0/0x5c0 1 acpi_ev_create_gpe_block+0x120/0x9e0 41 con_insert_unipair+0x100/0x1c0 1 uart_open+0x1c0/0xba0 1 dma_pool_create+0xe0/0x340 2 neigh_table_init_no_netlink+0x260/0x4c0 6 neigh_parms_alloc+0x30/0x200 1 netlink_kernel_create+0x130/0x320 5 fz_hash_alloc+0x50/0xe0 2 sn_common_hubdev_init+0xd0/0x6e0 28 kernel_param_sysfs_setup+0x30/0x180 72 process_zones+0x70/0x2e0 cat /sys/slab/kmalloc-128/free_calls 558 <not-available> 3 sn_io_slot_fixup+0x600/0x700 84 free_fdtable_rcu+0x120/0x260 2 seq_release+0x40/0x60 6 kmem_free+0x70/0xc0 24 free_as_io_context+0x20/0x200 1 acpi_get_object_info+0x3a0/0x3e0 1 acpi_add_single_object+0xcf0/0x1e40 2 con_release_unimap+0x80/0x140 1 free+0x20/0x40 SLAB_STORE_USER must be enabled for a slab cache by either booting with "slab_debug" or enabling user tracking specifically for the slab of interest. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:45 +08:00
len += sysfs_emit_at(buf, len, "%7ld ", l->count);
SLUB: include lifetime stats and sets of cpus / nodes in tracking output We have information about how long an object existed and about the nodes and cpus where the allocations and frees took place. Add that information to the tracking output in /sys/slab/xx/alloc_calls and /sys/slab/free_calls This will then enable slabinfo to output nice reports like this: christoph@qirst:~/slub$ ./slabinfo kmalloc-128 Slabcache: kmalloc-128 Aliases: 0 Order : 0 Sizes (bytes) Slabs Debug Memory ------------------------------------------------------------------------ Object : 128 Total : 12 Sanity Checks : On Total: 49152 SlabObj: 200 Full : 7 Redzoning : On Used : 24832 SlabSiz: 4096 Partial: 4 Poisoning : On Loss : 24320 Loss : 72 CpuSlab: 1 Tracking : On Lalig: 13968 Align : 8 Objects: 20 Tracing : Off Lpadd: 1152 kmalloc-128 has no kmem_cache operations kmalloc-128: Kernel object allocation ----------------------------------------------------------------------- 6 param_sysfs_setup+0x71/0x130 age=284512/284512/284512 pid=1 nodes=0-1,3 11 percpu_populate+0x39/0x80 age=283914/284428/284512 pid=1 nodes=0 21 __register_chrdev_region+0x31/0x170 age=282896/284347/284473 pid=1-1705 nodes=0-2 1 sys_inotify_init+0x76/0x1c0 age=283423 pid=1004 nodes=0 19 as_get_io_context+0x32/0xd0 age=6/247567/283988 pid=1-11782 nodes=0,2 10 ida_pre_get+0x4a/0x80 age=277666/283773/284526 pid=0-2177 nodes=0,2 24 kobject_kset_add_dir+0x37/0xb0 age=282727/283860/284472 pid=1-1723 nodes=0-2 1 acpi_ds_build_internal_buffer_obj+0xd3/0x11d age=284508 pid=1 nodes=0 24 con_insert_unipair+0xd7/0x110 age=284438/284438/284438 pid=1 nodes=0,2 1 uart_open+0x2d2/0x4b0 age=283896 pid=1 nodes=0 26 dma_pool_create+0x73/0x1a0 age=282762/282833/282916 pid=1705-1723 nodes=0 1 neigh_table_init_no_netlink+0xd2/0x210 age=284461 pid=1 nodes=0 2 neigh_parms_alloc+0x2b/0xe0 age=284410/284411/284412 pid=1 nodes=2 2 neigh_resolve_output+0x1e1/0x280 age=276289/276291/276293 pid=0-2443 nodes=0 1 netlink_kernel_create+0x90/0x170 age=284472 pid=1 nodes=0 4 xt_alloc_table_info+0x39/0xf0 age=283958/283958/283959 pid=1 nodes=1 3 fn_hash_insert+0x473/0x720 age=277653/277661/277666 pid=2177-2185 nodes=0 1 get_mtrr_state+0x285/0x2a0 age=284526 pid=0 nodes=0 1 cacheinfo_cpu_callback+0x26d/0x3e0 age=284458 pid=1 nodes=0 29 kernel_param_sysfs_setup+0x25/0x90 age=284511/284511/284512 pid=1 nodes=0-1,3 5 process_zones+0x5e/0x170 age=284546/284546/284546 pid=0 nodes=0 1 drm_core_init+0x48/0x160 age=284421 pid=1 nodes=2 kmalloc-128: Kernel object freeing ------------------------------------------------------------------------ 163 <not-available> age=4295176847 pid=0 nodes=0-3 1 __vunmap+0x6e/0xf0 age=282907 pid=1723 nodes=0 28 free_as_io_context+0x12/0x90 age=9243/262197/283474 pid=42-11754 nodes=0 1 acpi_get_object_info+0x1b7/0x1d4 age=284475 pid=1 nodes=0 1 do_acpi_find_child+0x45/0x4e age=284475 pid=1 nodes=0 NUMA nodes : 0 1 2 3 ------------------------------------------ All slabs 7 2 2 1 Partial slabs 2 2 0 0 Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 17:32:45 +08:00
if (l->addr)
len += sysfs_emit_at(buf, len, "%pS", (void *)l->addr);
slub: add ability to list alloc / free callers per slab This patch enables listing the callers who allocated or freed objects in a cache. For example to list the allocators for kmalloc-128 do cat /sys/slab/kmalloc-128/alloc_calls 7 sn_io_slot_fixup+0x40/0x700 7 sn_io_slot_fixup+0x80/0x700 9 sn_bus_fixup+0xe0/0x380 6 param_sysfs_setup+0xf0/0x280 276 percpu_populate+0xf0/0x1a0 19 __register_chrdev_region+0x30/0x360 8 expand_files+0x2e0/0x6e0 1 sys_epoll_create+0x60/0x200 1 __mounts_open+0x140/0x2c0 65 kmem_alloc+0x110/0x280 3 alloc_disk_node+0xe0/0x200 33 as_get_io_context+0x90/0x280 74 kobject_kset_add_dir+0x40/0x140 12 pci_create_bus+0x2a0/0x5c0 1 acpi_ev_create_gpe_block+0x120/0x9e0 41 con_insert_unipair+0x100/0x1c0 1 uart_open+0x1c0/0xba0 1 dma_pool_create+0xe0/0x340 2 neigh_table_init_no_netlink+0x260/0x4c0 6 neigh_parms_alloc+0x30/0x200 1 netlink_kernel_create+0x130/0x320 5 fz_hash_alloc+0x50/0xe0 2 sn_common_hubdev_init+0xd0/0x6e0 28 kernel_param_sysfs_setup+0x30/0x180 72 process_zones+0x70/0x2e0 cat /sys/slab/kmalloc-128/free_calls 558 <not-available> 3 sn_io_slot_fixup+0x600/0x700 84 free_fdtable_rcu+0x120/0x260 2 seq_release+0x40/0x60 6 kmem_free+0x70/0xc0 24 free_as_io_context+0x20/0x200 1 acpi_get_object_info+0x3a0/0x3e0 1 acpi_add_single_object+0xcf0/0x1e40 2 con_release_unimap+0x80/0x140 1 free+0x20/0x40 SLAB_STORE_USER must be enabled for a slab cache by either booting with "slab_debug" or enabling user tracking specifically for the slab of interest. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:45 +08:00
else
len += sysfs_emit_at(buf, len, "<not-available>");
if (l->sum_time != l->min_time)
len += sysfs_emit_at(buf, len, " age=%ld/%ld/%ld",
l->min_time,
(long)div_u64(l->sum_time,
l->count),
l->max_time);
else
len += sysfs_emit_at(buf, len, " age=%ld", l->min_time);
SLUB: include lifetime stats and sets of cpus / nodes in tracking output We have information about how long an object existed and about the nodes and cpus where the allocations and frees took place. Add that information to the tracking output in /sys/slab/xx/alloc_calls and /sys/slab/free_calls This will then enable slabinfo to output nice reports like this: christoph@qirst:~/slub$ ./slabinfo kmalloc-128 Slabcache: kmalloc-128 Aliases: 0 Order : 0 Sizes (bytes) Slabs Debug Memory ------------------------------------------------------------------------ Object : 128 Total : 12 Sanity Checks : On Total: 49152 SlabObj: 200 Full : 7 Redzoning : On Used : 24832 SlabSiz: 4096 Partial: 4 Poisoning : On Loss : 24320 Loss : 72 CpuSlab: 1 Tracking : On Lalig: 13968 Align : 8 Objects: 20 Tracing : Off Lpadd: 1152 kmalloc-128 has no kmem_cache operations kmalloc-128: Kernel object allocation ----------------------------------------------------------------------- 6 param_sysfs_setup+0x71/0x130 age=284512/284512/284512 pid=1 nodes=0-1,3 11 percpu_populate+0x39/0x80 age=283914/284428/284512 pid=1 nodes=0 21 __register_chrdev_region+0x31/0x170 age=282896/284347/284473 pid=1-1705 nodes=0-2 1 sys_inotify_init+0x76/0x1c0 age=283423 pid=1004 nodes=0 19 as_get_io_context+0x32/0xd0 age=6/247567/283988 pid=1-11782 nodes=0,2 10 ida_pre_get+0x4a/0x80 age=277666/283773/284526 pid=0-2177 nodes=0,2 24 kobject_kset_add_dir+0x37/0xb0 age=282727/283860/284472 pid=1-1723 nodes=0-2 1 acpi_ds_build_internal_buffer_obj+0xd3/0x11d age=284508 pid=1 nodes=0 24 con_insert_unipair+0xd7/0x110 age=284438/284438/284438 pid=1 nodes=0,2 1 uart_open+0x2d2/0x4b0 age=283896 pid=1 nodes=0 26 dma_pool_create+0x73/0x1a0 age=282762/282833/282916 pid=1705-1723 nodes=0 1 neigh_table_init_no_netlink+0xd2/0x210 age=284461 pid=1 nodes=0 2 neigh_parms_alloc+0x2b/0xe0 age=284410/284411/284412 pid=1 nodes=2 2 neigh_resolve_output+0x1e1/0x280 age=276289/276291/276293 pid=0-2443 nodes=0 1 netlink_kernel_create+0x90/0x170 age=284472 pid=1 nodes=0 4 xt_alloc_table_info+0x39/0xf0 age=283958/283958/283959 pid=1 nodes=1 3 fn_hash_insert+0x473/0x720 age=277653/277661/277666 pid=2177-2185 nodes=0 1 get_mtrr_state+0x285/0x2a0 age=284526 pid=0 nodes=0 1 cacheinfo_cpu_callback+0x26d/0x3e0 age=284458 pid=1 nodes=0 29 kernel_param_sysfs_setup+0x25/0x90 age=284511/284511/284512 pid=1 nodes=0-1,3 5 process_zones+0x5e/0x170 age=284546/284546/284546 pid=0 nodes=0 1 drm_core_init+0x48/0x160 age=284421 pid=1 nodes=2 kmalloc-128: Kernel object freeing ------------------------------------------------------------------------ 163 <not-available> age=4295176847 pid=0 nodes=0-3 1 __vunmap+0x6e/0xf0 age=282907 pid=1723 nodes=0 28 free_as_io_context+0x12/0x90 age=9243/262197/283474 pid=42-11754 nodes=0 1 acpi_get_object_info+0x1b7/0x1d4 age=284475 pid=1 nodes=0 1 do_acpi_find_child+0x45/0x4e age=284475 pid=1 nodes=0 NUMA nodes : 0 1 2 3 ------------------------------------------ All slabs 7 2 2 1 Partial slabs 2 2 0 0 Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 17:32:45 +08:00
if (l->min_pid != l->max_pid)
len += sysfs_emit_at(buf, len, " pid=%ld-%ld",
l->min_pid, l->max_pid);
SLUB: include lifetime stats and sets of cpus / nodes in tracking output We have information about how long an object existed and about the nodes and cpus where the allocations and frees took place. Add that information to the tracking output in /sys/slab/xx/alloc_calls and /sys/slab/free_calls This will then enable slabinfo to output nice reports like this: christoph@qirst:~/slub$ ./slabinfo kmalloc-128 Slabcache: kmalloc-128 Aliases: 0 Order : 0 Sizes (bytes) Slabs Debug Memory ------------------------------------------------------------------------ Object : 128 Total : 12 Sanity Checks : On Total: 49152 SlabObj: 200 Full : 7 Redzoning : On Used : 24832 SlabSiz: 4096 Partial: 4 Poisoning : On Loss : 24320 Loss : 72 CpuSlab: 1 Tracking : On Lalig: 13968 Align : 8 Objects: 20 Tracing : Off Lpadd: 1152 kmalloc-128 has no kmem_cache operations kmalloc-128: Kernel object allocation ----------------------------------------------------------------------- 6 param_sysfs_setup+0x71/0x130 age=284512/284512/284512 pid=1 nodes=0-1,3 11 percpu_populate+0x39/0x80 age=283914/284428/284512 pid=1 nodes=0 21 __register_chrdev_region+0x31/0x170 age=282896/284347/284473 pid=1-1705 nodes=0-2 1 sys_inotify_init+0x76/0x1c0 age=283423 pid=1004 nodes=0 19 as_get_io_context+0x32/0xd0 age=6/247567/283988 pid=1-11782 nodes=0,2 10 ida_pre_get+0x4a/0x80 age=277666/283773/284526 pid=0-2177 nodes=0,2 24 kobject_kset_add_dir+0x37/0xb0 age=282727/283860/284472 pid=1-1723 nodes=0-2 1 acpi_ds_build_internal_buffer_obj+0xd3/0x11d age=284508 pid=1 nodes=0 24 con_insert_unipair+0xd7/0x110 age=284438/284438/284438 pid=1 nodes=0,2 1 uart_open+0x2d2/0x4b0 age=283896 pid=1 nodes=0 26 dma_pool_create+0x73/0x1a0 age=282762/282833/282916 pid=1705-1723 nodes=0 1 neigh_table_init_no_netlink+0xd2/0x210 age=284461 pid=1 nodes=0 2 neigh_parms_alloc+0x2b/0xe0 age=284410/284411/284412 pid=1 nodes=2 2 neigh_resolve_output+0x1e1/0x280 age=276289/276291/276293 pid=0-2443 nodes=0 1 netlink_kernel_create+0x90/0x170 age=284472 pid=1 nodes=0 4 xt_alloc_table_info+0x39/0xf0 age=283958/283958/283959 pid=1 nodes=1 3 fn_hash_insert+0x473/0x720 age=277653/277661/277666 pid=2177-2185 nodes=0 1 get_mtrr_state+0x285/0x2a0 age=284526 pid=0 nodes=0 1 cacheinfo_cpu_callback+0x26d/0x3e0 age=284458 pid=1 nodes=0 29 kernel_param_sysfs_setup+0x25/0x90 age=284511/284511/284512 pid=1 nodes=0-1,3 5 process_zones+0x5e/0x170 age=284546/284546/284546 pid=0 nodes=0 1 drm_core_init+0x48/0x160 age=284421 pid=1 nodes=2 kmalloc-128: Kernel object freeing ------------------------------------------------------------------------ 163 <not-available> age=4295176847 pid=0 nodes=0-3 1 __vunmap+0x6e/0xf0 age=282907 pid=1723 nodes=0 28 free_as_io_context+0x12/0x90 age=9243/262197/283474 pid=42-11754 nodes=0 1 acpi_get_object_info+0x1b7/0x1d4 age=284475 pid=1 nodes=0 1 do_acpi_find_child+0x45/0x4e age=284475 pid=1 nodes=0 NUMA nodes : 0 1 2 3 ------------------------------------------ All slabs 7 2 2 1 Partial slabs 2 2 0 0 Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 17:32:45 +08:00
else
len += sysfs_emit_at(buf, len, " pid=%ld",
l->min_pid);
SLUB: include lifetime stats and sets of cpus / nodes in tracking output We have information about how long an object existed and about the nodes and cpus where the allocations and frees took place. Add that information to the tracking output in /sys/slab/xx/alloc_calls and /sys/slab/free_calls This will then enable slabinfo to output nice reports like this: christoph@qirst:~/slub$ ./slabinfo kmalloc-128 Slabcache: kmalloc-128 Aliases: 0 Order : 0 Sizes (bytes) Slabs Debug Memory ------------------------------------------------------------------------ Object : 128 Total : 12 Sanity Checks : On Total: 49152 SlabObj: 200 Full : 7 Redzoning : On Used : 24832 SlabSiz: 4096 Partial: 4 Poisoning : On Loss : 24320 Loss : 72 CpuSlab: 1 Tracking : On Lalig: 13968 Align : 8 Objects: 20 Tracing : Off Lpadd: 1152 kmalloc-128 has no kmem_cache operations kmalloc-128: Kernel object allocation ----------------------------------------------------------------------- 6 param_sysfs_setup+0x71/0x130 age=284512/284512/284512 pid=1 nodes=0-1,3 11 percpu_populate+0x39/0x80 age=283914/284428/284512 pid=1 nodes=0 21 __register_chrdev_region+0x31/0x170 age=282896/284347/284473 pid=1-1705 nodes=0-2 1 sys_inotify_init+0x76/0x1c0 age=283423 pid=1004 nodes=0 19 as_get_io_context+0x32/0xd0 age=6/247567/283988 pid=1-11782 nodes=0,2 10 ida_pre_get+0x4a/0x80 age=277666/283773/284526 pid=0-2177 nodes=0,2 24 kobject_kset_add_dir+0x37/0xb0 age=282727/283860/284472 pid=1-1723 nodes=0-2 1 acpi_ds_build_internal_buffer_obj+0xd3/0x11d age=284508 pid=1 nodes=0 24 con_insert_unipair+0xd7/0x110 age=284438/284438/284438 pid=1 nodes=0,2 1 uart_open+0x2d2/0x4b0 age=283896 pid=1 nodes=0 26 dma_pool_create+0x73/0x1a0 age=282762/282833/282916 pid=1705-1723 nodes=0 1 neigh_table_init_no_netlink+0xd2/0x210 age=284461 pid=1 nodes=0 2 neigh_parms_alloc+0x2b/0xe0 age=284410/284411/284412 pid=1 nodes=2 2 neigh_resolve_output+0x1e1/0x280 age=276289/276291/276293 pid=0-2443 nodes=0 1 netlink_kernel_create+0x90/0x170 age=284472 pid=1 nodes=0 4 xt_alloc_table_info+0x39/0xf0 age=283958/283958/283959 pid=1 nodes=1 3 fn_hash_insert+0x473/0x720 age=277653/277661/277666 pid=2177-2185 nodes=0 1 get_mtrr_state+0x285/0x2a0 age=284526 pid=0 nodes=0 1 cacheinfo_cpu_callback+0x26d/0x3e0 age=284458 pid=1 nodes=0 29 kernel_param_sysfs_setup+0x25/0x90 age=284511/284511/284512 pid=1 nodes=0-1,3 5 process_zones+0x5e/0x170 age=284546/284546/284546 pid=0 nodes=0 1 drm_core_init+0x48/0x160 age=284421 pid=1 nodes=2 kmalloc-128: Kernel object freeing ------------------------------------------------------------------------ 163 <not-available> age=4295176847 pid=0 nodes=0-3 1 __vunmap+0x6e/0xf0 age=282907 pid=1723 nodes=0 28 free_as_io_context+0x12/0x90 age=9243/262197/283474 pid=42-11754 nodes=0 1 acpi_get_object_info+0x1b7/0x1d4 age=284475 pid=1 nodes=0 1 do_acpi_find_child+0x45/0x4e age=284475 pid=1 nodes=0 NUMA nodes : 0 1 2 3 ------------------------------------------ All slabs 7 2 2 1 Partial slabs 2 2 0 0 Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 17:32:45 +08:00
if (num_online_cpus() > 1 &&
!cpumask_empty(to_cpumask(l->cpus)))
len += sysfs_emit_at(buf, len, " cpus=%*pbl",
cpumask_pr_args(to_cpumask(l->cpus)));
if (nr_online_nodes > 1 && !nodes_empty(l->nodes))
len += sysfs_emit_at(buf, len, " nodes=%*pbl",
nodemask_pr_args(&l->nodes));
len += sysfs_emit_at(buf, len, "\n");
slub: add ability to list alloc / free callers per slab This patch enables listing the callers who allocated or freed objects in a cache. For example to list the allocators for kmalloc-128 do cat /sys/slab/kmalloc-128/alloc_calls 7 sn_io_slot_fixup+0x40/0x700 7 sn_io_slot_fixup+0x80/0x700 9 sn_bus_fixup+0xe0/0x380 6 param_sysfs_setup+0xf0/0x280 276 percpu_populate+0xf0/0x1a0 19 __register_chrdev_region+0x30/0x360 8 expand_files+0x2e0/0x6e0 1 sys_epoll_create+0x60/0x200 1 __mounts_open+0x140/0x2c0 65 kmem_alloc+0x110/0x280 3 alloc_disk_node+0xe0/0x200 33 as_get_io_context+0x90/0x280 74 kobject_kset_add_dir+0x40/0x140 12 pci_create_bus+0x2a0/0x5c0 1 acpi_ev_create_gpe_block+0x120/0x9e0 41 con_insert_unipair+0x100/0x1c0 1 uart_open+0x1c0/0xba0 1 dma_pool_create+0xe0/0x340 2 neigh_table_init_no_netlink+0x260/0x4c0 6 neigh_parms_alloc+0x30/0x200 1 netlink_kernel_create+0x130/0x320 5 fz_hash_alloc+0x50/0xe0 2 sn_common_hubdev_init+0xd0/0x6e0 28 kernel_param_sysfs_setup+0x30/0x180 72 process_zones+0x70/0x2e0 cat /sys/slab/kmalloc-128/free_calls 558 <not-available> 3 sn_io_slot_fixup+0x600/0x700 84 free_fdtable_rcu+0x120/0x260 2 seq_release+0x40/0x60 6 kmem_free+0x70/0xc0 24 free_as_io_context+0x20/0x200 1 acpi_get_object_info+0x3a0/0x3e0 1 acpi_add_single_object+0xcf0/0x1e40 2 con_release_unimap+0x80/0x140 1 free+0x20/0x40 SLAB_STORE_USER must be enabled for a slab cache by either booting with "slab_debug" or enabling user tracking specifically for the slab of interest. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:45 +08:00
}
free_loc_track(&t);
if (!t.count)
len += sysfs_emit_at(buf, len, "No data\n");
return len;
slub: add ability to list alloc / free callers per slab This patch enables listing the callers who allocated or freed objects in a cache. For example to list the allocators for kmalloc-128 do cat /sys/slab/kmalloc-128/alloc_calls 7 sn_io_slot_fixup+0x40/0x700 7 sn_io_slot_fixup+0x80/0x700 9 sn_bus_fixup+0xe0/0x380 6 param_sysfs_setup+0xf0/0x280 276 percpu_populate+0xf0/0x1a0 19 __register_chrdev_region+0x30/0x360 8 expand_files+0x2e0/0x6e0 1 sys_epoll_create+0x60/0x200 1 __mounts_open+0x140/0x2c0 65 kmem_alloc+0x110/0x280 3 alloc_disk_node+0xe0/0x200 33 as_get_io_context+0x90/0x280 74 kobject_kset_add_dir+0x40/0x140 12 pci_create_bus+0x2a0/0x5c0 1 acpi_ev_create_gpe_block+0x120/0x9e0 41 con_insert_unipair+0x100/0x1c0 1 uart_open+0x1c0/0xba0 1 dma_pool_create+0xe0/0x340 2 neigh_table_init_no_netlink+0x260/0x4c0 6 neigh_parms_alloc+0x30/0x200 1 netlink_kernel_create+0x130/0x320 5 fz_hash_alloc+0x50/0xe0 2 sn_common_hubdev_init+0xd0/0x6e0 28 kernel_param_sysfs_setup+0x30/0x180 72 process_zones+0x70/0x2e0 cat /sys/slab/kmalloc-128/free_calls 558 <not-available> 3 sn_io_slot_fixup+0x600/0x700 84 free_fdtable_rcu+0x120/0x260 2 seq_release+0x40/0x60 6 kmem_free+0x70/0xc0 24 free_as_io_context+0x20/0x200 1 acpi_get_object_info+0x3a0/0x3e0 1 acpi_add_single_object+0xcf0/0x1e40 2 con_release_unimap+0x80/0x140 1 free+0x20/0x40 SLAB_STORE_USER must be enabled for a slab cache by either booting with "slab_debug" or enabling user tracking specifically for the slab of interest. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:45 +08:00
}
#endif /* CONFIG_SLUB_DEBUG */
slub: add ability to list alloc / free callers per slab This patch enables listing the callers who allocated or freed objects in a cache. For example to list the allocators for kmalloc-128 do cat /sys/slab/kmalloc-128/alloc_calls 7 sn_io_slot_fixup+0x40/0x700 7 sn_io_slot_fixup+0x80/0x700 9 sn_bus_fixup+0xe0/0x380 6 param_sysfs_setup+0xf0/0x280 276 percpu_populate+0xf0/0x1a0 19 __register_chrdev_region+0x30/0x360 8 expand_files+0x2e0/0x6e0 1 sys_epoll_create+0x60/0x200 1 __mounts_open+0x140/0x2c0 65 kmem_alloc+0x110/0x280 3 alloc_disk_node+0xe0/0x200 33 as_get_io_context+0x90/0x280 74 kobject_kset_add_dir+0x40/0x140 12 pci_create_bus+0x2a0/0x5c0 1 acpi_ev_create_gpe_block+0x120/0x9e0 41 con_insert_unipair+0x100/0x1c0 1 uart_open+0x1c0/0xba0 1 dma_pool_create+0xe0/0x340 2 neigh_table_init_no_netlink+0x260/0x4c0 6 neigh_parms_alloc+0x30/0x200 1 netlink_kernel_create+0x130/0x320 5 fz_hash_alloc+0x50/0xe0 2 sn_common_hubdev_init+0xd0/0x6e0 28 kernel_param_sysfs_setup+0x30/0x180 72 process_zones+0x70/0x2e0 cat /sys/slab/kmalloc-128/free_calls 558 <not-available> 3 sn_io_slot_fixup+0x600/0x700 84 free_fdtable_rcu+0x120/0x260 2 seq_release+0x40/0x60 6 kmem_free+0x70/0xc0 24 free_as_io_context+0x20/0x200 1 acpi_get_object_info+0x3a0/0x3e0 1 acpi_add_single_object+0xcf0/0x1e40 2 con_release_unimap+0x80/0x140 1 free+0x20/0x40 SLAB_STORE_USER must be enabled for a slab cache by either booting with "slab_debug" or enabling user tracking specifically for the slab of interest. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:45 +08:00
#ifdef SLUB_RESILIENCY_TEST
static void __init resiliency_test(void)
{
u8 *p;
mm, slab: combine kmalloc_caches and kmalloc_dma_caches Patch series "kmalloc-reclaimable caches", v4. As discussed at LSF/MM [1] here's a patchset that introduces kmalloc-reclaimable caches (more details in the second patch) and uses them for dcache external names. That allows us to repurpose the NR_INDIRECTLY_RECLAIMABLE_BYTES counter later in the series. With patch 3/6, dcache external names are allocated from kmalloc-rcl-* caches, eliminating the need for manual accounting. More importantly, it also ensures the reclaimable kmalloc allocations are grouped in pages separate from the regular kmalloc allocations. The need for proper accounting of dcache external names has shown it's easy for misbehaving process to allocate lots of them, causing premature OOMs. Without the added grouping, it's likely that a similar workload can interleave the dcache external names allocations with regular kmalloc allocations (note: I haven't searched myself for an example of such regular kmalloc allocation, but I would be very surprised if there wasn't some). A pathological case would be e.g. one 64byte regular allocations with 63 external dcache names in a page (64x64=4096), which means the page is not freed even after reclaiming after all dcache names, and the process can thus "steal" the whole page with single 64byte allocation. If other kmalloc users similar to dcache external names become identified, they can also benefit from the new functionality simply by adding __GFP_RECLAIMABLE to the kmalloc calls. Side benefits of the patchset (that could be also merged separately) include removed branch for detecting __GFP_DMA kmalloc(), and shortening kmalloc cache names in /proc/slabinfo output. The latter is potentially an ABI break in case there are tools parsing the names and expecting the values to be in bytes. This is how /proc/slabinfo looks like after booting in virtme: ... kmalloc-rcl-4M 0 0 4194304 1 1024 : tunables 1 1 0 : slabdata 0 0 0 ... kmalloc-rcl-96 7 32 128 32 1 : tunables 120 60 8 : slabdata 1 1 0 kmalloc-rcl-64 25 128 64 64 1 : tunables 120 60 8 : slabdata 2 2 0 kmalloc-rcl-32 0 0 32 124 1 : tunables 120 60 8 : slabdata 0 0 0 kmalloc-4M 0 0 4194304 1 1024 : tunables 1 1 0 : slabdata 0 0 0 kmalloc-2M 0 0 2097152 1 512 : tunables 1 1 0 : slabdata 0 0 0 kmalloc-1M 0 0 1048576 1 256 : tunables 1 1 0 : slabdata 0 0 0 ... /proc/vmstat with renamed nr_indirectly_reclaimable_bytes counter: ... nr_slab_reclaimable 2817 nr_slab_unreclaimable 1781 ... nr_kernel_misc_reclaimable 0 ... /proc/meminfo with new KReclaimable counter: ... Shmem: 564 kB KReclaimable: 11260 kB Slab: 18368 kB SReclaimable: 11260 kB SUnreclaim: 7108 kB KernelStack: 1248 kB ... This patch (of 6): The kmalloc caches currently mainain separate (optional) array kmalloc_dma_caches for __GFP_DMA allocations. There are tests for __GFP_DMA in the allocation hotpaths. We can avoid the branches by combining kmalloc_caches and kmalloc_dma_caches into a single two-dimensional array where the outer dimension is cache "type". This will also allow to add kmalloc-reclaimable caches as a third type. Link: http://lkml.kernel.org/r/20180731090649.16028-2-vbabka@suse.cz Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Acked-by: Mel Gorman <mgorman@techsingularity.net> Acked-by: Christoph Lameter <cl@linux.com> Acked-by: Roman Gushchin <guro@fb.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: Laura Abbott <labbott@redhat.com> Cc: Sumit Semwal <sumit.semwal@linaro.org> Cc: Vijayanand Jitta <vjitta@codeaurora.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-27 06:05:34 +08:00
int type = KMALLOC_NORMAL;
BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
pr_err("SLUB resiliency testing\n");
pr_err("-----------------------\n");
pr_err("A. Corruption after allocation\n");
p = kzalloc(16, GFP_KERNEL);
p[16] = 0x12;
pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
p + 16);
mm, slab: combine kmalloc_caches and kmalloc_dma_caches Patch series "kmalloc-reclaimable caches", v4. As discussed at LSF/MM [1] here's a patchset that introduces kmalloc-reclaimable caches (more details in the second patch) and uses them for dcache external names. That allows us to repurpose the NR_INDIRECTLY_RECLAIMABLE_BYTES counter later in the series. With patch 3/6, dcache external names are allocated from kmalloc-rcl-* caches, eliminating the need for manual accounting. More importantly, it also ensures the reclaimable kmalloc allocations are grouped in pages separate from the regular kmalloc allocations. The need for proper accounting of dcache external names has shown it's easy for misbehaving process to allocate lots of them, causing premature OOMs. Without the added grouping, it's likely that a similar workload can interleave the dcache external names allocations with regular kmalloc allocations (note: I haven't searched myself for an example of such regular kmalloc allocation, but I would be very surprised if there wasn't some). A pathological case would be e.g. one 64byte regular allocations with 63 external dcache names in a page (64x64=4096), which means the page is not freed even after reclaiming after all dcache names, and the process can thus "steal" the whole page with single 64byte allocation. If other kmalloc users similar to dcache external names become identified, they can also benefit from the new functionality simply by adding __GFP_RECLAIMABLE to the kmalloc calls. Side benefits of the patchset (that could be also merged separately) include removed branch for detecting __GFP_DMA kmalloc(), and shortening kmalloc cache names in /proc/slabinfo output. The latter is potentially an ABI break in case there are tools parsing the names and expecting the values to be in bytes. This is how /proc/slabinfo looks like after booting in virtme: ... kmalloc-rcl-4M 0 0 4194304 1 1024 : tunables 1 1 0 : slabdata 0 0 0 ... kmalloc-rcl-96 7 32 128 32 1 : tunables 120 60 8 : slabdata 1 1 0 kmalloc-rcl-64 25 128 64 64 1 : tunables 120 60 8 : slabdata 2 2 0 kmalloc-rcl-32 0 0 32 124 1 : tunables 120 60 8 : slabdata 0 0 0 kmalloc-4M 0 0 4194304 1 1024 : tunables 1 1 0 : slabdata 0 0 0 kmalloc-2M 0 0 2097152 1 512 : tunables 1 1 0 : slabdata 0 0 0 kmalloc-1M 0 0 1048576 1 256 : tunables 1 1 0 : slabdata 0 0 0 ... /proc/vmstat with renamed nr_indirectly_reclaimable_bytes counter: ... nr_slab_reclaimable 2817 nr_slab_unreclaimable 1781 ... nr_kernel_misc_reclaimable 0 ... /proc/meminfo with new KReclaimable counter: ... Shmem: 564 kB KReclaimable: 11260 kB Slab: 18368 kB SReclaimable: 11260 kB SUnreclaim: 7108 kB KernelStack: 1248 kB ... This patch (of 6): The kmalloc caches currently mainain separate (optional) array kmalloc_dma_caches for __GFP_DMA allocations. There are tests for __GFP_DMA in the allocation hotpaths. We can avoid the branches by combining kmalloc_caches and kmalloc_dma_caches into a single two-dimensional array where the outer dimension is cache "type". This will also allow to add kmalloc-reclaimable caches as a third type. Link: http://lkml.kernel.org/r/20180731090649.16028-2-vbabka@suse.cz Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Acked-by: Mel Gorman <mgorman@techsingularity.net> Acked-by: Christoph Lameter <cl@linux.com> Acked-by: Roman Gushchin <guro@fb.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: Laura Abbott <labbott@redhat.com> Cc: Sumit Semwal <sumit.semwal@linaro.org> Cc: Vijayanand Jitta <vjitta@codeaurora.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-27 06:05:34 +08:00
validate_slab_cache(kmalloc_caches[type][4]);
/* Hmmm... The next two are dangerous */
p = kzalloc(32, GFP_KERNEL);
p[32 + sizeof(void *)] = 0x34;
pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
p);
pr_err("If allocated object is overwritten then not detectable\n\n");
mm, slab: combine kmalloc_caches and kmalloc_dma_caches Patch series "kmalloc-reclaimable caches", v4. As discussed at LSF/MM [1] here's a patchset that introduces kmalloc-reclaimable caches (more details in the second patch) and uses them for dcache external names. That allows us to repurpose the NR_INDIRECTLY_RECLAIMABLE_BYTES counter later in the series. With patch 3/6, dcache external names are allocated from kmalloc-rcl-* caches, eliminating the need for manual accounting. More importantly, it also ensures the reclaimable kmalloc allocations are grouped in pages separate from the regular kmalloc allocations. The need for proper accounting of dcache external names has shown it's easy for misbehaving process to allocate lots of them, causing premature OOMs. Without the added grouping, it's likely that a similar workload can interleave the dcache external names allocations with regular kmalloc allocations (note: I haven't searched myself for an example of such regular kmalloc allocation, but I would be very surprised if there wasn't some). A pathological case would be e.g. one 64byte regular allocations with 63 external dcache names in a page (64x64=4096), which means the page is not freed even after reclaiming after all dcache names, and the process can thus "steal" the whole page with single 64byte allocation. If other kmalloc users similar to dcache external names become identified, they can also benefit from the new functionality simply by adding __GFP_RECLAIMABLE to the kmalloc calls. Side benefits of the patchset (that could be also merged separately) include removed branch for detecting __GFP_DMA kmalloc(), and shortening kmalloc cache names in /proc/slabinfo output. The latter is potentially an ABI break in case there are tools parsing the names and expecting the values to be in bytes. This is how /proc/slabinfo looks like after booting in virtme: ... kmalloc-rcl-4M 0 0 4194304 1 1024 : tunables 1 1 0 : slabdata 0 0 0 ... kmalloc-rcl-96 7 32 128 32 1 : tunables 120 60 8 : slabdata 1 1 0 kmalloc-rcl-64 25 128 64 64 1 : tunables 120 60 8 : slabdata 2 2 0 kmalloc-rcl-32 0 0 32 124 1 : tunables 120 60 8 : slabdata 0 0 0 kmalloc-4M 0 0 4194304 1 1024 : tunables 1 1 0 : slabdata 0 0 0 kmalloc-2M 0 0 2097152 1 512 : tunables 1 1 0 : slabdata 0 0 0 kmalloc-1M 0 0 1048576 1 256 : tunables 1 1 0 : slabdata 0 0 0 ... /proc/vmstat with renamed nr_indirectly_reclaimable_bytes counter: ... nr_slab_reclaimable 2817 nr_slab_unreclaimable 1781 ... nr_kernel_misc_reclaimable 0 ... /proc/meminfo with new KReclaimable counter: ... Shmem: 564 kB KReclaimable: 11260 kB Slab: 18368 kB SReclaimable: 11260 kB SUnreclaim: 7108 kB KernelStack: 1248 kB ... This patch (of 6): The kmalloc caches currently mainain separate (optional) array kmalloc_dma_caches for __GFP_DMA allocations. There are tests for __GFP_DMA in the allocation hotpaths. We can avoid the branches by combining kmalloc_caches and kmalloc_dma_caches into a single two-dimensional array where the outer dimension is cache "type". This will also allow to add kmalloc-reclaimable caches as a third type. Link: http://lkml.kernel.org/r/20180731090649.16028-2-vbabka@suse.cz Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Acked-by: Mel Gorman <mgorman@techsingularity.net> Acked-by: Christoph Lameter <cl@linux.com> Acked-by: Roman Gushchin <guro@fb.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: Laura Abbott <labbott@redhat.com> Cc: Sumit Semwal <sumit.semwal@linaro.org> Cc: Vijayanand Jitta <vjitta@codeaurora.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-27 06:05:34 +08:00
validate_slab_cache(kmalloc_caches[type][5]);
p = kzalloc(64, GFP_KERNEL);
p += 64 + (get_cycles() & 0xff) * sizeof(void *);
*p = 0x56;
pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
p);
pr_err("If allocated object is overwritten then not detectable\n\n");
mm, slab: combine kmalloc_caches and kmalloc_dma_caches Patch series "kmalloc-reclaimable caches", v4. As discussed at LSF/MM [1] here's a patchset that introduces kmalloc-reclaimable caches (more details in the second patch) and uses them for dcache external names. That allows us to repurpose the NR_INDIRECTLY_RECLAIMABLE_BYTES counter later in the series. With patch 3/6, dcache external names are allocated from kmalloc-rcl-* caches, eliminating the need for manual accounting. More importantly, it also ensures the reclaimable kmalloc allocations are grouped in pages separate from the regular kmalloc allocations. The need for proper accounting of dcache external names has shown it's easy for misbehaving process to allocate lots of them, causing premature OOMs. Without the added grouping, it's likely that a similar workload can interleave the dcache external names allocations with regular kmalloc allocations (note: I haven't searched myself for an example of such regular kmalloc allocation, but I would be very surprised if there wasn't some). A pathological case would be e.g. one 64byte regular allocations with 63 external dcache names in a page (64x64=4096), which means the page is not freed even after reclaiming after all dcache names, and the process can thus "steal" the whole page with single 64byte allocation. If other kmalloc users similar to dcache external names become identified, they can also benefit from the new functionality simply by adding __GFP_RECLAIMABLE to the kmalloc calls. Side benefits of the patchset (that could be also merged separately) include removed branch for detecting __GFP_DMA kmalloc(), and shortening kmalloc cache names in /proc/slabinfo output. The latter is potentially an ABI break in case there are tools parsing the names and expecting the values to be in bytes. This is how /proc/slabinfo looks like after booting in virtme: ... kmalloc-rcl-4M 0 0 4194304 1 1024 : tunables 1 1 0 : slabdata 0 0 0 ... kmalloc-rcl-96 7 32 128 32 1 : tunables 120 60 8 : slabdata 1 1 0 kmalloc-rcl-64 25 128 64 64 1 : tunables 120 60 8 : slabdata 2 2 0 kmalloc-rcl-32 0 0 32 124 1 : tunables 120 60 8 : slabdata 0 0 0 kmalloc-4M 0 0 4194304 1 1024 : tunables 1 1 0 : slabdata 0 0 0 kmalloc-2M 0 0 2097152 1 512 : tunables 1 1 0 : slabdata 0 0 0 kmalloc-1M 0 0 1048576 1 256 : tunables 1 1 0 : slabdata 0 0 0 ... /proc/vmstat with renamed nr_indirectly_reclaimable_bytes counter: ... nr_slab_reclaimable 2817 nr_slab_unreclaimable 1781 ... nr_kernel_misc_reclaimable 0 ... /proc/meminfo with new KReclaimable counter: ... Shmem: 564 kB KReclaimable: 11260 kB Slab: 18368 kB SReclaimable: 11260 kB SUnreclaim: 7108 kB KernelStack: 1248 kB ... This patch (of 6): The kmalloc caches currently mainain separate (optional) array kmalloc_dma_caches for __GFP_DMA allocations. There are tests for __GFP_DMA in the allocation hotpaths. We can avoid the branches by combining kmalloc_caches and kmalloc_dma_caches into a single two-dimensional array where the outer dimension is cache "type". This will also allow to add kmalloc-reclaimable caches as a third type. Link: http://lkml.kernel.org/r/20180731090649.16028-2-vbabka@suse.cz Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Acked-by: Mel Gorman <mgorman@techsingularity.net> Acked-by: Christoph Lameter <cl@linux.com> Acked-by: Roman Gushchin <guro@fb.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: Laura Abbott <labbott@redhat.com> Cc: Sumit Semwal <sumit.semwal@linaro.org> Cc: Vijayanand Jitta <vjitta@codeaurora.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-27 06:05:34 +08:00
validate_slab_cache(kmalloc_caches[type][6]);
pr_err("\nB. Corruption after free\n");
p = kzalloc(128, GFP_KERNEL);
kfree(p);
*p = 0x78;
pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
mm, slab: combine kmalloc_caches and kmalloc_dma_caches Patch series "kmalloc-reclaimable caches", v4. As discussed at LSF/MM [1] here's a patchset that introduces kmalloc-reclaimable caches (more details in the second patch) and uses them for dcache external names. That allows us to repurpose the NR_INDIRECTLY_RECLAIMABLE_BYTES counter later in the series. With patch 3/6, dcache external names are allocated from kmalloc-rcl-* caches, eliminating the need for manual accounting. More importantly, it also ensures the reclaimable kmalloc allocations are grouped in pages separate from the regular kmalloc allocations. The need for proper accounting of dcache external names has shown it's easy for misbehaving process to allocate lots of them, causing premature OOMs. Without the added grouping, it's likely that a similar workload can interleave the dcache external names allocations with regular kmalloc allocations (note: I haven't searched myself for an example of such regular kmalloc allocation, but I would be very surprised if there wasn't some). A pathological case would be e.g. one 64byte regular allocations with 63 external dcache names in a page (64x64=4096), which means the page is not freed even after reclaiming after all dcache names, and the process can thus "steal" the whole page with single 64byte allocation. If other kmalloc users similar to dcache external names become identified, they can also benefit from the new functionality simply by adding __GFP_RECLAIMABLE to the kmalloc calls. Side benefits of the patchset (that could be also merged separately) include removed branch for detecting __GFP_DMA kmalloc(), and shortening kmalloc cache names in /proc/slabinfo output. The latter is potentially an ABI break in case there are tools parsing the names and expecting the values to be in bytes. This is how /proc/slabinfo looks like after booting in virtme: ... kmalloc-rcl-4M 0 0 4194304 1 1024 : tunables 1 1 0 : slabdata 0 0 0 ... kmalloc-rcl-96 7 32 128 32 1 : tunables 120 60 8 : slabdata 1 1 0 kmalloc-rcl-64 25 128 64 64 1 : tunables 120 60 8 : slabdata 2 2 0 kmalloc-rcl-32 0 0 32 124 1 : tunables 120 60 8 : slabdata 0 0 0 kmalloc-4M 0 0 4194304 1 1024 : tunables 1 1 0 : slabdata 0 0 0 kmalloc-2M 0 0 2097152 1 512 : tunables 1 1 0 : slabdata 0 0 0 kmalloc-1M 0 0 1048576 1 256 : tunables 1 1 0 : slabdata 0 0 0 ... /proc/vmstat with renamed nr_indirectly_reclaimable_bytes counter: ... nr_slab_reclaimable 2817 nr_slab_unreclaimable 1781 ... nr_kernel_misc_reclaimable 0 ... /proc/meminfo with new KReclaimable counter: ... Shmem: 564 kB KReclaimable: 11260 kB Slab: 18368 kB SReclaimable: 11260 kB SUnreclaim: 7108 kB KernelStack: 1248 kB ... This patch (of 6): The kmalloc caches currently mainain separate (optional) array kmalloc_dma_caches for __GFP_DMA allocations. There are tests for __GFP_DMA in the allocation hotpaths. We can avoid the branches by combining kmalloc_caches and kmalloc_dma_caches into a single two-dimensional array where the outer dimension is cache "type". This will also allow to add kmalloc-reclaimable caches as a third type. Link: http://lkml.kernel.org/r/20180731090649.16028-2-vbabka@suse.cz Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Acked-by: Mel Gorman <mgorman@techsingularity.net> Acked-by: Christoph Lameter <cl@linux.com> Acked-by: Roman Gushchin <guro@fb.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: Laura Abbott <labbott@redhat.com> Cc: Sumit Semwal <sumit.semwal@linaro.org> Cc: Vijayanand Jitta <vjitta@codeaurora.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-27 06:05:34 +08:00
validate_slab_cache(kmalloc_caches[type][7]);
p = kzalloc(256, GFP_KERNEL);
kfree(p);
p[50] = 0x9a;
pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
mm, slab: combine kmalloc_caches and kmalloc_dma_caches Patch series "kmalloc-reclaimable caches", v4. As discussed at LSF/MM [1] here's a patchset that introduces kmalloc-reclaimable caches (more details in the second patch) and uses them for dcache external names. That allows us to repurpose the NR_INDIRECTLY_RECLAIMABLE_BYTES counter later in the series. With patch 3/6, dcache external names are allocated from kmalloc-rcl-* caches, eliminating the need for manual accounting. More importantly, it also ensures the reclaimable kmalloc allocations are grouped in pages separate from the regular kmalloc allocations. The need for proper accounting of dcache external names has shown it's easy for misbehaving process to allocate lots of them, causing premature OOMs. Without the added grouping, it's likely that a similar workload can interleave the dcache external names allocations with regular kmalloc allocations (note: I haven't searched myself for an example of such regular kmalloc allocation, but I would be very surprised if there wasn't some). A pathological case would be e.g. one 64byte regular allocations with 63 external dcache names in a page (64x64=4096), which means the page is not freed even after reclaiming after all dcache names, and the process can thus "steal" the whole page with single 64byte allocation. If other kmalloc users similar to dcache external names become identified, they can also benefit from the new functionality simply by adding __GFP_RECLAIMABLE to the kmalloc calls. Side benefits of the patchset (that could be also merged separately) include removed branch for detecting __GFP_DMA kmalloc(), and shortening kmalloc cache names in /proc/slabinfo output. The latter is potentially an ABI break in case there are tools parsing the names and expecting the values to be in bytes. This is how /proc/slabinfo looks like after booting in virtme: ... kmalloc-rcl-4M 0 0 4194304 1 1024 : tunables 1 1 0 : slabdata 0 0 0 ... kmalloc-rcl-96 7 32 128 32 1 : tunables 120 60 8 : slabdata 1 1 0 kmalloc-rcl-64 25 128 64 64 1 : tunables 120 60 8 : slabdata 2 2 0 kmalloc-rcl-32 0 0 32 124 1 : tunables 120 60 8 : slabdata 0 0 0 kmalloc-4M 0 0 4194304 1 1024 : tunables 1 1 0 : slabdata 0 0 0 kmalloc-2M 0 0 2097152 1 512 : tunables 1 1 0 : slabdata 0 0 0 kmalloc-1M 0 0 1048576 1 256 : tunables 1 1 0 : slabdata 0 0 0 ... /proc/vmstat with renamed nr_indirectly_reclaimable_bytes counter: ... nr_slab_reclaimable 2817 nr_slab_unreclaimable 1781 ... nr_kernel_misc_reclaimable 0 ... /proc/meminfo with new KReclaimable counter: ... Shmem: 564 kB KReclaimable: 11260 kB Slab: 18368 kB SReclaimable: 11260 kB SUnreclaim: 7108 kB KernelStack: 1248 kB ... This patch (of 6): The kmalloc caches currently mainain separate (optional) array kmalloc_dma_caches for __GFP_DMA allocations. There are tests for __GFP_DMA in the allocation hotpaths. We can avoid the branches by combining kmalloc_caches and kmalloc_dma_caches into a single two-dimensional array where the outer dimension is cache "type". This will also allow to add kmalloc-reclaimable caches as a third type. Link: http://lkml.kernel.org/r/20180731090649.16028-2-vbabka@suse.cz Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Acked-by: Mel Gorman <mgorman@techsingularity.net> Acked-by: Christoph Lameter <cl@linux.com> Acked-by: Roman Gushchin <guro@fb.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: Laura Abbott <labbott@redhat.com> Cc: Sumit Semwal <sumit.semwal@linaro.org> Cc: Vijayanand Jitta <vjitta@codeaurora.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-27 06:05:34 +08:00
validate_slab_cache(kmalloc_caches[type][8]);
p = kzalloc(512, GFP_KERNEL);
kfree(p);
p[512] = 0xab;
pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
mm, slab: combine kmalloc_caches and kmalloc_dma_caches Patch series "kmalloc-reclaimable caches", v4. As discussed at LSF/MM [1] here's a patchset that introduces kmalloc-reclaimable caches (more details in the second patch) and uses them for dcache external names. That allows us to repurpose the NR_INDIRECTLY_RECLAIMABLE_BYTES counter later in the series. With patch 3/6, dcache external names are allocated from kmalloc-rcl-* caches, eliminating the need for manual accounting. More importantly, it also ensures the reclaimable kmalloc allocations are grouped in pages separate from the regular kmalloc allocations. The need for proper accounting of dcache external names has shown it's easy for misbehaving process to allocate lots of them, causing premature OOMs. Without the added grouping, it's likely that a similar workload can interleave the dcache external names allocations with regular kmalloc allocations (note: I haven't searched myself for an example of such regular kmalloc allocation, but I would be very surprised if there wasn't some). A pathological case would be e.g. one 64byte regular allocations with 63 external dcache names in a page (64x64=4096), which means the page is not freed even after reclaiming after all dcache names, and the process can thus "steal" the whole page with single 64byte allocation. If other kmalloc users similar to dcache external names become identified, they can also benefit from the new functionality simply by adding __GFP_RECLAIMABLE to the kmalloc calls. Side benefits of the patchset (that could be also merged separately) include removed branch for detecting __GFP_DMA kmalloc(), and shortening kmalloc cache names in /proc/slabinfo output. The latter is potentially an ABI break in case there are tools parsing the names and expecting the values to be in bytes. This is how /proc/slabinfo looks like after booting in virtme: ... kmalloc-rcl-4M 0 0 4194304 1 1024 : tunables 1 1 0 : slabdata 0 0 0 ... kmalloc-rcl-96 7 32 128 32 1 : tunables 120 60 8 : slabdata 1 1 0 kmalloc-rcl-64 25 128 64 64 1 : tunables 120 60 8 : slabdata 2 2 0 kmalloc-rcl-32 0 0 32 124 1 : tunables 120 60 8 : slabdata 0 0 0 kmalloc-4M 0 0 4194304 1 1024 : tunables 1 1 0 : slabdata 0 0 0 kmalloc-2M 0 0 2097152 1 512 : tunables 1 1 0 : slabdata 0 0 0 kmalloc-1M 0 0 1048576 1 256 : tunables 1 1 0 : slabdata 0 0 0 ... /proc/vmstat with renamed nr_indirectly_reclaimable_bytes counter: ... nr_slab_reclaimable 2817 nr_slab_unreclaimable 1781 ... nr_kernel_misc_reclaimable 0 ... /proc/meminfo with new KReclaimable counter: ... Shmem: 564 kB KReclaimable: 11260 kB Slab: 18368 kB SReclaimable: 11260 kB SUnreclaim: 7108 kB KernelStack: 1248 kB ... This patch (of 6): The kmalloc caches currently mainain separate (optional) array kmalloc_dma_caches for __GFP_DMA allocations. There are tests for __GFP_DMA in the allocation hotpaths. We can avoid the branches by combining kmalloc_caches and kmalloc_dma_caches into a single two-dimensional array where the outer dimension is cache "type". This will also allow to add kmalloc-reclaimable caches as a third type. Link: http://lkml.kernel.org/r/20180731090649.16028-2-vbabka@suse.cz Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Acked-by: Mel Gorman <mgorman@techsingularity.net> Acked-by: Christoph Lameter <cl@linux.com> Acked-by: Roman Gushchin <guro@fb.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: Laura Abbott <labbott@redhat.com> Cc: Sumit Semwal <sumit.semwal@linaro.org> Cc: Vijayanand Jitta <vjitta@codeaurora.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-27 06:05:34 +08:00
validate_slab_cache(kmalloc_caches[type][9]);
}
#else
#ifdef CONFIG_SYSFS
static void resiliency_test(void) {};
#endif
#endif /* SLUB_RESILIENCY_TEST */
#ifdef CONFIG_SYSFS
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
enum slab_stat_type {
SL_ALL, /* All slabs */
SL_PARTIAL, /* Only partially allocated slabs */
SL_CPU, /* Only slabs used for cpu caches */
SL_OBJECTS, /* Determine allocated objects not slabs */
SL_TOTAL /* Determine object capacity not slabs */
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
};
#define SO_ALL (1 << SL_ALL)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
#define SO_PARTIAL (1 << SL_PARTIAL)
#define SO_CPU (1 << SL_CPU)
#define SO_OBJECTS (1 << SL_OBJECTS)
#define SO_TOTAL (1 << SL_TOTAL)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
static ssize_t show_slab_objects(struct kmem_cache *s,
char *buf, unsigned long flags)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
{
unsigned long total = 0;
int node;
int x;
unsigned long *nodes;
int len = 0;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
treewide: kzalloc() -> kcalloc() The kzalloc() function has a 2-factor argument form, kcalloc(). This patch replaces cases of: kzalloc(a * b, gfp) with: kcalloc(a * b, gfp) as well as handling cases of: kzalloc(a * b * c, gfp) with: kzalloc(array3_size(a, b, c), gfp) as it's slightly less ugly than: kzalloc_array(array_size(a, b), c, gfp) This does, however, attempt to ignore constant size factors like: kzalloc(4 * 1024, gfp) though any constants defined via macros get caught up in the conversion. Any factors with a sizeof() of "unsigned char", "char", and "u8" were dropped, since they're redundant. The Coccinelle script used for this was: // Fix redundant parens around sizeof(). @@ type TYPE; expression THING, E; @@ ( kzalloc( - (sizeof(TYPE)) * E + sizeof(TYPE) * E , ...) | kzalloc( - (sizeof(THING)) * E + sizeof(THING) * E , ...) ) // Drop single-byte sizes and redundant parens. @@ expression COUNT; typedef u8; typedef __u8; @@ ( kzalloc( - sizeof(u8) * (COUNT) + COUNT , ...) | kzalloc( - sizeof(__u8) * (COUNT) + COUNT , ...) | kzalloc( - sizeof(char) * (COUNT) + COUNT , ...) | kzalloc( - sizeof(unsigned char) * (COUNT) + COUNT , ...) | kzalloc( - sizeof(u8) * COUNT + COUNT , ...) | kzalloc( - sizeof(__u8) * COUNT + COUNT , ...) | kzalloc( - sizeof(char) * COUNT + COUNT , ...) | kzalloc( - sizeof(unsigned char) * COUNT + COUNT , ...) ) // 2-factor product with sizeof(type/expression) and identifier or constant. @@ type TYPE; expression THING; identifier COUNT_ID; constant COUNT_CONST; @@ ( - kzalloc + kcalloc ( - sizeof(TYPE) * (COUNT_ID) + COUNT_ID, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * COUNT_ID + COUNT_ID, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * (COUNT_CONST) + COUNT_CONST, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * COUNT_CONST + COUNT_CONST, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * (COUNT_ID) + COUNT_ID, sizeof(THING) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * COUNT_ID + COUNT_ID, sizeof(THING) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * (COUNT_CONST) + COUNT_CONST, sizeof(THING) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * COUNT_CONST + COUNT_CONST, sizeof(THING) , ...) ) // 2-factor product, only identifiers. @@ identifier SIZE, COUNT; @@ - kzalloc + kcalloc ( - SIZE * COUNT + COUNT, SIZE , ...) // 3-factor product with 1 sizeof(type) or sizeof(expression), with // redundant parens removed. @@ expression THING; identifier STRIDE, COUNT; type TYPE; @@ ( kzalloc( - sizeof(TYPE) * (COUNT) * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc( - sizeof(TYPE) * (COUNT) * STRIDE + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc( - sizeof(TYPE) * COUNT * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc( - sizeof(TYPE) * COUNT * STRIDE + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kzalloc( - sizeof(THING) * (COUNT) * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kzalloc( - sizeof(THING) * (COUNT) * STRIDE + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kzalloc( - sizeof(THING) * COUNT * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kzalloc( - sizeof(THING) * COUNT * STRIDE + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) ) // 3-factor product with 2 sizeof(variable), with redundant parens removed. @@ expression THING1, THING2; identifier COUNT; type TYPE1, TYPE2; @@ ( kzalloc( - sizeof(TYPE1) * sizeof(TYPE2) * COUNT + array3_size(COUNT, sizeof(TYPE1), sizeof(TYPE2)) , ...) | kzalloc( - sizeof(TYPE1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(TYPE1), sizeof(TYPE2)) , ...) | kzalloc( - sizeof(THING1) * sizeof(THING2) * COUNT + array3_size(COUNT, sizeof(THING1), sizeof(THING2)) , ...) | kzalloc( - sizeof(THING1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(THING1), sizeof(THING2)) , ...) | kzalloc( - sizeof(TYPE1) * sizeof(THING2) * COUNT + array3_size(COUNT, sizeof(TYPE1), sizeof(THING2)) , ...) | kzalloc( - sizeof(TYPE1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(TYPE1), sizeof(THING2)) , ...) ) // 3-factor product, only identifiers, with redundant parens removed. @@ identifier STRIDE, SIZE, COUNT; @@ ( kzalloc( - (COUNT) * STRIDE * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - COUNT * (STRIDE) * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - COUNT * STRIDE * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - (COUNT) * (STRIDE) * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - COUNT * (STRIDE) * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - (COUNT) * STRIDE * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - (COUNT) * (STRIDE) * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kzalloc( - COUNT * STRIDE * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) ) // Any remaining multi-factor products, first at least 3-factor products, // when they're not all constants... @@ expression E1, E2, E3; constant C1, C2, C3; @@ ( kzalloc(C1 * C2 * C3, ...) | kzalloc( - (E1) * E2 * E3 + array3_size(E1, E2, E3) , ...) | kzalloc( - (E1) * (E2) * E3 + array3_size(E1, E2, E3) , ...) | kzalloc( - (E1) * (E2) * (E3) + array3_size(E1, E2, E3) , ...) | kzalloc( - E1 * E2 * E3 + array3_size(E1, E2, E3) , ...) ) // And then all remaining 2 factors products when they're not all constants, // keeping sizeof() as the second factor argument. @@ expression THING, E1, E2; type TYPE; constant C1, C2, C3; @@ ( kzalloc(sizeof(THING) * C2, ...) | kzalloc(sizeof(TYPE) * C2, ...) | kzalloc(C1 * C2 * C3, ...) | kzalloc(C1 * C2, ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * (E2) + E2, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(TYPE) * E2 + E2, sizeof(TYPE) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * (E2) + E2, sizeof(THING) , ...) | - kzalloc + kcalloc ( - sizeof(THING) * E2 + E2, sizeof(THING) , ...) | - kzalloc + kcalloc ( - (E1) * E2 + E1, E2 , ...) | - kzalloc + kcalloc ( - (E1) * (E2) + E1, E2 , ...) | - kzalloc + kcalloc ( - E1 * E2 + E1, E2 , ...) ) Signed-off-by: Kees Cook <keescook@chromium.org>
2018-06-13 05:03:40 +08:00
nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
if (!nodes)
return -ENOMEM;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (flags & SO_CPU) {
int cpu;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
for_each_possible_cpu(cpu) {
struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
cpu);
int node;
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
struct page *page;
SLUB: Avoid page struct cacheline bouncing due to remote frees to cpu slab A remote free may access the same page struct that also contains the lockless freelist for the cpu slab. If objects have a short lifetime and are freed by a different processor then remote frees back to the slab from which we are currently allocating are frequent. The cacheline with the page struct needs to be repeately acquired in exclusive mode by both the allocating thread and the freeing thread. If this is frequent enough then performance will suffer because of cacheline bouncing. This patchset puts the lockless_freelist pointer in its own cacheline. In order to make that happen we introduce a per cpu structure called kmem_cache_cpu. Instead of keeping an array of pointers to page structs we now keep an array to a per cpu structure that--among other things--contains the pointer to the lockless freelist. The freeing thread can then keep possession of exclusive access to the page struct cacheline while the allocating thread keeps its exclusive access to the cacheline containing the per cpu structure. This works as long as the allocating cpu is able to service its request from the lockless freelist. If the lockless freelist runs empty then the allocating thread needs to acquire exclusive access to the cacheline with the page struct lock the slab. The allocating thread will then check if new objects were freed to the per cpu slab. If so it will keep the slab as the cpu slab and continue with the recently remote freed objects. So the allocating thread can take a series of just freed remote pages and dish them out again. Ideally allocations could be just recycling objects in the same slab this way which will lead to an ideal allocation / remote free pattern. The number of objects that can be handled in this way is limited by the capacity of one slab. Increasing slab size via slub_min_objects/ slub_max_order may increase the number of objects and therefore performance. If the allocating thread runs out of objects and finds that no objects were put back by the remote processor then it will retrieve a new slab (from the partial lists or from the page allocator) and start with a whole new set of objects while the remote thread may still be freeing objects to the old cpu slab. This may then repeat until the new slab is also exhausted. If remote freeing has freed objects in the earlier slab then that earlier slab will now be on the partial freelist and the allocating thread will pick that slab next for allocation. So the loop is extended. However, both threads need to take the list_lock to make the swizzling via the partial list happen. It is likely that this kind of scheme will keep the objects being passed around to a small set that can be kept in the cpu caches leading to increased performance. More code cleanups become possible: - Instead of passing a cpu we can now pass a kmem_cache_cpu structure around. Allows reducing the number of parameters to various functions. - Can define a new node_match() function for NUMA to encapsulate locality checks. Effect on allocations: Cachelines touched before this patch: Write: page cache struct and first cacheline of object Cachelines touched after this patch: Write: kmem_cache_cpu cacheline and first cacheline of object Read: page cache struct (but see later patch that avoids touching that cacheline) The handling when the lockless alloc list runs empty gets to be a bit more complicated since another cacheline has now to be written to. But that is halfway out of the hot path. Effect on freeing: Cachelines touched before this patch: Write: page_struct and first cacheline of object Cachelines touched after this patch depending on how we free: Write(to cpu_slab): kmem_cache_cpu struct and first cacheline of object Write(to other): page struct and first cacheline of object Read(to cpu_slab): page struct to id slab etc. (but see later patch that avoids touching the page struct on free) Read(to other): cpu local kmem_cache_cpu struct to verify its not the cpu slab. Summary: Pro: - Distinct cachelines so that concurrent remote frees and local allocs on a cpuslab can occur without cacheline bouncing. - Avoids potential bouncing cachelines because of neighboring per cpu pointer updates in kmem_cache's cpu_slab structure since it now grows to a cacheline (Therefore remove the comment that talks about that concern). Cons: - Freeing objects now requires the reading of one additional cacheline. That can be mitigated for some cases by the following patches but its not possible to completely eliminate these references. - Memory usage grows slightly. The size of each per cpu object is blown up from one word (pointing to the page_struct) to one cacheline with various data. So this is NR_CPUS*NR_SLABS*L1_BYTES more memory use. Lets say NR_SLABS is 100 and a cache line size of 128 then we have just increased SLAB metadata requirements by 12.8k per cpu. (Another later patch reduces these requirements) Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-16 16:26:05 +08:00
page = READ_ONCE(c->page);
if (!page)
continue;
node = page_to_nid(page);
if (flags & SO_TOTAL)
x = page->objects;
else if (flags & SO_OBJECTS)
x = page->inuse;
else
x = 1;
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
total += x;
nodes[node] += x;
page = slub_percpu_partial_read_once(c);
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
if (page) {
node = page_to_nid(page);
if (flags & SO_TOTAL)
WARN_ON_ONCE(1);
else if (flags & SO_OBJECTS)
WARN_ON_ONCE(1);
else
x = page->pages;
total += x;
nodes[node] += x;
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
}
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
}
mm/slub: fix a deadlock in show_slab_objects() A long time ago we fixed a similar deadlock in show_slab_objects() [1]. However, it is apparently due to the commits like 01fb58bcba63 ("slab: remove synchronous synchronize_sched() from memcg cache deactivation path") and 03afc0e25f7f ("slab: get_online_mems for kmem_cache_{create,destroy,shrink}"), this kind of deadlock is back by just reading files in /sys/kernel/slab which will generate a lockdep splat below. Since the "mem_hotplug_lock" here is only to obtain a stable online node mask while racing with NUMA node hotplug, in the worst case, the results may me miscalculated while doing NUMA node hotplug, but they shall be corrected by later reads of the same files. WARNING: possible circular locking dependency detected ------------------------------------------------------ cat/5224 is trying to acquire lock: ffff900012ac3120 (mem_hotplug_lock.rw_sem){++++}, at: show_slab_objects+0x94/0x3a8 but task is already holding lock: b8ff009693eee398 (kn->count#45){++++}, at: kernfs_seq_start+0x44/0xf0 which lock already depends on the new lock. the existing dependency chain (in reverse order) is: -> #2 (kn->count#45){++++}: lock_acquire+0x31c/0x360 __kernfs_remove+0x290/0x490 kernfs_remove+0x30/0x44 sysfs_remove_dir+0x70/0x88 kobject_del+0x50/0xb0 sysfs_slab_unlink+0x2c/0x38 shutdown_cache+0xa0/0xf0 kmemcg_cache_shutdown_fn+0x1c/0x34 kmemcg_workfn+0x44/0x64 process_one_work+0x4f4/0x950 worker_thread+0x390/0x4bc kthread+0x1cc/0x1e8 ret_from_fork+0x10/0x18 -> #1 (slab_mutex){+.+.}: lock_acquire+0x31c/0x360 __mutex_lock_common+0x16c/0xf78 mutex_lock_nested+0x40/0x50 memcg_create_kmem_cache+0x38/0x16c memcg_kmem_cache_create_func+0x3c/0x70 process_one_work+0x4f4/0x950 worker_thread+0x390/0x4bc kthread+0x1cc/0x1e8 ret_from_fork+0x10/0x18 -> #0 (mem_hotplug_lock.rw_sem){++++}: validate_chain+0xd10/0x2bcc __lock_acquire+0x7f4/0xb8c lock_acquire+0x31c/0x360 get_online_mems+0x54/0x150 show_slab_objects+0x94/0x3a8 total_objects_show+0x28/0x34 slab_attr_show+0x38/0x54 sysfs_kf_seq_show+0x198/0x2d4 kernfs_seq_show+0xa4/0xcc seq_read+0x30c/0x8a8 kernfs_fop_read+0xa8/0x314 __vfs_read+0x88/0x20c vfs_read+0xd8/0x10c ksys_read+0xb0/0x120 __arm64_sys_read+0x54/0x88 el0_svc_handler+0x170/0x240 el0_svc+0x8/0xc other info that might help us debug this: Chain exists of: mem_hotplug_lock.rw_sem --> slab_mutex --> kn->count#45 Possible unsafe locking scenario: CPU0 CPU1 ---- ---- lock(kn->count#45); lock(slab_mutex); lock(kn->count#45); lock(mem_hotplug_lock.rw_sem); *** DEADLOCK *** 3 locks held by cat/5224: #0: 9eff00095b14b2a0 (&p->lock){+.+.}, at: seq_read+0x4c/0x8a8 #1: 0eff008997041480 (&of->mutex){+.+.}, at: kernfs_seq_start+0x34/0xf0 #2: b8ff009693eee398 (kn->count#45){++++}, at: kernfs_seq_start+0x44/0xf0 stack backtrace: Call trace: dump_backtrace+0x0/0x248 show_stack+0x20/0x2c dump_stack+0xd0/0x140 print_circular_bug+0x368/0x380 check_noncircular+0x248/0x250 validate_chain+0xd10/0x2bcc __lock_acquire+0x7f4/0xb8c lock_acquire+0x31c/0x360 get_online_mems+0x54/0x150 show_slab_objects+0x94/0x3a8 total_objects_show+0x28/0x34 slab_attr_show+0x38/0x54 sysfs_kf_seq_show+0x198/0x2d4 kernfs_seq_show+0xa4/0xcc seq_read+0x30c/0x8a8 kernfs_fop_read+0xa8/0x314 __vfs_read+0x88/0x20c vfs_read+0xd8/0x10c ksys_read+0xb0/0x120 __arm64_sys_read+0x54/0x88 el0_svc_handler+0x170/0x240 el0_svc+0x8/0xc I think it is important to mention that this doesn't expose the show_slab_objects to use-after-free. There is only a single path that might really race here and that is the slab hotplug notifier callback __kmem_cache_shrink (via slab_mem_going_offline_callback) but that path doesn't really destroy kmem_cache_node data structures. [1] http://lkml.iu.edu/hypermail/linux/kernel/1101.0/02850.html [akpm@linux-foundation.org: add comment explaining why we don't need mem_hotplug_lock] Link: http://lkml.kernel.org/r/1570192309-10132-1-git-send-email-cai@lca.pw Fixes: 01fb58bcba63 ("slab: remove synchronous synchronize_sched() from memcg cache deactivation path") Fixes: 03afc0e25f7f ("slab: get_online_mems for kmem_cache_{create,destroy,shrink}") Signed-off-by: Qian Cai <cai@lca.pw> Acked-by: Michal Hocko <mhocko@suse.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Tejun Heo <tj@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Cc: Roman Gushchin <guro@fb.com> Cc: <stable@vger.kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-10-15 05:11:51 +08:00
/*
* It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
* already held which will conflict with an existing lock order:
*
* mem_hotplug_lock->slab_mutex->kernfs_mutex
*
* We don't really need mem_hotplug_lock (to hold off
* slab_mem_going_offline_callback) here because slab's memory hot
* unplug code doesn't destroy the kmem_cache->node[] data.
*/
#ifdef CONFIG_SLUB_DEBUG
if (flags & SO_ALL) {
struct kmem_cache_node *n;
for_each_kmem_cache_node(s, node, n) {
if (flags & SO_TOTAL)
x = atomic_long_read(&n->total_objects);
else if (flags & SO_OBJECTS)
x = atomic_long_read(&n->total_objects) -
count_partial(n, count_free);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
else
x = atomic_long_read(&n->nr_slabs);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
total += x;
nodes[node] += x;
}
} else
#endif
if (flags & SO_PARTIAL) {
struct kmem_cache_node *n;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
for_each_kmem_cache_node(s, node, n) {
if (flags & SO_TOTAL)
x = count_partial(n, count_total);
else if (flags & SO_OBJECTS)
x = count_partial(n, count_inuse);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
else
x = n->nr_partial;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
total += x;
nodes[node] += x;
}
}
len += sysfs_emit_at(buf, len, "%lu", total);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
#ifdef CONFIG_NUMA
for (node = 0; node < nr_node_ids; node++) {
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (nodes[node])
len += sysfs_emit_at(buf, len, " N%d=%lu",
node, nodes[node]);
}
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
#endif
len += sysfs_emit_at(buf, len, "\n");
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
kfree(nodes);
return len;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
#define to_slab(n) container_of(n, struct kmem_cache, kobj)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
struct slab_attribute {
struct attribute attr;
ssize_t (*show)(struct kmem_cache *s, char *buf);
ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
};
#define SLAB_ATTR_RO(_name) \
mm: restrict access to slab files under procfs and sysfs Historically /proc/slabinfo and files under /sys/kernel/slab/* have world read permissions and are accessible to the world. slabinfo contains rather private information related both to the kernel and userspace tasks. Depending on the situation, it might reveal either private information per se or information useful to make another targeted attack. Some examples of what can be learned by reading/watching for /proc/slabinfo entries: 1) dentry (and different *inode*) number might reveal other processes fs activity. The number of dentry "active objects" doesn't strictly show file count opened/touched by a process, however, there is a good correlation between them. The patch "proc: force dcache drop on unauthorized access" relies on the privacy of dentry count. 2) different inode entries might reveal the same information as (1), but these are more fine granted counters. If a filesystem is mounted in a private mount point (or even a private namespace) and fs type differs from other mounted fs types, fs activity in this mount point/namespace is revealed. If there is a single ecryptfs mount point, the whole fs activity of a single user is revealed. Number of files in ecryptfs mount point is a private information per se. 3) fuse_* reveals number of files / fs activity of a user in a user private mount point. It is approx. the same severity as ecryptfs infoleak in (2). 4) sysfs_dir_cache similar to (2) reveals devices' addition/removal, which can be otherwise hidden by "chmod 0700 /sys/". With 0444 slabinfo the precise number of sysfs files is known to the world. 5) buffer_head might reveal some kernel activity. With other information leaks an attacker might identify what specific kernel routines generate buffer_head activity. 6) *kmalloc* infoleaks are very situational. Attacker should watch for the specific kmalloc size entry and filter the noise related to the unrelated kernel activity. If an attacker has relatively silent victim system, he might get rather precise counters. Additional information sources might significantly increase the slabinfo infoleak benefits. E.g. if an attacker knows that the processes activity on the system is very low (only core daemons like syslog and cron), he may run setxid binaries / trigger local daemon activity / trigger network services activity / await sporadic cron jobs activity / etc. and get rather precise counters for fs and network activity of these privileged tasks, which is unknown otherwise. Also hiding slabinfo and /sys/kernel/slab/* is a one step to complicate exploitation of kernel heap overflows (and possibly, other bugs). The related discussion: http://thread.gmane.org/gmane.linux.kernel/1108378 To keep compatibility with old permission model where non-root monitoring daemon could watch for kernel memleaks though slabinfo one should do: groupadd slabinfo usermod -a -G slabinfo $MONITOR_USER And add the following commands to init scripts (to mountall.conf in Ubuntu's upstart case): chmod g+r /proc/slabinfo /sys/kernel/slab/*/* chgrp slabinfo /proc/slabinfo /sys/kernel/slab/*/* Signed-off-by: Vasiliy Kulikov <segoon@openwall.com> Reviewed-by: Kees Cook <kees@ubuntu.com> Reviewed-by: Dave Hansen <dave@linux.vnet.ibm.com> Acked-by: Christoph Lameter <cl@gentwo.org> Acked-by: David Rientjes <rientjes@google.com> CC: Valdis.Kletnieks@vt.edu CC: Linus Torvalds <torvalds@linux-foundation.org> CC: Alan Cox <alan@linux.intel.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-09-28 01:54:53 +08:00
static struct slab_attribute _name##_attr = \
__ATTR(_name, 0400, _name##_show, NULL)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
#define SLAB_ATTR(_name) \
static struct slab_attribute _name##_attr = \
mm: restrict access to slab files under procfs and sysfs Historically /proc/slabinfo and files under /sys/kernel/slab/* have world read permissions and are accessible to the world. slabinfo contains rather private information related both to the kernel and userspace tasks. Depending on the situation, it might reveal either private information per se or information useful to make another targeted attack. Some examples of what can be learned by reading/watching for /proc/slabinfo entries: 1) dentry (and different *inode*) number might reveal other processes fs activity. The number of dentry "active objects" doesn't strictly show file count opened/touched by a process, however, there is a good correlation between them. The patch "proc: force dcache drop on unauthorized access" relies on the privacy of dentry count. 2) different inode entries might reveal the same information as (1), but these are more fine granted counters. If a filesystem is mounted in a private mount point (or even a private namespace) and fs type differs from other mounted fs types, fs activity in this mount point/namespace is revealed. If there is a single ecryptfs mount point, the whole fs activity of a single user is revealed. Number of files in ecryptfs mount point is a private information per se. 3) fuse_* reveals number of files / fs activity of a user in a user private mount point. It is approx. the same severity as ecryptfs infoleak in (2). 4) sysfs_dir_cache similar to (2) reveals devices' addition/removal, which can be otherwise hidden by "chmod 0700 /sys/". With 0444 slabinfo the precise number of sysfs files is known to the world. 5) buffer_head might reveal some kernel activity. With other information leaks an attacker might identify what specific kernel routines generate buffer_head activity. 6) *kmalloc* infoleaks are very situational. Attacker should watch for the specific kmalloc size entry and filter the noise related to the unrelated kernel activity. If an attacker has relatively silent victim system, he might get rather precise counters. Additional information sources might significantly increase the slabinfo infoleak benefits. E.g. if an attacker knows that the processes activity on the system is very low (only core daemons like syslog and cron), he may run setxid binaries / trigger local daemon activity / trigger network services activity / await sporadic cron jobs activity / etc. and get rather precise counters for fs and network activity of these privileged tasks, which is unknown otherwise. Also hiding slabinfo and /sys/kernel/slab/* is a one step to complicate exploitation of kernel heap overflows (and possibly, other bugs). The related discussion: http://thread.gmane.org/gmane.linux.kernel/1108378 To keep compatibility with old permission model where non-root monitoring daemon could watch for kernel memleaks though slabinfo one should do: groupadd slabinfo usermod -a -G slabinfo $MONITOR_USER And add the following commands to init scripts (to mountall.conf in Ubuntu's upstart case): chmod g+r /proc/slabinfo /sys/kernel/slab/*/* chgrp slabinfo /proc/slabinfo /sys/kernel/slab/*/* Signed-off-by: Vasiliy Kulikov <segoon@openwall.com> Reviewed-by: Kees Cook <kees@ubuntu.com> Reviewed-by: Dave Hansen <dave@linux.vnet.ibm.com> Acked-by: Christoph Lameter <cl@gentwo.org> Acked-by: David Rientjes <rientjes@google.com> CC: Valdis.Kletnieks@vt.edu CC: Linus Torvalds <torvalds@linux-foundation.org> CC: Alan Cox <alan@linux.intel.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-09-28 01:54:53 +08:00
__ATTR(_name, 0600, _name##_show, _name##_store)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
{
return sysfs_emit(buf, "%u\n", s->size);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
SLAB_ATTR_RO(slab_size);
static ssize_t align_show(struct kmem_cache *s, char *buf)
{
return sysfs_emit(buf, "%u\n", s->align);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
SLAB_ATTR_RO(align);
static ssize_t object_size_show(struct kmem_cache *s, char *buf)
{
return sysfs_emit(buf, "%u\n", s->object_size);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
SLAB_ATTR_RO(object_size);
static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
{
return sysfs_emit(buf, "%u\n", oo_objects(s->oo));
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
SLAB_ATTR_RO(objs_per_slab);
static ssize_t order_show(struct kmem_cache *s, char *buf)
{
return sysfs_emit(buf, "%u\n", oo_order(s->oo));
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
mm, slub: remove runtime allocation order changes SLUB allows runtime changing of page allocation order by writing into the /sys/kernel/slab/<cache>/order file. Jann has reported [1] that this interface allows the order to be set too small, leading to crashes. While it's possible to fix the immediate issue, closer inspection reveals potential races. Storing the new order calls calculate_sizes() which non-atomically updates a lot of kmem_cache fields while the cache is still in use. Unexpected behavior might occur even if the fields are set to the same value as they were. This could be fixed by splitting out the part of calculate_sizes() that depends on forced_order, so that we only update kmem_cache.oo field. This could still race with init_cache_random_seq(), shuffle_freelist(), allocate_slab(). Perhaps it's possible to audit and e.g. add some READ_ONCE/WRITE_ONCE accesses, it might be easier just to remove the runtime order changes, which is what this patch does. If there are valid usecases for per-cache order setting, we could e.g. extend the boot parameters to do that. [1] https://lore.kernel.org/r/CAG48ez31PP--h6_FzVyfJ4H86QYczAFPdxtJHUEEan+7VJETAQ@mail.gmail.com Reported-by: Jann Horn <jannh@google.com> Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Kees Cook <keescook@chromium.org> Acked-by: Christoph Lameter <cl@linux.com> Acked-by: Roman Gushchin <guro@fb.com> Cc: Vijayanand Jitta <vjitta@codeaurora.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Pekka Enberg <penberg@kernel.org> Link: http://lkml.kernel.org/r/20200610163135.17364-4-vbabka@suse.cz Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-07 14:18:41 +08:00
SLAB_ATTR_RO(order);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
slub: add min_partial sysfs tunable Now that a cache's min_partial has been moved to struct kmem_cache, it's possible to easily tune it from userspace by adding a sysfs attribute. It may not be desirable to keep a large number of partial slabs around if a cache is used infrequently and memory, especially when constrained by a cgroup, is scarce. It's better to allow userspace to set the minimum policy per cache instead of relying explicitly on kmem_cache_shrink(). The memory savings from simply moving min_partial from struct kmem_cache_node to struct kmem_cache is obviously not significant (unless maybe you're from SGI or something), at the largest it's # allocated caches * (MAX_NUMNODES - 1) * sizeof(unsigned long) The true savings occurs when userspace reduces the number of partial slabs that would otherwise be wasted, especially on machines with a large number of nodes (ia64 with CONFIG_NODES_SHIFT at 10 for default?). As well as the kernel estimates ideal values for n->min_partial and ensures it's within a sane range, userspace has no other input other than writing to /sys/kernel/slab/cache/shrink. There simply isn't any better heuristic to add when calculating the partial values for a better estimate that works for all possible caches. And since it's currently a static value, the user really has no way of reclaiming that wasted space, which can be significant when constrained by a cgroup (either cpusets or, later, memory controller slab limits) without shrinking it entirely. This also allows the user to specify that increased fragmentation and more partial slabs are actually desired to avoid the cost of allocating new slabs at runtime for specific caches. There's also no reason why this should be a per-struct kmem_cache_node value in the first place. You could argue that a machine would have such node size asymmetries that it should be specified on a per-node basis, but we know nobody is doing that right now since it's a purely static value at the moment and there's no convenient way to tune that via slub's sysfs interface. Cc: Christoph Lameter <cl@linux-foundation.org> Signed-off-by: David Rientjes <rientjes@google.com> Signed-off-by: Pekka Enberg <penberg@cs.helsinki.fi>
2009-02-23 09:40:09 +08:00
static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
{
return sysfs_emit(buf, "%lu\n", s->min_partial);
slub: add min_partial sysfs tunable Now that a cache's min_partial has been moved to struct kmem_cache, it's possible to easily tune it from userspace by adding a sysfs attribute. It may not be desirable to keep a large number of partial slabs around if a cache is used infrequently and memory, especially when constrained by a cgroup, is scarce. It's better to allow userspace to set the minimum policy per cache instead of relying explicitly on kmem_cache_shrink(). The memory savings from simply moving min_partial from struct kmem_cache_node to struct kmem_cache is obviously not significant (unless maybe you're from SGI or something), at the largest it's # allocated caches * (MAX_NUMNODES - 1) * sizeof(unsigned long) The true savings occurs when userspace reduces the number of partial slabs that would otherwise be wasted, especially on machines with a large number of nodes (ia64 with CONFIG_NODES_SHIFT at 10 for default?). As well as the kernel estimates ideal values for n->min_partial and ensures it's within a sane range, userspace has no other input other than writing to /sys/kernel/slab/cache/shrink. There simply isn't any better heuristic to add when calculating the partial values for a better estimate that works for all possible caches. And since it's currently a static value, the user really has no way of reclaiming that wasted space, which can be significant when constrained by a cgroup (either cpusets or, later, memory controller slab limits) without shrinking it entirely. This also allows the user to specify that increased fragmentation and more partial slabs are actually desired to avoid the cost of allocating new slabs at runtime for specific caches. There's also no reason why this should be a per-struct kmem_cache_node value in the first place. You could argue that a machine would have such node size asymmetries that it should be specified on a per-node basis, but we know nobody is doing that right now since it's a purely static value at the moment and there's no convenient way to tune that via slub's sysfs interface. Cc: Christoph Lameter <cl@linux-foundation.org> Signed-off-by: David Rientjes <rientjes@google.com> Signed-off-by: Pekka Enberg <penberg@cs.helsinki.fi>
2009-02-23 09:40:09 +08:00
}
static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
size_t length)
{
unsigned long min;
int err;
err = kstrtoul(buf, 10, &min);
slub: add min_partial sysfs tunable Now that a cache's min_partial has been moved to struct kmem_cache, it's possible to easily tune it from userspace by adding a sysfs attribute. It may not be desirable to keep a large number of partial slabs around if a cache is used infrequently and memory, especially when constrained by a cgroup, is scarce. It's better to allow userspace to set the minimum policy per cache instead of relying explicitly on kmem_cache_shrink(). The memory savings from simply moving min_partial from struct kmem_cache_node to struct kmem_cache is obviously not significant (unless maybe you're from SGI or something), at the largest it's # allocated caches * (MAX_NUMNODES - 1) * sizeof(unsigned long) The true savings occurs when userspace reduces the number of partial slabs that would otherwise be wasted, especially on machines with a large number of nodes (ia64 with CONFIG_NODES_SHIFT at 10 for default?). As well as the kernel estimates ideal values for n->min_partial and ensures it's within a sane range, userspace has no other input other than writing to /sys/kernel/slab/cache/shrink. There simply isn't any better heuristic to add when calculating the partial values for a better estimate that works for all possible caches. And since it's currently a static value, the user really has no way of reclaiming that wasted space, which can be significant when constrained by a cgroup (either cpusets or, later, memory controller slab limits) without shrinking it entirely. This also allows the user to specify that increased fragmentation and more partial slabs are actually desired to avoid the cost of allocating new slabs at runtime for specific caches. There's also no reason why this should be a per-struct kmem_cache_node value in the first place. You could argue that a machine would have such node size asymmetries that it should be specified on a per-node basis, but we know nobody is doing that right now since it's a purely static value at the moment and there's no convenient way to tune that via slub's sysfs interface. Cc: Christoph Lameter <cl@linux-foundation.org> Signed-off-by: David Rientjes <rientjes@google.com> Signed-off-by: Pekka Enberg <penberg@cs.helsinki.fi>
2009-02-23 09:40:09 +08:00
if (err)
return err;
set_min_partial(s, min);
slub: add min_partial sysfs tunable Now that a cache's min_partial has been moved to struct kmem_cache, it's possible to easily tune it from userspace by adding a sysfs attribute. It may not be desirable to keep a large number of partial slabs around if a cache is used infrequently and memory, especially when constrained by a cgroup, is scarce. It's better to allow userspace to set the minimum policy per cache instead of relying explicitly on kmem_cache_shrink(). The memory savings from simply moving min_partial from struct kmem_cache_node to struct kmem_cache is obviously not significant (unless maybe you're from SGI or something), at the largest it's # allocated caches * (MAX_NUMNODES - 1) * sizeof(unsigned long) The true savings occurs when userspace reduces the number of partial slabs that would otherwise be wasted, especially on machines with a large number of nodes (ia64 with CONFIG_NODES_SHIFT at 10 for default?). As well as the kernel estimates ideal values for n->min_partial and ensures it's within a sane range, userspace has no other input other than writing to /sys/kernel/slab/cache/shrink. There simply isn't any better heuristic to add when calculating the partial values for a better estimate that works for all possible caches. And since it's currently a static value, the user really has no way of reclaiming that wasted space, which can be significant when constrained by a cgroup (either cpusets or, later, memory controller slab limits) without shrinking it entirely. This also allows the user to specify that increased fragmentation and more partial slabs are actually desired to avoid the cost of allocating new slabs at runtime for specific caches. There's also no reason why this should be a per-struct kmem_cache_node value in the first place. You could argue that a machine would have such node size asymmetries that it should be specified on a per-node basis, but we know nobody is doing that right now since it's a purely static value at the moment and there's no convenient way to tune that via slub's sysfs interface. Cc: Christoph Lameter <cl@linux-foundation.org> Signed-off-by: David Rientjes <rientjes@google.com> Signed-off-by: Pekka Enberg <penberg@cs.helsinki.fi>
2009-02-23 09:40:09 +08:00
return length;
}
SLAB_ATTR(min_partial);
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
{
return sysfs_emit(buf, "%u\n", slub_cpu_partial(s));
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
}
static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
size_t length)
{
unsigned int objects;
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
int err;
err = kstrtouint(buf, 10, &objects);
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
if (err)
return err;
if (objects && !kmem_cache_has_cpu_partial(s))
return -EINVAL;
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
slub_set_cpu_partial(s, objects);
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
flush_all(s);
return length;
}
SLAB_ATTR(cpu_partial);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
static ssize_t ctor_show(struct kmem_cache *s, char *buf)
{
if (!s->ctor)
return 0;
return sysfs_emit(buf, "%pS\n", s->ctor);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
SLAB_ATTR_RO(ctor);
static ssize_t aliases_show(struct kmem_cache *s, char *buf)
{
return sysfs_emit(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
SLAB_ATTR_RO(aliases);
static ssize_t partial_show(struct kmem_cache *s, char *buf)
{
return show_slab_objects(s, buf, SO_PARTIAL);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
SLAB_ATTR_RO(partial);
static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
{
return show_slab_objects(s, buf, SO_CPU);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
SLAB_ATTR_RO(cpu_slabs);
static ssize_t objects_show(struct kmem_cache *s, char *buf)
{
return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
SLAB_ATTR_RO(objects);
static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
{
return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
}
SLAB_ATTR_RO(objects_partial);
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
{
int objects = 0;
int pages = 0;
int cpu;
int len = 0;
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
for_each_online_cpu(cpu) {
struct page *page;
page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
if (page) {
pages += page->pages;
objects += page->pobjects;
}
}
len += sysfs_emit_at(buf, len, "%d(%d)", objects, pages);
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
#ifdef CONFIG_SMP
for_each_online_cpu(cpu) {
struct page *page;
page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
if (page)
len += sysfs_emit_at(buf, len, " C%d=%d(%d)",
cpu, page->pobjects, page->pages);
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
}
#endif
len += sysfs_emit_at(buf, len, "\n");
return len;
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
}
SLAB_ATTR_RO(slabs_cpu_partial);
static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
{
return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
}
SLAB_ATTR_RO(reclaim_account);
static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
{
return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
}
SLAB_ATTR_RO(hwcache_align);
#ifdef CONFIG_ZONE_DMA
static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
{
return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
}
SLAB_ATTR_RO(cache_dma);
#endif
usercopy: Prepare for usercopy whitelisting This patch prepares the slab allocator to handle caches having annotations (useroffset and usersize) defining usercopy regions. This patch is modified from Brad Spengler/PaX Team's PAX_USERCOPY whitelisting code in the last public patch of grsecurity/PaX based on my understanding of the code. Changes or omissions from the original code are mine and don't reflect the original grsecurity/PaX code. Currently, hardened usercopy performs dynamic bounds checking on slab cache objects. This is good, but still leaves a lot of kernel memory available to be copied to/from userspace in the face of bugs. To further restrict what memory is available for copying, this creates a way to whitelist specific areas of a given slab cache object for copying to/from userspace, allowing much finer granularity of access control. Slab caches that are never exposed to userspace can declare no whitelist for their objects, thereby keeping them unavailable to userspace via dynamic copy operations. (Note, an implicit form of whitelisting is the use of constant sizes in usercopy operations and get_user()/put_user(); these bypass hardened usercopy checks since these sizes cannot change at runtime.) To support this whitelist annotation, usercopy region offset and size members are added to struct kmem_cache. The slab allocator receives a new function, kmem_cache_create_usercopy(), that creates a new cache with a usercopy region defined, suitable for declaring spans of fields within the objects that get copied to/from userspace. In this patch, the default kmem_cache_create() marks the entire allocation as whitelisted, leaving it semantically unchanged. Once all fine-grained whitelists have been added (in subsequent patches), this will be changed to a usersize of 0, making caches created with kmem_cache_create() not copyable to/from userspace. After the entire usercopy whitelist series is applied, less than 15% of the slab cache memory remains exposed to potential usercopy bugs after a fresh boot: Total Slab Memory: 48074720 Usercopyable Memory: 6367532 13.2% task_struct 0.2% 4480/1630720 RAW 0.3% 300/96000 RAWv6 2.1% 1408/64768 ext4_inode_cache 3.0% 269760/8740224 dentry 11.1% 585984/5273856 mm_struct 29.1% 54912/188448 kmalloc-8 100.0% 24576/24576 kmalloc-16 100.0% 28672/28672 kmalloc-32 100.0% 81920/81920 kmalloc-192 100.0% 96768/96768 kmalloc-128 100.0% 143360/143360 names_cache 100.0% 163840/163840 kmalloc-64 100.0% 167936/167936 kmalloc-256 100.0% 339968/339968 kmalloc-512 100.0% 350720/350720 kmalloc-96 100.0% 455616/455616 kmalloc-8192 100.0% 655360/655360 kmalloc-1024 100.0% 812032/812032 kmalloc-4096 100.0% 819200/819200 kmalloc-2048 100.0% 1310720/1310720 After some kernel build workloads, the percentage (mainly driven by dentry and inode caches expanding) drops under 10%: Total Slab Memory: 95516184 Usercopyable Memory: 8497452 8.8% task_struct 0.2% 4000/1456000 RAW 0.3% 300/96000 RAWv6 2.1% 1408/64768 ext4_inode_cache 3.0% 1217280/39439872 dentry 11.1% 1623200/14608800 mm_struct 29.1% 73216/251264 kmalloc-8 100.0% 24576/24576 kmalloc-16 100.0% 28672/28672 kmalloc-32 100.0% 94208/94208 kmalloc-192 100.0% 96768/96768 kmalloc-128 100.0% 143360/143360 names_cache 100.0% 163840/163840 kmalloc-64 100.0% 245760/245760 kmalloc-256 100.0% 339968/339968 kmalloc-512 100.0% 350720/350720 kmalloc-96 100.0% 563520/563520 kmalloc-8192 100.0% 655360/655360 kmalloc-1024 100.0% 794624/794624 kmalloc-4096 100.0% 819200/819200 kmalloc-2048 100.0% 1257472/1257472 Signed-off-by: David Windsor <dave@nullcore.net> [kees: adjust commit log, split out a few extra kmalloc hunks] [kees: add field names to function declarations] [kees: convert BUGs to WARNs and fail closed] [kees: add attack surface reduction analysis to commit log] Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: linux-mm@kvack.org Cc: linux-xfs@vger.kernel.org Signed-off-by: Kees Cook <keescook@chromium.org> Acked-by: Christoph Lameter <cl@linux.com>
2017-06-11 10:50:28 +08:00
static ssize_t usersize_show(struct kmem_cache *s, char *buf)
{
return sysfs_emit(buf, "%u\n", s->usersize);
usercopy: Prepare for usercopy whitelisting This patch prepares the slab allocator to handle caches having annotations (useroffset and usersize) defining usercopy regions. This patch is modified from Brad Spengler/PaX Team's PAX_USERCOPY whitelisting code in the last public patch of grsecurity/PaX based on my understanding of the code. Changes or omissions from the original code are mine and don't reflect the original grsecurity/PaX code. Currently, hardened usercopy performs dynamic bounds checking on slab cache objects. This is good, but still leaves a lot of kernel memory available to be copied to/from userspace in the face of bugs. To further restrict what memory is available for copying, this creates a way to whitelist specific areas of a given slab cache object for copying to/from userspace, allowing much finer granularity of access control. Slab caches that are never exposed to userspace can declare no whitelist for their objects, thereby keeping them unavailable to userspace via dynamic copy operations. (Note, an implicit form of whitelisting is the use of constant sizes in usercopy operations and get_user()/put_user(); these bypass hardened usercopy checks since these sizes cannot change at runtime.) To support this whitelist annotation, usercopy region offset and size members are added to struct kmem_cache. The slab allocator receives a new function, kmem_cache_create_usercopy(), that creates a new cache with a usercopy region defined, suitable for declaring spans of fields within the objects that get copied to/from userspace. In this patch, the default kmem_cache_create() marks the entire allocation as whitelisted, leaving it semantically unchanged. Once all fine-grained whitelists have been added (in subsequent patches), this will be changed to a usersize of 0, making caches created with kmem_cache_create() not copyable to/from userspace. After the entire usercopy whitelist series is applied, less than 15% of the slab cache memory remains exposed to potential usercopy bugs after a fresh boot: Total Slab Memory: 48074720 Usercopyable Memory: 6367532 13.2% task_struct 0.2% 4480/1630720 RAW 0.3% 300/96000 RAWv6 2.1% 1408/64768 ext4_inode_cache 3.0% 269760/8740224 dentry 11.1% 585984/5273856 mm_struct 29.1% 54912/188448 kmalloc-8 100.0% 24576/24576 kmalloc-16 100.0% 28672/28672 kmalloc-32 100.0% 81920/81920 kmalloc-192 100.0% 96768/96768 kmalloc-128 100.0% 143360/143360 names_cache 100.0% 163840/163840 kmalloc-64 100.0% 167936/167936 kmalloc-256 100.0% 339968/339968 kmalloc-512 100.0% 350720/350720 kmalloc-96 100.0% 455616/455616 kmalloc-8192 100.0% 655360/655360 kmalloc-1024 100.0% 812032/812032 kmalloc-4096 100.0% 819200/819200 kmalloc-2048 100.0% 1310720/1310720 After some kernel build workloads, the percentage (mainly driven by dentry and inode caches expanding) drops under 10%: Total Slab Memory: 95516184 Usercopyable Memory: 8497452 8.8% task_struct 0.2% 4000/1456000 RAW 0.3% 300/96000 RAWv6 2.1% 1408/64768 ext4_inode_cache 3.0% 1217280/39439872 dentry 11.1% 1623200/14608800 mm_struct 29.1% 73216/251264 kmalloc-8 100.0% 24576/24576 kmalloc-16 100.0% 28672/28672 kmalloc-32 100.0% 94208/94208 kmalloc-192 100.0% 96768/96768 kmalloc-128 100.0% 143360/143360 names_cache 100.0% 163840/163840 kmalloc-64 100.0% 245760/245760 kmalloc-256 100.0% 339968/339968 kmalloc-512 100.0% 350720/350720 kmalloc-96 100.0% 563520/563520 kmalloc-8192 100.0% 655360/655360 kmalloc-1024 100.0% 794624/794624 kmalloc-4096 100.0% 819200/819200 kmalloc-2048 100.0% 1257472/1257472 Signed-off-by: David Windsor <dave@nullcore.net> [kees: adjust commit log, split out a few extra kmalloc hunks] [kees: add field names to function declarations] [kees: convert BUGs to WARNs and fail closed] [kees: add attack surface reduction analysis to commit log] Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: linux-mm@kvack.org Cc: linux-xfs@vger.kernel.org Signed-off-by: Kees Cook <keescook@chromium.org> Acked-by: Christoph Lameter <cl@linux.com>
2017-06-11 10:50:28 +08:00
}
SLAB_ATTR_RO(usersize);
static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
{
return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
}
SLAB_ATTR_RO(destroy_by_rcu);
#ifdef CONFIG_SLUB_DEBUG
static ssize_t slabs_show(struct kmem_cache *s, char *buf)
{
return show_slab_objects(s, buf, SO_ALL);
}
SLAB_ATTR_RO(slabs);
static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
{
return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
}
SLAB_ATTR_RO(total_objects);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
{
return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
SLAB_ATTR_RO(sanity_checks);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
static ssize_t trace_show(struct kmem_cache *s, char *buf)
{
return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TRACE));
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
SLAB_ATTR_RO(trace);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
{
return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
mm, slub: make some slub_debug related attributes read-only SLUB_DEBUG creates several files under /sys/kernel/slab/<cache>/ that can be read to check if the respective debugging options are enabled for given cache. The options can be also toggled at runtime by writing into the files. Some of those, namely red_zone, poison, and store_user can be toggled only when no objects yet exist in the cache. Vijayanand reports [1] that there is a problem with freelist randomization if changing the debugging option's state results in different number of objects per page, and the random sequence cache needs thus needs to be recomputed. However, another problem is that the check for "no objects yet exist in the cache" is racy, as noted by Jann [2] and fixing that would add overhead or otherwise complicate the allocation/freeing paths. Thus it would be much simpler just to remove the runtime toggling support. The documentation describes it's "In case you forgot to enable debugging on the kernel command line", but the neccessity of having no objects limits its usefulness anyway for many caches. Vijayanand describes an use case [3] where debugging is enabled for all but zram caches for memory overhead reasons, and using the runtime toggles was the only way to achieve such configuration. After the previous patch it's now possible to do that directly from the kernel boot option, so we can remove the dangerous runtime toggles by making the /sys attribute files read-only. While updating it, also improve the documentation of the debugging /sys files. [1] https://lkml.kernel.org/r/1580379523-32272-1-git-send-email-vjitta@codeaurora.org [2] https://lore.kernel.org/r/CAG48ez31PP--h6_FzVyfJ4H86QYczAFPdxtJHUEEan+7VJETAQ@mail.gmail.com [3] https://lore.kernel.org/r/1383cd32-1ddc-4dac-b5f8-9c42282fa81c@codeaurora.org Reported-by: Vijayanand Jitta <vjitta@codeaurora.org> Reported-by: Jann Horn <jannh@google.com> Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Kees Cook <keescook@chromium.org> Acked-by: Roman Gushchin <guro@fb.com> Cc: Christoph Lameter <cl@linux.com> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Pekka Enberg <penberg@kernel.org> Link: http://lkml.kernel.org/r/20200610163135.17364-3-vbabka@suse.cz Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-07 14:18:38 +08:00
SLAB_ATTR_RO(red_zone);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
static ssize_t poison_show(struct kmem_cache *s, char *buf)
{
return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_POISON));
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
mm, slub: make some slub_debug related attributes read-only SLUB_DEBUG creates several files under /sys/kernel/slab/<cache>/ that can be read to check if the respective debugging options are enabled for given cache. The options can be also toggled at runtime by writing into the files. Some of those, namely red_zone, poison, and store_user can be toggled only when no objects yet exist in the cache. Vijayanand reports [1] that there is a problem with freelist randomization if changing the debugging option's state results in different number of objects per page, and the random sequence cache needs thus needs to be recomputed. However, another problem is that the check for "no objects yet exist in the cache" is racy, as noted by Jann [2] and fixing that would add overhead or otherwise complicate the allocation/freeing paths. Thus it would be much simpler just to remove the runtime toggling support. The documentation describes it's "In case you forgot to enable debugging on the kernel command line", but the neccessity of having no objects limits its usefulness anyway for many caches. Vijayanand describes an use case [3] where debugging is enabled for all but zram caches for memory overhead reasons, and using the runtime toggles was the only way to achieve such configuration. After the previous patch it's now possible to do that directly from the kernel boot option, so we can remove the dangerous runtime toggles by making the /sys attribute files read-only. While updating it, also improve the documentation of the debugging /sys files. [1] https://lkml.kernel.org/r/1580379523-32272-1-git-send-email-vjitta@codeaurora.org [2] https://lore.kernel.org/r/CAG48ez31PP--h6_FzVyfJ4H86QYczAFPdxtJHUEEan+7VJETAQ@mail.gmail.com [3] https://lore.kernel.org/r/1383cd32-1ddc-4dac-b5f8-9c42282fa81c@codeaurora.org Reported-by: Vijayanand Jitta <vjitta@codeaurora.org> Reported-by: Jann Horn <jannh@google.com> Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Kees Cook <keescook@chromium.org> Acked-by: Roman Gushchin <guro@fb.com> Cc: Christoph Lameter <cl@linux.com> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Pekka Enberg <penberg@kernel.org> Link: http://lkml.kernel.org/r/20200610163135.17364-3-vbabka@suse.cz Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-07 14:18:38 +08:00
SLAB_ATTR_RO(poison);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
static ssize_t store_user_show(struct kmem_cache *s, char *buf)
{
return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
mm, slub: make some slub_debug related attributes read-only SLUB_DEBUG creates several files under /sys/kernel/slab/<cache>/ that can be read to check if the respective debugging options are enabled for given cache. The options can be also toggled at runtime by writing into the files. Some of those, namely red_zone, poison, and store_user can be toggled only when no objects yet exist in the cache. Vijayanand reports [1] that there is a problem with freelist randomization if changing the debugging option's state results in different number of objects per page, and the random sequence cache needs thus needs to be recomputed. However, another problem is that the check for "no objects yet exist in the cache" is racy, as noted by Jann [2] and fixing that would add overhead or otherwise complicate the allocation/freeing paths. Thus it would be much simpler just to remove the runtime toggling support. The documentation describes it's "In case you forgot to enable debugging on the kernel command line", but the neccessity of having no objects limits its usefulness anyway for many caches. Vijayanand describes an use case [3] where debugging is enabled for all but zram caches for memory overhead reasons, and using the runtime toggles was the only way to achieve such configuration. After the previous patch it's now possible to do that directly from the kernel boot option, so we can remove the dangerous runtime toggles by making the /sys attribute files read-only. While updating it, also improve the documentation of the debugging /sys files. [1] https://lkml.kernel.org/r/1580379523-32272-1-git-send-email-vjitta@codeaurora.org [2] https://lore.kernel.org/r/CAG48ez31PP--h6_FzVyfJ4H86QYczAFPdxtJHUEEan+7VJETAQ@mail.gmail.com [3] https://lore.kernel.org/r/1383cd32-1ddc-4dac-b5f8-9c42282fa81c@codeaurora.org Reported-by: Vijayanand Jitta <vjitta@codeaurora.org> Reported-by: Jann Horn <jannh@google.com> Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Kees Cook <keescook@chromium.org> Acked-by: Roman Gushchin <guro@fb.com> Cc: Christoph Lameter <cl@linux.com> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Pekka Enberg <penberg@kernel.org> Link: http://lkml.kernel.org/r/20200610163135.17364-3-vbabka@suse.cz Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-07 14:18:38 +08:00
SLAB_ATTR_RO(store_user);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
slub: validation of slabs (metadata and guard zones) This enables validation of slab. Validation means that all objects are checked to see if there are redzone violations, if padding has been overwritten or any pointers have been corrupted. Also checks the consistency of slab counters. Validation enables the detection of metadata corruption without the kernel having to execute code that actually uses (allocs/frees) and object. It allows one to make sure that the slab metainformation and the guard values around an object have not been compromised. A single slabcache can be checked by writing a 1 to the "validate" file. i.e. echo 1 >/sys/slab/kmalloc-128/validate or use the slabinfo tool to check all slabs slabinfo -v Error messages will show up in the syslog. Note that validation can only reach slabs that are on a list. This means that we are usually restricted to partial slabs and active slabs unless SLAB_STORE_USER is active which will build a full slab list and allows validation of slabs that are fully in use. Booting with "slub_debug" set will enable SLAB_STORE_USER and then full diagnostic are available. Note that we attempt to push cpu slabs back to the lists when we start the check. If the cpu slab is reactivated before we get to it (another processor grabs it before we get to it) then it cannot be checked. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:43 +08:00
static ssize_t validate_show(struct kmem_cache *s, char *buf)
{
return 0;
}
static ssize_t validate_store(struct kmem_cache *s,
const char *buf, size_t length)
{
int ret = -EINVAL;
if (buf[0] == '1') {
ret = validate_slab_cache(s);
if (ret >= 0)
ret = length;
}
return ret;
slub: validation of slabs (metadata and guard zones) This enables validation of slab. Validation means that all objects are checked to see if there are redzone violations, if padding has been overwritten or any pointers have been corrupted. Also checks the consistency of slab counters. Validation enables the detection of metadata corruption without the kernel having to execute code that actually uses (allocs/frees) and object. It allows one to make sure that the slab metainformation and the guard values around an object have not been compromised. A single slabcache can be checked by writing a 1 to the "validate" file. i.e. echo 1 >/sys/slab/kmalloc-128/validate or use the slabinfo tool to check all slabs slabinfo -v Error messages will show up in the syslog. Note that validation can only reach slabs that are on a list. This means that we are usually restricted to partial slabs and active slabs unless SLAB_STORE_USER is active which will build a full slab list and allows validation of slabs that are fully in use. Booting with "slub_debug" set will enable SLAB_STORE_USER and then full diagnostic are available. Note that we attempt to push cpu slabs back to the lists when we start the check. If the cpu slab is reactivated before we get to it (another processor grabs it before we get to it) then it cannot be checked. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:43 +08:00
}
SLAB_ATTR(validate);
static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
{
if (!(s->flags & SLAB_STORE_USER))
return -ENOSYS;
return list_locations(s, buf, TRACK_ALLOC);
}
SLAB_ATTR_RO(alloc_calls);
static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
{
if (!(s->flags & SLAB_STORE_USER))
return -ENOSYS;
return list_locations(s, buf, TRACK_FREE);
}
SLAB_ATTR_RO(free_calls);
#endif /* CONFIG_SLUB_DEBUG */
#ifdef CONFIG_FAILSLAB
static ssize_t failslab_show(struct kmem_cache *s, char *buf)
{
return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
}
SLAB_ATTR_RO(failslab);
#endif
slub: validation of slabs (metadata and guard zones) This enables validation of slab. Validation means that all objects are checked to see if there are redzone violations, if padding has been overwritten or any pointers have been corrupted. Also checks the consistency of slab counters. Validation enables the detection of metadata corruption without the kernel having to execute code that actually uses (allocs/frees) and object. It allows one to make sure that the slab metainformation and the guard values around an object have not been compromised. A single slabcache can be checked by writing a 1 to the "validate" file. i.e. echo 1 >/sys/slab/kmalloc-128/validate or use the slabinfo tool to check all slabs slabinfo -v Error messages will show up in the syslog. Note that validation can only reach slabs that are on a list. This means that we are usually restricted to partial slabs and active slabs unless SLAB_STORE_USER is active which will build a full slab list and allows validation of slabs that are fully in use. Booting with "slub_debug" set will enable SLAB_STORE_USER and then full diagnostic are available. Note that we attempt to push cpu slabs back to the lists when we start the check. If the cpu slab is reactivated before we get to it (another processor grabs it before we get to it) then it cannot be checked. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:43 +08:00
static ssize_t shrink_show(struct kmem_cache *s, char *buf)
{
return 0;
}
static ssize_t shrink_store(struct kmem_cache *s,
const char *buf, size_t length)
{
slub: never fail to shrink cache SLUB's version of __kmem_cache_shrink() not only removes empty slabs, but also tries to rearrange the partial lists to place slabs filled up most to the head to cope with fragmentation. To achieve that, it allocates a temporary array of lists used to sort slabs by the number of objects in use. If the allocation fails, the whole procedure is aborted. This is unacceptable for the kernel memory accounting extension of the memory cgroup, where we want to make sure that kmem_cache_shrink() successfully discarded empty slabs. Although the allocation failure is utterly unlikely with the current page allocator implementation, which retries GFP_KERNEL allocations of order <= 2 infinitely, it is better not to rely on that. This patch therefore makes __kmem_cache_shrink() allocate the array on stack instead of calling kmalloc, which may fail. The array size is chosen to be equal to 32, because most SLUB caches store not more than 32 objects per slab page. Slab pages with <= 32 free objects are sorted using the array by the number of objects in use and promoted to the head of the partial list, while slab pages with > 32 free objects are left in the end of the list without any ordering imposed on them. Signed-off-by: Vladimir Davydov <vdavydov@parallels.com> Acked-by: Christoph Lameter <cl@linux.com> Acked-by: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Huang Ying <ying.huang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-13 06:59:41 +08:00
if (buf[0] == '1')
kmem_cache_shrink(s);
slub: never fail to shrink cache SLUB's version of __kmem_cache_shrink() not only removes empty slabs, but also tries to rearrange the partial lists to place slabs filled up most to the head to cope with fragmentation. To achieve that, it allocates a temporary array of lists used to sort slabs by the number of objects in use. If the allocation fails, the whole procedure is aborted. This is unacceptable for the kernel memory accounting extension of the memory cgroup, where we want to make sure that kmem_cache_shrink() successfully discarded empty slabs. Although the allocation failure is utterly unlikely with the current page allocator implementation, which retries GFP_KERNEL allocations of order <= 2 infinitely, it is better not to rely on that. This patch therefore makes __kmem_cache_shrink() allocate the array on stack instead of calling kmalloc, which may fail. The array size is chosen to be equal to 32, because most SLUB caches store not more than 32 objects per slab page. Slab pages with <= 32 free objects are sorted using the array by the number of objects in use and promoted to the head of the partial list, while slab pages with > 32 free objects are left in the end of the list without any ordering imposed on them. Signed-off-by: Vladimir Davydov <vdavydov@parallels.com> Acked-by: Christoph Lameter <cl@linux.com> Acked-by: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Huang Ying <ying.huang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-13 06:59:41 +08:00
else
return -EINVAL;
return length;
}
SLAB_ATTR(shrink);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
#ifdef CONFIG_NUMA
static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
{
return sysfs_emit(buf, "%u\n", s->remote_node_defrag_ratio / 10);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
const char *buf, size_t length)
{
unsigned int ratio;
int err;
err = kstrtouint(buf, 10, &ratio);
if (err)
return err;
if (ratio > 100)
return -ERANGE;
s->remote_node_defrag_ratio = ratio * 10;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
return length;
}
SLAB_ATTR(remote_node_defrag_ratio);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
#endif
SLUB: Support for performance statistics The statistics provided here allow the monitoring of allocator behavior but at the cost of some (minimal) loss of performance. Counters are placed in SLUB's per cpu data structure. The per cpu structure may be extended by the statistics to grow larger than one cacheline which will increase the cache footprint of SLUB. There is a compile option to enable/disable the inclusion of the runtime statistics and its off by default. The slabinfo tool is enhanced to support these statistics via two options: -D Switches the line of information displayed for a slab from size mode to activity mode. -A Sorts the slabs displayed by activity. This allows the display of the slabs most important to the performance of a certain load. -r Report option will report detailed statistics on Example (tbench load): slabinfo -AD ->Shows the most active slabs Name Objects Alloc Free %Fast skbuff_fclone_cache 33 111953835 111953835 99 99 :0000192 2666 5283688 5281047 99 99 :0001024 849 5247230 5246389 83 83 vm_area_struct 1349 119642 118355 91 22 :0004096 15 66753 66751 98 98 :0000064 2067 25297 23383 98 78 dentry 10259 28635 18464 91 45 :0000080 11004 18950 8089 98 98 :0000096 1703 12358 10784 99 98 :0000128 762 10582 9875 94 18 :0000512 184 9807 9647 95 81 :0002048 479 9669 9195 83 65 anon_vma 777 9461 9002 99 71 kmalloc-8 6492 9981 5624 99 97 :0000768 258 7174 6931 58 15 So the skbuff_fclone_cache is of highest importance for the tbench load. Pretty high load on the 192 sized slab. Look for the aliases slabinfo -a | grep 000192 :0000192 <- xfs_btree_cur filp kmalloc-192 uid_cache tw_sock_TCP request_sock_TCPv6 tw_sock_TCPv6 skbuff_head_cache xfs_ili Likely skbuff_head_cache. Looking into the statistics of the skbuff_fclone_cache is possible through slabinfo skbuff_fclone_cache ->-r option implied if cache name is mentioned .... Usual output ... Slab Perf Counter Alloc Free %Al %Fr -------------------------------------------------- Fastpath 111953360 111946981 99 99 Slowpath 1044 7423 0 0 Page Alloc 272 264 0 0 Add partial 25 325 0 0 Remove partial 86 264 0 0 RemoteObj/SlabFrozen 350 4832 0 0 Total 111954404 111954404 Flushes 49 Refill 0 Deactivate Full=325(92%) Empty=0(0%) ToHead=24(6%) ToTail=1(0%) Looks good because the fastpath is overwhelmingly taken. skbuff_head_cache: Slab Perf Counter Alloc Free %Al %Fr -------------------------------------------------- Fastpath 5297262 5259882 99 99 Slowpath 4477 39586 0 0 Page Alloc 937 824 0 0 Add partial 0 2515 0 0 Remove partial 1691 824 0 0 RemoteObj/SlabFrozen 2621 9684 0 0 Total 5301739 5299468 Deactivate Full=2620(100%) Empty=0(0%) ToHead=0(0%) ToTail=0(0%) Descriptions of the output: Total: The total number of allocation and frees that occurred for a slab Fastpath: The number of allocations/frees that used the fastpath. Slowpath: Other allocations Page Alloc: Number of calls to the page allocator as a result of slowpath processing Add Partial: Number of slabs added to the partial list through free or alloc (occurs during cpuslab flushes) Remove Partial: Number of slabs removed from the partial list as a result of allocations retrieving a partial slab or by a free freeing the last object of a slab. RemoteObj/Froz: How many times were remotely freed object encountered when a slab was about to be deactivated. Frozen: How many times was free able to skip list processing because the slab was in use as the cpuslab of another processor. Flushes: Number of times the cpuslab was flushed on request (kmem_cache_shrink, may result from races in __slab_alloc) Refill: Number of times we were able to refill the cpuslab from remotely freed objects for the same slab. Deactivate: Statistics how slabs were deactivated. Shows how they were put onto the partial list. In general fastpath is very good. Slowpath without partial list processing is also desirable. Any touching of partial list uses node specific locks which may potentially cause list lock contention. Signed-off-by: Christoph Lameter <clameter@sgi.com>
2008-02-08 09:47:41 +08:00
#ifdef CONFIG_SLUB_STATS
static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
{
unsigned long sum = 0;
int cpu;
int len = 0;
treewide: kmalloc() -> kmalloc_array() The kmalloc() function has a 2-factor argument form, kmalloc_array(). This patch replaces cases of: kmalloc(a * b, gfp) with: kmalloc_array(a * b, gfp) as well as handling cases of: kmalloc(a * b * c, gfp) with: kmalloc(array3_size(a, b, c), gfp) as it's slightly less ugly than: kmalloc_array(array_size(a, b), c, gfp) This does, however, attempt to ignore constant size factors like: kmalloc(4 * 1024, gfp) though any constants defined via macros get caught up in the conversion. Any factors with a sizeof() of "unsigned char", "char", and "u8" were dropped, since they're redundant. The tools/ directory was manually excluded, since it has its own implementation of kmalloc(). The Coccinelle script used for this was: // Fix redundant parens around sizeof(). @@ type TYPE; expression THING, E; @@ ( kmalloc( - (sizeof(TYPE)) * E + sizeof(TYPE) * E , ...) | kmalloc( - (sizeof(THING)) * E + sizeof(THING) * E , ...) ) // Drop single-byte sizes and redundant parens. @@ expression COUNT; typedef u8; typedef __u8; @@ ( kmalloc( - sizeof(u8) * (COUNT) + COUNT , ...) | kmalloc( - sizeof(__u8) * (COUNT) + COUNT , ...) | kmalloc( - sizeof(char) * (COUNT) + COUNT , ...) | kmalloc( - sizeof(unsigned char) * (COUNT) + COUNT , ...) | kmalloc( - sizeof(u8) * COUNT + COUNT , ...) | kmalloc( - sizeof(__u8) * COUNT + COUNT , ...) | kmalloc( - sizeof(char) * COUNT + COUNT , ...) | kmalloc( - sizeof(unsigned char) * COUNT + COUNT , ...) ) // 2-factor product with sizeof(type/expression) and identifier or constant. @@ type TYPE; expression THING; identifier COUNT_ID; constant COUNT_CONST; @@ ( - kmalloc + kmalloc_array ( - sizeof(TYPE) * (COUNT_ID) + COUNT_ID, sizeof(TYPE) , ...) | - kmalloc + kmalloc_array ( - sizeof(TYPE) * COUNT_ID + COUNT_ID, sizeof(TYPE) , ...) | - kmalloc + kmalloc_array ( - sizeof(TYPE) * (COUNT_CONST) + COUNT_CONST, sizeof(TYPE) , ...) | - kmalloc + kmalloc_array ( - sizeof(TYPE) * COUNT_CONST + COUNT_CONST, sizeof(TYPE) , ...) | - kmalloc + kmalloc_array ( - sizeof(THING) * (COUNT_ID) + COUNT_ID, sizeof(THING) , ...) | - kmalloc + kmalloc_array ( - sizeof(THING) * COUNT_ID + COUNT_ID, sizeof(THING) , ...) | - kmalloc + kmalloc_array ( - sizeof(THING) * (COUNT_CONST) + COUNT_CONST, sizeof(THING) , ...) | - kmalloc + kmalloc_array ( - sizeof(THING) * COUNT_CONST + COUNT_CONST, sizeof(THING) , ...) ) // 2-factor product, only identifiers. @@ identifier SIZE, COUNT; @@ - kmalloc + kmalloc_array ( - SIZE * COUNT + COUNT, SIZE , ...) // 3-factor product with 1 sizeof(type) or sizeof(expression), with // redundant parens removed. @@ expression THING; identifier STRIDE, COUNT; type TYPE; @@ ( kmalloc( - sizeof(TYPE) * (COUNT) * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kmalloc( - sizeof(TYPE) * (COUNT) * STRIDE + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kmalloc( - sizeof(TYPE) * COUNT * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kmalloc( - sizeof(TYPE) * COUNT * STRIDE + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kmalloc( - sizeof(THING) * (COUNT) * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kmalloc( - sizeof(THING) * (COUNT) * STRIDE + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kmalloc( - sizeof(THING) * COUNT * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kmalloc( - sizeof(THING) * COUNT * STRIDE + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) ) // 3-factor product with 2 sizeof(variable), with redundant parens removed. @@ expression THING1, THING2; identifier COUNT; type TYPE1, TYPE2; @@ ( kmalloc( - sizeof(TYPE1) * sizeof(TYPE2) * COUNT + array3_size(COUNT, sizeof(TYPE1), sizeof(TYPE2)) , ...) | kmalloc( - sizeof(TYPE1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(TYPE1), sizeof(TYPE2)) , ...) | kmalloc( - sizeof(THING1) * sizeof(THING2) * COUNT + array3_size(COUNT, sizeof(THING1), sizeof(THING2)) , ...) | kmalloc( - sizeof(THING1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(THING1), sizeof(THING2)) , ...) | kmalloc( - sizeof(TYPE1) * sizeof(THING2) * COUNT + array3_size(COUNT, sizeof(TYPE1), sizeof(THING2)) , ...) | kmalloc( - sizeof(TYPE1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(TYPE1), sizeof(THING2)) , ...) ) // 3-factor product, only identifiers, with redundant parens removed. @@ identifier STRIDE, SIZE, COUNT; @@ ( kmalloc( - (COUNT) * STRIDE * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kmalloc( - COUNT * (STRIDE) * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kmalloc( - COUNT * STRIDE * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kmalloc( - (COUNT) * (STRIDE) * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kmalloc( - COUNT * (STRIDE) * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kmalloc( - (COUNT) * STRIDE * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kmalloc( - (COUNT) * (STRIDE) * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kmalloc( - COUNT * STRIDE * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) ) // Any remaining multi-factor products, first at least 3-factor products, // when they're not all constants... @@ expression E1, E2, E3; constant C1, C2, C3; @@ ( kmalloc(C1 * C2 * C3, ...) | kmalloc( - (E1) * E2 * E3 + array3_size(E1, E2, E3) , ...) | kmalloc( - (E1) * (E2) * E3 + array3_size(E1, E2, E3) , ...) | kmalloc( - (E1) * (E2) * (E3) + array3_size(E1, E2, E3) , ...) | kmalloc( - E1 * E2 * E3 + array3_size(E1, E2, E3) , ...) ) // And then all remaining 2 factors products when they're not all constants, // keeping sizeof() as the second factor argument. @@ expression THING, E1, E2; type TYPE; constant C1, C2, C3; @@ ( kmalloc(sizeof(THING) * C2, ...) | kmalloc(sizeof(TYPE) * C2, ...) | kmalloc(C1 * C2 * C3, ...) | kmalloc(C1 * C2, ...) | - kmalloc + kmalloc_array ( - sizeof(TYPE) * (E2) + E2, sizeof(TYPE) , ...) | - kmalloc + kmalloc_array ( - sizeof(TYPE) * E2 + E2, sizeof(TYPE) , ...) | - kmalloc + kmalloc_array ( - sizeof(THING) * (E2) + E2, sizeof(THING) , ...) | - kmalloc + kmalloc_array ( - sizeof(THING) * E2 + E2, sizeof(THING) , ...) | - kmalloc + kmalloc_array ( - (E1) * E2 + E1, E2 , ...) | - kmalloc + kmalloc_array ( - (E1) * (E2) + E1, E2 , ...) | - kmalloc + kmalloc_array ( - E1 * E2 + E1, E2 , ...) ) Signed-off-by: Kees Cook <keescook@chromium.org>
2018-06-13 04:55:00 +08:00
int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
SLUB: Support for performance statistics The statistics provided here allow the monitoring of allocator behavior but at the cost of some (minimal) loss of performance. Counters are placed in SLUB's per cpu data structure. The per cpu structure may be extended by the statistics to grow larger than one cacheline which will increase the cache footprint of SLUB. There is a compile option to enable/disable the inclusion of the runtime statistics and its off by default. The slabinfo tool is enhanced to support these statistics via two options: -D Switches the line of information displayed for a slab from size mode to activity mode. -A Sorts the slabs displayed by activity. This allows the display of the slabs most important to the performance of a certain load. -r Report option will report detailed statistics on Example (tbench load): slabinfo -AD ->Shows the most active slabs Name Objects Alloc Free %Fast skbuff_fclone_cache 33 111953835 111953835 99 99 :0000192 2666 5283688 5281047 99 99 :0001024 849 5247230 5246389 83 83 vm_area_struct 1349 119642 118355 91 22 :0004096 15 66753 66751 98 98 :0000064 2067 25297 23383 98 78 dentry 10259 28635 18464 91 45 :0000080 11004 18950 8089 98 98 :0000096 1703 12358 10784 99 98 :0000128 762 10582 9875 94 18 :0000512 184 9807 9647 95 81 :0002048 479 9669 9195 83 65 anon_vma 777 9461 9002 99 71 kmalloc-8 6492 9981 5624 99 97 :0000768 258 7174 6931 58 15 So the skbuff_fclone_cache is of highest importance for the tbench load. Pretty high load on the 192 sized slab. Look for the aliases slabinfo -a | grep 000192 :0000192 <- xfs_btree_cur filp kmalloc-192 uid_cache tw_sock_TCP request_sock_TCPv6 tw_sock_TCPv6 skbuff_head_cache xfs_ili Likely skbuff_head_cache. Looking into the statistics of the skbuff_fclone_cache is possible through slabinfo skbuff_fclone_cache ->-r option implied if cache name is mentioned .... Usual output ... Slab Perf Counter Alloc Free %Al %Fr -------------------------------------------------- Fastpath 111953360 111946981 99 99 Slowpath 1044 7423 0 0 Page Alloc 272 264 0 0 Add partial 25 325 0 0 Remove partial 86 264 0 0 RemoteObj/SlabFrozen 350 4832 0 0 Total 111954404 111954404 Flushes 49 Refill 0 Deactivate Full=325(92%) Empty=0(0%) ToHead=24(6%) ToTail=1(0%) Looks good because the fastpath is overwhelmingly taken. skbuff_head_cache: Slab Perf Counter Alloc Free %Al %Fr -------------------------------------------------- Fastpath 5297262 5259882 99 99 Slowpath 4477 39586 0 0 Page Alloc 937 824 0 0 Add partial 0 2515 0 0 Remove partial 1691 824 0 0 RemoteObj/SlabFrozen 2621 9684 0 0 Total 5301739 5299468 Deactivate Full=2620(100%) Empty=0(0%) ToHead=0(0%) ToTail=0(0%) Descriptions of the output: Total: The total number of allocation and frees that occurred for a slab Fastpath: The number of allocations/frees that used the fastpath. Slowpath: Other allocations Page Alloc: Number of calls to the page allocator as a result of slowpath processing Add Partial: Number of slabs added to the partial list through free or alloc (occurs during cpuslab flushes) Remove Partial: Number of slabs removed from the partial list as a result of allocations retrieving a partial slab or by a free freeing the last object of a slab. RemoteObj/Froz: How many times were remotely freed object encountered when a slab was about to be deactivated. Frozen: How many times was free able to skip list processing because the slab was in use as the cpuslab of another processor. Flushes: Number of times the cpuslab was flushed on request (kmem_cache_shrink, may result from races in __slab_alloc) Refill: Number of times we were able to refill the cpuslab from remotely freed objects for the same slab. Deactivate: Statistics how slabs were deactivated. Shows how they were put onto the partial list. In general fastpath is very good. Slowpath without partial list processing is also desirable. Any touching of partial list uses node specific locks which may potentially cause list lock contention. Signed-off-by: Christoph Lameter <clameter@sgi.com>
2008-02-08 09:47:41 +08:00
if (!data)
return -ENOMEM;
for_each_online_cpu(cpu) {
unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
SLUB: Support for performance statistics The statistics provided here allow the monitoring of allocator behavior but at the cost of some (minimal) loss of performance. Counters are placed in SLUB's per cpu data structure. The per cpu structure may be extended by the statistics to grow larger than one cacheline which will increase the cache footprint of SLUB. There is a compile option to enable/disable the inclusion of the runtime statistics and its off by default. The slabinfo tool is enhanced to support these statistics via two options: -D Switches the line of information displayed for a slab from size mode to activity mode. -A Sorts the slabs displayed by activity. This allows the display of the slabs most important to the performance of a certain load. -r Report option will report detailed statistics on Example (tbench load): slabinfo -AD ->Shows the most active slabs Name Objects Alloc Free %Fast skbuff_fclone_cache 33 111953835 111953835 99 99 :0000192 2666 5283688 5281047 99 99 :0001024 849 5247230 5246389 83 83 vm_area_struct 1349 119642 118355 91 22 :0004096 15 66753 66751 98 98 :0000064 2067 25297 23383 98 78 dentry 10259 28635 18464 91 45 :0000080 11004 18950 8089 98 98 :0000096 1703 12358 10784 99 98 :0000128 762 10582 9875 94 18 :0000512 184 9807 9647 95 81 :0002048 479 9669 9195 83 65 anon_vma 777 9461 9002 99 71 kmalloc-8 6492 9981 5624 99 97 :0000768 258 7174 6931 58 15 So the skbuff_fclone_cache is of highest importance for the tbench load. Pretty high load on the 192 sized slab. Look for the aliases slabinfo -a | grep 000192 :0000192 <- xfs_btree_cur filp kmalloc-192 uid_cache tw_sock_TCP request_sock_TCPv6 tw_sock_TCPv6 skbuff_head_cache xfs_ili Likely skbuff_head_cache. Looking into the statistics of the skbuff_fclone_cache is possible through slabinfo skbuff_fclone_cache ->-r option implied if cache name is mentioned .... Usual output ... Slab Perf Counter Alloc Free %Al %Fr -------------------------------------------------- Fastpath 111953360 111946981 99 99 Slowpath 1044 7423 0 0 Page Alloc 272 264 0 0 Add partial 25 325 0 0 Remove partial 86 264 0 0 RemoteObj/SlabFrozen 350 4832 0 0 Total 111954404 111954404 Flushes 49 Refill 0 Deactivate Full=325(92%) Empty=0(0%) ToHead=24(6%) ToTail=1(0%) Looks good because the fastpath is overwhelmingly taken. skbuff_head_cache: Slab Perf Counter Alloc Free %Al %Fr -------------------------------------------------- Fastpath 5297262 5259882 99 99 Slowpath 4477 39586 0 0 Page Alloc 937 824 0 0 Add partial 0 2515 0 0 Remove partial 1691 824 0 0 RemoteObj/SlabFrozen 2621 9684 0 0 Total 5301739 5299468 Deactivate Full=2620(100%) Empty=0(0%) ToHead=0(0%) ToTail=0(0%) Descriptions of the output: Total: The total number of allocation and frees that occurred for a slab Fastpath: The number of allocations/frees that used the fastpath. Slowpath: Other allocations Page Alloc: Number of calls to the page allocator as a result of slowpath processing Add Partial: Number of slabs added to the partial list through free or alloc (occurs during cpuslab flushes) Remove Partial: Number of slabs removed from the partial list as a result of allocations retrieving a partial slab or by a free freeing the last object of a slab. RemoteObj/Froz: How many times were remotely freed object encountered when a slab was about to be deactivated. Frozen: How many times was free able to skip list processing because the slab was in use as the cpuslab of another processor. Flushes: Number of times the cpuslab was flushed on request (kmem_cache_shrink, may result from races in __slab_alloc) Refill: Number of times we were able to refill the cpuslab from remotely freed objects for the same slab. Deactivate: Statistics how slabs were deactivated. Shows how they were put onto the partial list. In general fastpath is very good. Slowpath without partial list processing is also desirable. Any touching of partial list uses node specific locks which may potentially cause list lock contention. Signed-off-by: Christoph Lameter <clameter@sgi.com>
2008-02-08 09:47:41 +08:00
data[cpu] = x;
sum += x;
}
len += sysfs_emit_at(buf, len, "%lu", sum);
SLUB: Support for performance statistics The statistics provided here allow the monitoring of allocator behavior but at the cost of some (minimal) loss of performance. Counters are placed in SLUB's per cpu data structure. The per cpu structure may be extended by the statistics to grow larger than one cacheline which will increase the cache footprint of SLUB. There is a compile option to enable/disable the inclusion of the runtime statistics and its off by default. The slabinfo tool is enhanced to support these statistics via two options: -D Switches the line of information displayed for a slab from size mode to activity mode. -A Sorts the slabs displayed by activity. This allows the display of the slabs most important to the performance of a certain load. -r Report option will report detailed statistics on Example (tbench load): slabinfo -AD ->Shows the most active slabs Name Objects Alloc Free %Fast skbuff_fclone_cache 33 111953835 111953835 99 99 :0000192 2666 5283688 5281047 99 99 :0001024 849 5247230 5246389 83 83 vm_area_struct 1349 119642 118355 91 22 :0004096 15 66753 66751 98 98 :0000064 2067 25297 23383 98 78 dentry 10259 28635 18464 91 45 :0000080 11004 18950 8089 98 98 :0000096 1703 12358 10784 99 98 :0000128 762 10582 9875 94 18 :0000512 184 9807 9647 95 81 :0002048 479 9669 9195 83 65 anon_vma 777 9461 9002 99 71 kmalloc-8 6492 9981 5624 99 97 :0000768 258 7174 6931 58 15 So the skbuff_fclone_cache is of highest importance for the tbench load. Pretty high load on the 192 sized slab. Look for the aliases slabinfo -a | grep 000192 :0000192 <- xfs_btree_cur filp kmalloc-192 uid_cache tw_sock_TCP request_sock_TCPv6 tw_sock_TCPv6 skbuff_head_cache xfs_ili Likely skbuff_head_cache. Looking into the statistics of the skbuff_fclone_cache is possible through slabinfo skbuff_fclone_cache ->-r option implied if cache name is mentioned .... Usual output ... Slab Perf Counter Alloc Free %Al %Fr -------------------------------------------------- Fastpath 111953360 111946981 99 99 Slowpath 1044 7423 0 0 Page Alloc 272 264 0 0 Add partial 25 325 0 0 Remove partial 86 264 0 0 RemoteObj/SlabFrozen 350 4832 0 0 Total 111954404 111954404 Flushes 49 Refill 0 Deactivate Full=325(92%) Empty=0(0%) ToHead=24(6%) ToTail=1(0%) Looks good because the fastpath is overwhelmingly taken. skbuff_head_cache: Slab Perf Counter Alloc Free %Al %Fr -------------------------------------------------- Fastpath 5297262 5259882 99 99 Slowpath 4477 39586 0 0 Page Alloc 937 824 0 0 Add partial 0 2515 0 0 Remove partial 1691 824 0 0 RemoteObj/SlabFrozen 2621 9684 0 0 Total 5301739 5299468 Deactivate Full=2620(100%) Empty=0(0%) ToHead=0(0%) ToTail=0(0%) Descriptions of the output: Total: The total number of allocation and frees that occurred for a slab Fastpath: The number of allocations/frees that used the fastpath. Slowpath: Other allocations Page Alloc: Number of calls to the page allocator as a result of slowpath processing Add Partial: Number of slabs added to the partial list through free or alloc (occurs during cpuslab flushes) Remove Partial: Number of slabs removed from the partial list as a result of allocations retrieving a partial slab or by a free freeing the last object of a slab. RemoteObj/Froz: How many times were remotely freed object encountered when a slab was about to be deactivated. Frozen: How many times was free able to skip list processing because the slab was in use as the cpuslab of another processor. Flushes: Number of times the cpuslab was flushed on request (kmem_cache_shrink, may result from races in __slab_alloc) Refill: Number of times we were able to refill the cpuslab from remotely freed objects for the same slab. Deactivate: Statistics how slabs were deactivated. Shows how they were put onto the partial list. In general fastpath is very good. Slowpath without partial list processing is also desirable. Any touching of partial list uses node specific locks which may potentially cause list lock contention. Signed-off-by: Christoph Lameter <clameter@sgi.com>
2008-02-08 09:47:41 +08:00
#ifdef CONFIG_SMP
SLUB: Support for performance statistics The statistics provided here allow the monitoring of allocator behavior but at the cost of some (minimal) loss of performance. Counters are placed in SLUB's per cpu data structure. The per cpu structure may be extended by the statistics to grow larger than one cacheline which will increase the cache footprint of SLUB. There is a compile option to enable/disable the inclusion of the runtime statistics and its off by default. The slabinfo tool is enhanced to support these statistics via two options: -D Switches the line of information displayed for a slab from size mode to activity mode. -A Sorts the slabs displayed by activity. This allows the display of the slabs most important to the performance of a certain load. -r Report option will report detailed statistics on Example (tbench load): slabinfo -AD ->Shows the most active slabs Name Objects Alloc Free %Fast skbuff_fclone_cache 33 111953835 111953835 99 99 :0000192 2666 5283688 5281047 99 99 :0001024 849 5247230 5246389 83 83 vm_area_struct 1349 119642 118355 91 22 :0004096 15 66753 66751 98 98 :0000064 2067 25297 23383 98 78 dentry 10259 28635 18464 91 45 :0000080 11004 18950 8089 98 98 :0000096 1703 12358 10784 99 98 :0000128 762 10582 9875 94 18 :0000512 184 9807 9647 95 81 :0002048 479 9669 9195 83 65 anon_vma 777 9461 9002 99 71 kmalloc-8 6492 9981 5624 99 97 :0000768 258 7174 6931 58 15 So the skbuff_fclone_cache is of highest importance for the tbench load. Pretty high load on the 192 sized slab. Look for the aliases slabinfo -a | grep 000192 :0000192 <- xfs_btree_cur filp kmalloc-192 uid_cache tw_sock_TCP request_sock_TCPv6 tw_sock_TCPv6 skbuff_head_cache xfs_ili Likely skbuff_head_cache. Looking into the statistics of the skbuff_fclone_cache is possible through slabinfo skbuff_fclone_cache ->-r option implied if cache name is mentioned .... Usual output ... Slab Perf Counter Alloc Free %Al %Fr -------------------------------------------------- Fastpath 111953360 111946981 99 99 Slowpath 1044 7423 0 0 Page Alloc 272 264 0 0 Add partial 25 325 0 0 Remove partial 86 264 0 0 RemoteObj/SlabFrozen 350 4832 0 0 Total 111954404 111954404 Flushes 49 Refill 0 Deactivate Full=325(92%) Empty=0(0%) ToHead=24(6%) ToTail=1(0%) Looks good because the fastpath is overwhelmingly taken. skbuff_head_cache: Slab Perf Counter Alloc Free %Al %Fr -------------------------------------------------- Fastpath 5297262 5259882 99 99 Slowpath 4477 39586 0 0 Page Alloc 937 824 0 0 Add partial 0 2515 0 0 Remove partial 1691 824 0 0 RemoteObj/SlabFrozen 2621 9684 0 0 Total 5301739 5299468 Deactivate Full=2620(100%) Empty=0(0%) ToHead=0(0%) ToTail=0(0%) Descriptions of the output: Total: The total number of allocation and frees that occurred for a slab Fastpath: The number of allocations/frees that used the fastpath. Slowpath: Other allocations Page Alloc: Number of calls to the page allocator as a result of slowpath processing Add Partial: Number of slabs added to the partial list through free or alloc (occurs during cpuslab flushes) Remove Partial: Number of slabs removed from the partial list as a result of allocations retrieving a partial slab or by a free freeing the last object of a slab. RemoteObj/Froz: How many times were remotely freed object encountered when a slab was about to be deactivated. Frozen: How many times was free able to skip list processing because the slab was in use as the cpuslab of another processor. Flushes: Number of times the cpuslab was flushed on request (kmem_cache_shrink, may result from races in __slab_alloc) Refill: Number of times we were able to refill the cpuslab from remotely freed objects for the same slab. Deactivate: Statistics how slabs were deactivated. Shows how they were put onto the partial list. In general fastpath is very good. Slowpath without partial list processing is also desirable. Any touching of partial list uses node specific locks which may potentially cause list lock contention. Signed-off-by: Christoph Lameter <clameter@sgi.com>
2008-02-08 09:47:41 +08:00
for_each_online_cpu(cpu) {
if (data[cpu])
len += sysfs_emit_at(buf, len, " C%d=%u",
cpu, data[cpu]);
SLUB: Support for performance statistics The statistics provided here allow the monitoring of allocator behavior but at the cost of some (minimal) loss of performance. Counters are placed in SLUB's per cpu data structure. The per cpu structure may be extended by the statistics to grow larger than one cacheline which will increase the cache footprint of SLUB. There is a compile option to enable/disable the inclusion of the runtime statistics and its off by default. The slabinfo tool is enhanced to support these statistics via two options: -D Switches the line of information displayed for a slab from size mode to activity mode. -A Sorts the slabs displayed by activity. This allows the display of the slabs most important to the performance of a certain load. -r Report option will report detailed statistics on Example (tbench load): slabinfo -AD ->Shows the most active slabs Name Objects Alloc Free %Fast skbuff_fclone_cache 33 111953835 111953835 99 99 :0000192 2666 5283688 5281047 99 99 :0001024 849 5247230 5246389 83 83 vm_area_struct 1349 119642 118355 91 22 :0004096 15 66753 66751 98 98 :0000064 2067 25297 23383 98 78 dentry 10259 28635 18464 91 45 :0000080 11004 18950 8089 98 98 :0000096 1703 12358 10784 99 98 :0000128 762 10582 9875 94 18 :0000512 184 9807 9647 95 81 :0002048 479 9669 9195 83 65 anon_vma 777 9461 9002 99 71 kmalloc-8 6492 9981 5624 99 97 :0000768 258 7174 6931 58 15 So the skbuff_fclone_cache is of highest importance for the tbench load. Pretty high load on the 192 sized slab. Look for the aliases slabinfo -a | grep 000192 :0000192 <- xfs_btree_cur filp kmalloc-192 uid_cache tw_sock_TCP request_sock_TCPv6 tw_sock_TCPv6 skbuff_head_cache xfs_ili Likely skbuff_head_cache. Looking into the statistics of the skbuff_fclone_cache is possible through slabinfo skbuff_fclone_cache ->-r option implied if cache name is mentioned .... Usual output ... Slab Perf Counter Alloc Free %Al %Fr -------------------------------------------------- Fastpath 111953360 111946981 99 99 Slowpath 1044 7423 0 0 Page Alloc 272 264 0 0 Add partial 25 325 0 0 Remove partial 86 264 0 0 RemoteObj/SlabFrozen 350 4832 0 0 Total 111954404 111954404 Flushes 49 Refill 0 Deactivate Full=325(92%) Empty=0(0%) ToHead=24(6%) ToTail=1(0%) Looks good because the fastpath is overwhelmingly taken. skbuff_head_cache: Slab Perf Counter Alloc Free %Al %Fr -------------------------------------------------- Fastpath 5297262 5259882 99 99 Slowpath 4477 39586 0 0 Page Alloc 937 824 0 0 Add partial 0 2515 0 0 Remove partial 1691 824 0 0 RemoteObj/SlabFrozen 2621 9684 0 0 Total 5301739 5299468 Deactivate Full=2620(100%) Empty=0(0%) ToHead=0(0%) ToTail=0(0%) Descriptions of the output: Total: The total number of allocation and frees that occurred for a slab Fastpath: The number of allocations/frees that used the fastpath. Slowpath: Other allocations Page Alloc: Number of calls to the page allocator as a result of slowpath processing Add Partial: Number of slabs added to the partial list through free or alloc (occurs during cpuslab flushes) Remove Partial: Number of slabs removed from the partial list as a result of allocations retrieving a partial slab or by a free freeing the last object of a slab. RemoteObj/Froz: How many times were remotely freed object encountered when a slab was about to be deactivated. Frozen: How many times was free able to skip list processing because the slab was in use as the cpuslab of another processor. Flushes: Number of times the cpuslab was flushed on request (kmem_cache_shrink, may result from races in __slab_alloc) Refill: Number of times we were able to refill the cpuslab from remotely freed objects for the same slab. Deactivate: Statistics how slabs were deactivated. Shows how they were put onto the partial list. In general fastpath is very good. Slowpath without partial list processing is also desirable. Any touching of partial list uses node specific locks which may potentially cause list lock contention. Signed-off-by: Christoph Lameter <clameter@sgi.com>
2008-02-08 09:47:41 +08:00
}
#endif
SLUB: Support for performance statistics The statistics provided here allow the monitoring of allocator behavior but at the cost of some (minimal) loss of performance. Counters are placed in SLUB's per cpu data structure. The per cpu structure may be extended by the statistics to grow larger than one cacheline which will increase the cache footprint of SLUB. There is a compile option to enable/disable the inclusion of the runtime statistics and its off by default. The slabinfo tool is enhanced to support these statistics via two options: -D Switches the line of information displayed for a slab from size mode to activity mode. -A Sorts the slabs displayed by activity. This allows the display of the slabs most important to the performance of a certain load. -r Report option will report detailed statistics on Example (tbench load): slabinfo -AD ->Shows the most active slabs Name Objects Alloc Free %Fast skbuff_fclone_cache 33 111953835 111953835 99 99 :0000192 2666 5283688 5281047 99 99 :0001024 849 5247230 5246389 83 83 vm_area_struct 1349 119642 118355 91 22 :0004096 15 66753 66751 98 98 :0000064 2067 25297 23383 98 78 dentry 10259 28635 18464 91 45 :0000080 11004 18950 8089 98 98 :0000096 1703 12358 10784 99 98 :0000128 762 10582 9875 94 18 :0000512 184 9807 9647 95 81 :0002048 479 9669 9195 83 65 anon_vma 777 9461 9002 99 71 kmalloc-8 6492 9981 5624 99 97 :0000768 258 7174 6931 58 15 So the skbuff_fclone_cache is of highest importance for the tbench load. Pretty high load on the 192 sized slab. Look for the aliases slabinfo -a | grep 000192 :0000192 <- xfs_btree_cur filp kmalloc-192 uid_cache tw_sock_TCP request_sock_TCPv6 tw_sock_TCPv6 skbuff_head_cache xfs_ili Likely skbuff_head_cache. Looking into the statistics of the skbuff_fclone_cache is possible through slabinfo skbuff_fclone_cache ->-r option implied if cache name is mentioned .... Usual output ... Slab Perf Counter Alloc Free %Al %Fr -------------------------------------------------- Fastpath 111953360 111946981 99 99 Slowpath 1044 7423 0 0 Page Alloc 272 264 0 0 Add partial 25 325 0 0 Remove partial 86 264 0 0 RemoteObj/SlabFrozen 350 4832 0 0 Total 111954404 111954404 Flushes 49 Refill 0 Deactivate Full=325(92%) Empty=0(0%) ToHead=24(6%) ToTail=1(0%) Looks good because the fastpath is overwhelmingly taken. skbuff_head_cache: Slab Perf Counter Alloc Free %Al %Fr -------------------------------------------------- Fastpath 5297262 5259882 99 99 Slowpath 4477 39586 0 0 Page Alloc 937 824 0 0 Add partial 0 2515 0 0 Remove partial 1691 824 0 0 RemoteObj/SlabFrozen 2621 9684 0 0 Total 5301739 5299468 Deactivate Full=2620(100%) Empty=0(0%) ToHead=0(0%) ToTail=0(0%) Descriptions of the output: Total: The total number of allocation and frees that occurred for a slab Fastpath: The number of allocations/frees that used the fastpath. Slowpath: Other allocations Page Alloc: Number of calls to the page allocator as a result of slowpath processing Add Partial: Number of slabs added to the partial list through free or alloc (occurs during cpuslab flushes) Remove Partial: Number of slabs removed from the partial list as a result of allocations retrieving a partial slab or by a free freeing the last object of a slab. RemoteObj/Froz: How many times were remotely freed object encountered when a slab was about to be deactivated. Frozen: How many times was free able to skip list processing because the slab was in use as the cpuslab of another processor. Flushes: Number of times the cpuslab was flushed on request (kmem_cache_shrink, may result from races in __slab_alloc) Refill: Number of times we were able to refill the cpuslab from remotely freed objects for the same slab. Deactivate: Statistics how slabs were deactivated. Shows how they were put onto the partial list. In general fastpath is very good. Slowpath without partial list processing is also desirable. Any touching of partial list uses node specific locks which may potentially cause list lock contention. Signed-off-by: Christoph Lameter <clameter@sgi.com>
2008-02-08 09:47:41 +08:00
kfree(data);
len += sysfs_emit_at(buf, len, "\n");
return len;
SLUB: Support for performance statistics The statistics provided here allow the monitoring of allocator behavior but at the cost of some (minimal) loss of performance. Counters are placed in SLUB's per cpu data structure. The per cpu structure may be extended by the statistics to grow larger than one cacheline which will increase the cache footprint of SLUB. There is a compile option to enable/disable the inclusion of the runtime statistics and its off by default. The slabinfo tool is enhanced to support these statistics via two options: -D Switches the line of information displayed for a slab from size mode to activity mode. -A Sorts the slabs displayed by activity. This allows the display of the slabs most important to the performance of a certain load. -r Report option will report detailed statistics on Example (tbench load): slabinfo -AD ->Shows the most active slabs Name Objects Alloc Free %Fast skbuff_fclone_cache 33 111953835 111953835 99 99 :0000192 2666 5283688 5281047 99 99 :0001024 849 5247230 5246389 83 83 vm_area_struct 1349 119642 118355 91 22 :0004096 15 66753 66751 98 98 :0000064 2067 25297 23383 98 78 dentry 10259 28635 18464 91 45 :0000080 11004 18950 8089 98 98 :0000096 1703 12358 10784 99 98 :0000128 762 10582 9875 94 18 :0000512 184 9807 9647 95 81 :0002048 479 9669 9195 83 65 anon_vma 777 9461 9002 99 71 kmalloc-8 6492 9981 5624 99 97 :0000768 258 7174 6931 58 15 So the skbuff_fclone_cache is of highest importance for the tbench load. Pretty high load on the 192 sized slab. Look for the aliases slabinfo -a | grep 000192 :0000192 <- xfs_btree_cur filp kmalloc-192 uid_cache tw_sock_TCP request_sock_TCPv6 tw_sock_TCPv6 skbuff_head_cache xfs_ili Likely skbuff_head_cache. Looking into the statistics of the skbuff_fclone_cache is possible through slabinfo skbuff_fclone_cache ->-r option implied if cache name is mentioned .... Usual output ... Slab Perf Counter Alloc Free %Al %Fr -------------------------------------------------- Fastpath 111953360 111946981 99 99 Slowpath 1044 7423 0 0 Page Alloc 272 264 0 0 Add partial 25 325 0 0 Remove partial 86 264 0 0 RemoteObj/SlabFrozen 350 4832 0 0 Total 111954404 111954404 Flushes 49 Refill 0 Deactivate Full=325(92%) Empty=0(0%) ToHead=24(6%) ToTail=1(0%) Looks good because the fastpath is overwhelmingly taken. skbuff_head_cache: Slab Perf Counter Alloc Free %Al %Fr -------------------------------------------------- Fastpath 5297262 5259882 99 99 Slowpath 4477 39586 0 0 Page Alloc 937 824 0 0 Add partial 0 2515 0 0 Remove partial 1691 824 0 0 RemoteObj/SlabFrozen 2621 9684 0 0 Total 5301739 5299468 Deactivate Full=2620(100%) Empty=0(0%) ToHead=0(0%) ToTail=0(0%) Descriptions of the output: Total: The total number of allocation and frees that occurred for a slab Fastpath: The number of allocations/frees that used the fastpath. Slowpath: Other allocations Page Alloc: Number of calls to the page allocator as a result of slowpath processing Add Partial: Number of slabs added to the partial list through free or alloc (occurs during cpuslab flushes) Remove Partial: Number of slabs removed from the partial list as a result of allocations retrieving a partial slab or by a free freeing the last object of a slab. RemoteObj/Froz: How many times were remotely freed object encountered when a slab was about to be deactivated. Frozen: How many times was free able to skip list processing because the slab was in use as the cpuslab of another processor. Flushes: Number of times the cpuslab was flushed on request (kmem_cache_shrink, may result from races in __slab_alloc) Refill: Number of times we were able to refill the cpuslab from remotely freed objects for the same slab. Deactivate: Statistics how slabs were deactivated. Shows how they were put onto the partial list. In general fastpath is very good. Slowpath without partial list processing is also desirable. Any touching of partial list uses node specific locks which may potentially cause list lock contention. Signed-off-by: Christoph Lameter <clameter@sgi.com>
2008-02-08 09:47:41 +08:00
}
static void clear_stat(struct kmem_cache *s, enum stat_item si)
{
int cpu;
for_each_online_cpu(cpu)
per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
}
SLUB: Support for performance statistics The statistics provided here allow the monitoring of allocator behavior but at the cost of some (minimal) loss of performance. Counters are placed in SLUB's per cpu data structure. The per cpu structure may be extended by the statistics to grow larger than one cacheline which will increase the cache footprint of SLUB. There is a compile option to enable/disable the inclusion of the runtime statistics and its off by default. The slabinfo tool is enhanced to support these statistics via two options: -D Switches the line of information displayed for a slab from size mode to activity mode. -A Sorts the slabs displayed by activity. This allows the display of the slabs most important to the performance of a certain load. -r Report option will report detailed statistics on Example (tbench load): slabinfo -AD ->Shows the most active slabs Name Objects Alloc Free %Fast skbuff_fclone_cache 33 111953835 111953835 99 99 :0000192 2666 5283688 5281047 99 99 :0001024 849 5247230 5246389 83 83 vm_area_struct 1349 119642 118355 91 22 :0004096 15 66753 66751 98 98 :0000064 2067 25297 23383 98 78 dentry 10259 28635 18464 91 45 :0000080 11004 18950 8089 98 98 :0000096 1703 12358 10784 99 98 :0000128 762 10582 9875 94 18 :0000512 184 9807 9647 95 81 :0002048 479 9669 9195 83 65 anon_vma 777 9461 9002 99 71 kmalloc-8 6492 9981 5624 99 97 :0000768 258 7174 6931 58 15 So the skbuff_fclone_cache is of highest importance for the tbench load. Pretty high load on the 192 sized slab. Look for the aliases slabinfo -a | grep 000192 :0000192 <- xfs_btree_cur filp kmalloc-192 uid_cache tw_sock_TCP request_sock_TCPv6 tw_sock_TCPv6 skbuff_head_cache xfs_ili Likely skbuff_head_cache. Looking into the statistics of the skbuff_fclone_cache is possible through slabinfo skbuff_fclone_cache ->-r option implied if cache name is mentioned .... Usual output ... Slab Perf Counter Alloc Free %Al %Fr -------------------------------------------------- Fastpath 111953360 111946981 99 99 Slowpath 1044 7423 0 0 Page Alloc 272 264 0 0 Add partial 25 325 0 0 Remove partial 86 264 0 0 RemoteObj/SlabFrozen 350 4832 0 0 Total 111954404 111954404 Flushes 49 Refill 0 Deactivate Full=325(92%) Empty=0(0%) ToHead=24(6%) ToTail=1(0%) Looks good because the fastpath is overwhelmingly taken. skbuff_head_cache: Slab Perf Counter Alloc Free %Al %Fr -------------------------------------------------- Fastpath 5297262 5259882 99 99 Slowpath 4477 39586 0 0 Page Alloc 937 824 0 0 Add partial 0 2515 0 0 Remove partial 1691 824 0 0 RemoteObj/SlabFrozen 2621 9684 0 0 Total 5301739 5299468 Deactivate Full=2620(100%) Empty=0(0%) ToHead=0(0%) ToTail=0(0%) Descriptions of the output: Total: The total number of allocation and frees that occurred for a slab Fastpath: The number of allocations/frees that used the fastpath. Slowpath: Other allocations Page Alloc: Number of calls to the page allocator as a result of slowpath processing Add Partial: Number of slabs added to the partial list through free or alloc (occurs during cpuslab flushes) Remove Partial: Number of slabs removed from the partial list as a result of allocations retrieving a partial slab or by a free freeing the last object of a slab. RemoteObj/Froz: How many times were remotely freed object encountered when a slab was about to be deactivated. Frozen: How many times was free able to skip list processing because the slab was in use as the cpuslab of another processor. Flushes: Number of times the cpuslab was flushed on request (kmem_cache_shrink, may result from races in __slab_alloc) Refill: Number of times we were able to refill the cpuslab from remotely freed objects for the same slab. Deactivate: Statistics how slabs were deactivated. Shows how they were put onto the partial list. In general fastpath is very good. Slowpath without partial list processing is also desirable. Any touching of partial list uses node specific locks which may potentially cause list lock contention. Signed-off-by: Christoph Lameter <clameter@sgi.com>
2008-02-08 09:47:41 +08:00
#define STAT_ATTR(si, text) \
static ssize_t text##_show(struct kmem_cache *s, char *buf) \
{ \
return show_stat(s, buf, si); \
} \
static ssize_t text##_store(struct kmem_cache *s, \
const char *buf, size_t length) \
{ \
if (buf[0] != '0') \
return -EINVAL; \
clear_stat(s, si); \
return length; \
} \
SLAB_ATTR(text); \
SLUB: Support for performance statistics The statistics provided here allow the monitoring of allocator behavior but at the cost of some (minimal) loss of performance. Counters are placed in SLUB's per cpu data structure. The per cpu structure may be extended by the statistics to grow larger than one cacheline which will increase the cache footprint of SLUB. There is a compile option to enable/disable the inclusion of the runtime statistics and its off by default. The slabinfo tool is enhanced to support these statistics via two options: -D Switches the line of information displayed for a slab from size mode to activity mode. -A Sorts the slabs displayed by activity. This allows the display of the slabs most important to the performance of a certain load. -r Report option will report detailed statistics on Example (tbench load): slabinfo -AD ->Shows the most active slabs Name Objects Alloc Free %Fast skbuff_fclone_cache 33 111953835 111953835 99 99 :0000192 2666 5283688 5281047 99 99 :0001024 849 5247230 5246389 83 83 vm_area_struct 1349 119642 118355 91 22 :0004096 15 66753 66751 98 98 :0000064 2067 25297 23383 98 78 dentry 10259 28635 18464 91 45 :0000080 11004 18950 8089 98 98 :0000096 1703 12358 10784 99 98 :0000128 762 10582 9875 94 18 :0000512 184 9807 9647 95 81 :0002048 479 9669 9195 83 65 anon_vma 777 9461 9002 99 71 kmalloc-8 6492 9981 5624 99 97 :0000768 258 7174 6931 58 15 So the skbuff_fclone_cache is of highest importance for the tbench load. Pretty high load on the 192 sized slab. Look for the aliases slabinfo -a | grep 000192 :0000192 <- xfs_btree_cur filp kmalloc-192 uid_cache tw_sock_TCP request_sock_TCPv6 tw_sock_TCPv6 skbuff_head_cache xfs_ili Likely skbuff_head_cache. Looking into the statistics of the skbuff_fclone_cache is possible through slabinfo skbuff_fclone_cache ->-r option implied if cache name is mentioned .... Usual output ... Slab Perf Counter Alloc Free %Al %Fr -------------------------------------------------- Fastpath 111953360 111946981 99 99 Slowpath 1044 7423 0 0 Page Alloc 272 264 0 0 Add partial 25 325 0 0 Remove partial 86 264 0 0 RemoteObj/SlabFrozen 350 4832 0 0 Total 111954404 111954404 Flushes 49 Refill 0 Deactivate Full=325(92%) Empty=0(0%) ToHead=24(6%) ToTail=1(0%) Looks good because the fastpath is overwhelmingly taken. skbuff_head_cache: Slab Perf Counter Alloc Free %Al %Fr -------------------------------------------------- Fastpath 5297262 5259882 99 99 Slowpath 4477 39586 0 0 Page Alloc 937 824 0 0 Add partial 0 2515 0 0 Remove partial 1691 824 0 0 RemoteObj/SlabFrozen 2621 9684 0 0 Total 5301739 5299468 Deactivate Full=2620(100%) Empty=0(0%) ToHead=0(0%) ToTail=0(0%) Descriptions of the output: Total: The total number of allocation and frees that occurred for a slab Fastpath: The number of allocations/frees that used the fastpath. Slowpath: Other allocations Page Alloc: Number of calls to the page allocator as a result of slowpath processing Add Partial: Number of slabs added to the partial list through free or alloc (occurs during cpuslab flushes) Remove Partial: Number of slabs removed from the partial list as a result of allocations retrieving a partial slab or by a free freeing the last object of a slab. RemoteObj/Froz: How many times were remotely freed object encountered when a slab was about to be deactivated. Frozen: How many times was free able to skip list processing because the slab was in use as the cpuslab of another processor. Flushes: Number of times the cpuslab was flushed on request (kmem_cache_shrink, may result from races in __slab_alloc) Refill: Number of times we were able to refill the cpuslab from remotely freed objects for the same slab. Deactivate: Statistics how slabs were deactivated. Shows how they were put onto the partial list. In general fastpath is very good. Slowpath without partial list processing is also desirable. Any touching of partial list uses node specific locks which may potentially cause list lock contention. Signed-off-by: Christoph Lameter <clameter@sgi.com>
2008-02-08 09:47:41 +08:00
STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
STAT_ATTR(FREE_FASTPATH, free_fastpath);
STAT_ATTR(FREE_SLOWPATH, free_slowpath);
STAT_ATTR(FREE_FROZEN, free_frozen);
STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
STAT_ATTR(ALLOC_SLAB, alloc_slab);
STAT_ATTR(ALLOC_REFILL, alloc_refill);
STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
SLUB: Support for performance statistics The statistics provided here allow the monitoring of allocator behavior but at the cost of some (minimal) loss of performance. Counters are placed in SLUB's per cpu data structure. The per cpu structure may be extended by the statistics to grow larger than one cacheline which will increase the cache footprint of SLUB. There is a compile option to enable/disable the inclusion of the runtime statistics and its off by default. The slabinfo tool is enhanced to support these statistics via two options: -D Switches the line of information displayed for a slab from size mode to activity mode. -A Sorts the slabs displayed by activity. This allows the display of the slabs most important to the performance of a certain load. -r Report option will report detailed statistics on Example (tbench load): slabinfo -AD ->Shows the most active slabs Name Objects Alloc Free %Fast skbuff_fclone_cache 33 111953835 111953835 99 99 :0000192 2666 5283688 5281047 99 99 :0001024 849 5247230 5246389 83 83 vm_area_struct 1349 119642 118355 91 22 :0004096 15 66753 66751 98 98 :0000064 2067 25297 23383 98 78 dentry 10259 28635 18464 91 45 :0000080 11004 18950 8089 98 98 :0000096 1703 12358 10784 99 98 :0000128 762 10582 9875 94 18 :0000512 184 9807 9647 95 81 :0002048 479 9669 9195 83 65 anon_vma 777 9461 9002 99 71 kmalloc-8 6492 9981 5624 99 97 :0000768 258 7174 6931 58 15 So the skbuff_fclone_cache is of highest importance for the tbench load. Pretty high load on the 192 sized slab. Look for the aliases slabinfo -a | grep 000192 :0000192 <- xfs_btree_cur filp kmalloc-192 uid_cache tw_sock_TCP request_sock_TCPv6 tw_sock_TCPv6 skbuff_head_cache xfs_ili Likely skbuff_head_cache. Looking into the statistics of the skbuff_fclone_cache is possible through slabinfo skbuff_fclone_cache ->-r option implied if cache name is mentioned .... Usual output ... Slab Perf Counter Alloc Free %Al %Fr -------------------------------------------------- Fastpath 111953360 111946981 99 99 Slowpath 1044 7423 0 0 Page Alloc 272 264 0 0 Add partial 25 325 0 0 Remove partial 86 264 0 0 RemoteObj/SlabFrozen 350 4832 0 0 Total 111954404 111954404 Flushes 49 Refill 0 Deactivate Full=325(92%) Empty=0(0%) ToHead=24(6%) ToTail=1(0%) Looks good because the fastpath is overwhelmingly taken. skbuff_head_cache: Slab Perf Counter Alloc Free %Al %Fr -------------------------------------------------- Fastpath 5297262 5259882 99 99 Slowpath 4477 39586 0 0 Page Alloc 937 824 0 0 Add partial 0 2515 0 0 Remove partial 1691 824 0 0 RemoteObj/SlabFrozen 2621 9684 0 0 Total 5301739 5299468 Deactivate Full=2620(100%) Empty=0(0%) ToHead=0(0%) ToTail=0(0%) Descriptions of the output: Total: The total number of allocation and frees that occurred for a slab Fastpath: The number of allocations/frees that used the fastpath. Slowpath: Other allocations Page Alloc: Number of calls to the page allocator as a result of slowpath processing Add Partial: Number of slabs added to the partial list through free or alloc (occurs during cpuslab flushes) Remove Partial: Number of slabs removed from the partial list as a result of allocations retrieving a partial slab or by a free freeing the last object of a slab. RemoteObj/Froz: How many times were remotely freed object encountered when a slab was about to be deactivated. Frozen: How many times was free able to skip list processing because the slab was in use as the cpuslab of another processor. Flushes: Number of times the cpuslab was flushed on request (kmem_cache_shrink, may result from races in __slab_alloc) Refill: Number of times we were able to refill the cpuslab from remotely freed objects for the same slab. Deactivate: Statistics how slabs were deactivated. Shows how they were put onto the partial list. In general fastpath is very good. Slowpath without partial list processing is also desirable. Any touching of partial list uses node specific locks which may potentially cause list lock contention. Signed-off-by: Christoph Lameter <clameter@sgi.com>
2008-02-08 09:47:41 +08:00
STAT_ATTR(FREE_SLAB, free_slab);
STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
STAT_ATTR(ORDER_FALLBACK, order_fallback);
STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
#endif /* CONFIG_SLUB_STATS */
SLUB: Support for performance statistics The statistics provided here allow the monitoring of allocator behavior but at the cost of some (minimal) loss of performance. Counters are placed in SLUB's per cpu data structure. The per cpu structure may be extended by the statistics to grow larger than one cacheline which will increase the cache footprint of SLUB. There is a compile option to enable/disable the inclusion of the runtime statistics and its off by default. The slabinfo tool is enhanced to support these statistics via two options: -D Switches the line of information displayed for a slab from size mode to activity mode. -A Sorts the slabs displayed by activity. This allows the display of the slabs most important to the performance of a certain load. -r Report option will report detailed statistics on Example (tbench load): slabinfo -AD ->Shows the most active slabs Name Objects Alloc Free %Fast skbuff_fclone_cache 33 111953835 111953835 99 99 :0000192 2666 5283688 5281047 99 99 :0001024 849 5247230 5246389 83 83 vm_area_struct 1349 119642 118355 91 22 :0004096 15 66753 66751 98 98 :0000064 2067 25297 23383 98 78 dentry 10259 28635 18464 91 45 :0000080 11004 18950 8089 98 98 :0000096 1703 12358 10784 99 98 :0000128 762 10582 9875 94 18 :0000512 184 9807 9647 95 81 :0002048 479 9669 9195 83 65 anon_vma 777 9461 9002 99 71 kmalloc-8 6492 9981 5624 99 97 :0000768 258 7174 6931 58 15 So the skbuff_fclone_cache is of highest importance for the tbench load. Pretty high load on the 192 sized slab. Look for the aliases slabinfo -a | grep 000192 :0000192 <- xfs_btree_cur filp kmalloc-192 uid_cache tw_sock_TCP request_sock_TCPv6 tw_sock_TCPv6 skbuff_head_cache xfs_ili Likely skbuff_head_cache. Looking into the statistics of the skbuff_fclone_cache is possible through slabinfo skbuff_fclone_cache ->-r option implied if cache name is mentioned .... Usual output ... Slab Perf Counter Alloc Free %Al %Fr -------------------------------------------------- Fastpath 111953360 111946981 99 99 Slowpath 1044 7423 0 0 Page Alloc 272 264 0 0 Add partial 25 325 0 0 Remove partial 86 264 0 0 RemoteObj/SlabFrozen 350 4832 0 0 Total 111954404 111954404 Flushes 49 Refill 0 Deactivate Full=325(92%) Empty=0(0%) ToHead=24(6%) ToTail=1(0%) Looks good because the fastpath is overwhelmingly taken. skbuff_head_cache: Slab Perf Counter Alloc Free %Al %Fr -------------------------------------------------- Fastpath 5297262 5259882 99 99 Slowpath 4477 39586 0 0 Page Alloc 937 824 0 0 Add partial 0 2515 0 0 Remove partial 1691 824 0 0 RemoteObj/SlabFrozen 2621 9684 0 0 Total 5301739 5299468 Deactivate Full=2620(100%) Empty=0(0%) ToHead=0(0%) ToTail=0(0%) Descriptions of the output: Total: The total number of allocation and frees that occurred for a slab Fastpath: The number of allocations/frees that used the fastpath. Slowpath: Other allocations Page Alloc: Number of calls to the page allocator as a result of slowpath processing Add Partial: Number of slabs added to the partial list through free or alloc (occurs during cpuslab flushes) Remove Partial: Number of slabs removed from the partial list as a result of allocations retrieving a partial slab or by a free freeing the last object of a slab. RemoteObj/Froz: How many times were remotely freed object encountered when a slab was about to be deactivated. Frozen: How many times was free able to skip list processing because the slab was in use as the cpuslab of another processor. Flushes: Number of times the cpuslab was flushed on request (kmem_cache_shrink, may result from races in __slab_alloc) Refill: Number of times we were able to refill the cpuslab from remotely freed objects for the same slab. Deactivate: Statistics how slabs were deactivated. Shows how they were put onto the partial list. In general fastpath is very good. Slowpath without partial list processing is also desirable. Any touching of partial list uses node specific locks which may potentially cause list lock contention. Signed-off-by: Christoph Lameter <clameter@sgi.com>
2008-02-08 09:47:41 +08:00
static struct attribute *slab_attrs[] = {
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
&slab_size_attr.attr,
&object_size_attr.attr,
&objs_per_slab_attr.attr,
&order_attr.attr,
slub: add min_partial sysfs tunable Now that a cache's min_partial has been moved to struct kmem_cache, it's possible to easily tune it from userspace by adding a sysfs attribute. It may not be desirable to keep a large number of partial slabs around if a cache is used infrequently and memory, especially when constrained by a cgroup, is scarce. It's better to allow userspace to set the minimum policy per cache instead of relying explicitly on kmem_cache_shrink(). The memory savings from simply moving min_partial from struct kmem_cache_node to struct kmem_cache is obviously not significant (unless maybe you're from SGI or something), at the largest it's # allocated caches * (MAX_NUMNODES - 1) * sizeof(unsigned long) The true savings occurs when userspace reduces the number of partial slabs that would otherwise be wasted, especially on machines with a large number of nodes (ia64 with CONFIG_NODES_SHIFT at 10 for default?). As well as the kernel estimates ideal values for n->min_partial and ensures it's within a sane range, userspace has no other input other than writing to /sys/kernel/slab/cache/shrink. There simply isn't any better heuristic to add when calculating the partial values for a better estimate that works for all possible caches. And since it's currently a static value, the user really has no way of reclaiming that wasted space, which can be significant when constrained by a cgroup (either cpusets or, later, memory controller slab limits) without shrinking it entirely. This also allows the user to specify that increased fragmentation and more partial slabs are actually desired to avoid the cost of allocating new slabs at runtime for specific caches. There's also no reason why this should be a per-struct kmem_cache_node value in the first place. You could argue that a machine would have such node size asymmetries that it should be specified on a per-node basis, but we know nobody is doing that right now since it's a purely static value at the moment and there's no convenient way to tune that via slub's sysfs interface. Cc: Christoph Lameter <cl@linux-foundation.org> Signed-off-by: David Rientjes <rientjes@google.com> Signed-off-by: Pekka Enberg <penberg@cs.helsinki.fi>
2009-02-23 09:40:09 +08:00
&min_partial_attr.attr,
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
&cpu_partial_attr.attr,
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
&objects_attr.attr,
&objects_partial_attr.attr,
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
&partial_attr.attr,
&cpu_slabs_attr.attr,
&ctor_attr.attr,
&aliases_attr.attr,
&align_attr.attr,
&hwcache_align_attr.attr,
&reclaim_account_attr.attr,
&destroy_by_rcu_attr.attr,
&shrink_attr.attr,
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
&slabs_cpu_partial_attr.attr,
#ifdef CONFIG_SLUB_DEBUG
&total_objects_attr.attr,
&slabs_attr.attr,
&sanity_checks_attr.attr,
&trace_attr.attr,
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
&red_zone_attr.attr,
&poison_attr.attr,
&store_user_attr.attr,
slub: validation of slabs (metadata and guard zones) This enables validation of slab. Validation means that all objects are checked to see if there are redzone violations, if padding has been overwritten or any pointers have been corrupted. Also checks the consistency of slab counters. Validation enables the detection of metadata corruption without the kernel having to execute code that actually uses (allocs/frees) and object. It allows one to make sure that the slab metainformation and the guard values around an object have not been compromised. A single slabcache can be checked by writing a 1 to the "validate" file. i.e. echo 1 >/sys/slab/kmalloc-128/validate or use the slabinfo tool to check all slabs slabinfo -v Error messages will show up in the syslog. Note that validation can only reach slabs that are on a list. This means that we are usually restricted to partial slabs and active slabs unless SLAB_STORE_USER is active which will build a full slab list and allows validation of slabs that are fully in use. Booting with "slub_debug" set will enable SLAB_STORE_USER and then full diagnostic are available. Note that we attempt to push cpu slabs back to the lists when we start the check. If the cpu slab is reactivated before we get to it (another processor grabs it before we get to it) then it cannot be checked. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:43 +08:00
&validate_attr.attr,
slub: add ability to list alloc / free callers per slab This patch enables listing the callers who allocated or freed objects in a cache. For example to list the allocators for kmalloc-128 do cat /sys/slab/kmalloc-128/alloc_calls 7 sn_io_slot_fixup+0x40/0x700 7 sn_io_slot_fixup+0x80/0x700 9 sn_bus_fixup+0xe0/0x380 6 param_sysfs_setup+0xf0/0x280 276 percpu_populate+0xf0/0x1a0 19 __register_chrdev_region+0x30/0x360 8 expand_files+0x2e0/0x6e0 1 sys_epoll_create+0x60/0x200 1 __mounts_open+0x140/0x2c0 65 kmem_alloc+0x110/0x280 3 alloc_disk_node+0xe0/0x200 33 as_get_io_context+0x90/0x280 74 kobject_kset_add_dir+0x40/0x140 12 pci_create_bus+0x2a0/0x5c0 1 acpi_ev_create_gpe_block+0x120/0x9e0 41 con_insert_unipair+0x100/0x1c0 1 uart_open+0x1c0/0xba0 1 dma_pool_create+0xe0/0x340 2 neigh_table_init_no_netlink+0x260/0x4c0 6 neigh_parms_alloc+0x30/0x200 1 netlink_kernel_create+0x130/0x320 5 fz_hash_alloc+0x50/0xe0 2 sn_common_hubdev_init+0xd0/0x6e0 28 kernel_param_sysfs_setup+0x30/0x180 72 process_zones+0x70/0x2e0 cat /sys/slab/kmalloc-128/free_calls 558 <not-available> 3 sn_io_slot_fixup+0x600/0x700 84 free_fdtable_rcu+0x120/0x260 2 seq_release+0x40/0x60 6 kmem_free+0x70/0xc0 24 free_as_io_context+0x20/0x200 1 acpi_get_object_info+0x3a0/0x3e0 1 acpi_add_single_object+0xcf0/0x1e40 2 con_release_unimap+0x80/0x140 1 free+0x20/0x40 SLAB_STORE_USER must be enabled for a slab cache by either booting with "slab_debug" or enabling user tracking specifically for the slab of interest. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:45 +08:00
&alloc_calls_attr.attr,
&free_calls_attr.attr,
#endif
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
#ifdef CONFIG_ZONE_DMA
&cache_dma_attr.attr,
#endif
#ifdef CONFIG_NUMA
&remote_node_defrag_ratio_attr.attr,
SLUB: Support for performance statistics The statistics provided here allow the monitoring of allocator behavior but at the cost of some (minimal) loss of performance. Counters are placed in SLUB's per cpu data structure. The per cpu structure may be extended by the statistics to grow larger than one cacheline which will increase the cache footprint of SLUB. There is a compile option to enable/disable the inclusion of the runtime statistics and its off by default. The slabinfo tool is enhanced to support these statistics via two options: -D Switches the line of information displayed for a slab from size mode to activity mode. -A Sorts the slabs displayed by activity. This allows the display of the slabs most important to the performance of a certain load. -r Report option will report detailed statistics on Example (tbench load): slabinfo -AD ->Shows the most active slabs Name Objects Alloc Free %Fast skbuff_fclone_cache 33 111953835 111953835 99 99 :0000192 2666 5283688 5281047 99 99 :0001024 849 5247230 5246389 83 83 vm_area_struct 1349 119642 118355 91 22 :0004096 15 66753 66751 98 98 :0000064 2067 25297 23383 98 78 dentry 10259 28635 18464 91 45 :0000080 11004 18950 8089 98 98 :0000096 1703 12358 10784 99 98 :0000128 762 10582 9875 94 18 :0000512 184 9807 9647 95 81 :0002048 479 9669 9195 83 65 anon_vma 777 9461 9002 99 71 kmalloc-8 6492 9981 5624 99 97 :0000768 258 7174 6931 58 15 So the skbuff_fclone_cache is of highest importance for the tbench load. Pretty high load on the 192 sized slab. Look for the aliases slabinfo -a | grep 000192 :0000192 <- xfs_btree_cur filp kmalloc-192 uid_cache tw_sock_TCP request_sock_TCPv6 tw_sock_TCPv6 skbuff_head_cache xfs_ili Likely skbuff_head_cache. Looking into the statistics of the skbuff_fclone_cache is possible through slabinfo skbuff_fclone_cache ->-r option implied if cache name is mentioned .... Usual output ... Slab Perf Counter Alloc Free %Al %Fr -------------------------------------------------- Fastpath 111953360 111946981 99 99 Slowpath 1044 7423 0 0 Page Alloc 272 264 0 0 Add partial 25 325 0 0 Remove partial 86 264 0 0 RemoteObj/SlabFrozen 350 4832 0 0 Total 111954404 111954404 Flushes 49 Refill 0 Deactivate Full=325(92%) Empty=0(0%) ToHead=24(6%) ToTail=1(0%) Looks good because the fastpath is overwhelmingly taken. skbuff_head_cache: Slab Perf Counter Alloc Free %Al %Fr -------------------------------------------------- Fastpath 5297262 5259882 99 99 Slowpath 4477 39586 0 0 Page Alloc 937 824 0 0 Add partial 0 2515 0 0 Remove partial 1691 824 0 0 RemoteObj/SlabFrozen 2621 9684 0 0 Total 5301739 5299468 Deactivate Full=2620(100%) Empty=0(0%) ToHead=0(0%) ToTail=0(0%) Descriptions of the output: Total: The total number of allocation and frees that occurred for a slab Fastpath: The number of allocations/frees that used the fastpath. Slowpath: Other allocations Page Alloc: Number of calls to the page allocator as a result of slowpath processing Add Partial: Number of slabs added to the partial list through free or alloc (occurs during cpuslab flushes) Remove Partial: Number of slabs removed from the partial list as a result of allocations retrieving a partial slab or by a free freeing the last object of a slab. RemoteObj/Froz: How many times were remotely freed object encountered when a slab was about to be deactivated. Frozen: How many times was free able to skip list processing because the slab was in use as the cpuslab of another processor. Flushes: Number of times the cpuslab was flushed on request (kmem_cache_shrink, may result from races in __slab_alloc) Refill: Number of times we were able to refill the cpuslab from remotely freed objects for the same slab. Deactivate: Statistics how slabs were deactivated. Shows how they were put onto the partial list. In general fastpath is very good. Slowpath without partial list processing is also desirable. Any touching of partial list uses node specific locks which may potentially cause list lock contention. Signed-off-by: Christoph Lameter <clameter@sgi.com>
2008-02-08 09:47:41 +08:00
#endif
#ifdef CONFIG_SLUB_STATS
&alloc_fastpath_attr.attr,
&alloc_slowpath_attr.attr,
&free_fastpath_attr.attr,
&free_slowpath_attr.attr,
&free_frozen_attr.attr,
&free_add_partial_attr.attr,
&free_remove_partial_attr.attr,
&alloc_from_partial_attr.attr,
&alloc_slab_attr.attr,
&alloc_refill_attr.attr,
&alloc_node_mismatch_attr.attr,
SLUB: Support for performance statistics The statistics provided here allow the monitoring of allocator behavior but at the cost of some (minimal) loss of performance. Counters are placed in SLUB's per cpu data structure. The per cpu structure may be extended by the statistics to grow larger than one cacheline which will increase the cache footprint of SLUB. There is a compile option to enable/disable the inclusion of the runtime statistics and its off by default. The slabinfo tool is enhanced to support these statistics via two options: -D Switches the line of information displayed for a slab from size mode to activity mode. -A Sorts the slabs displayed by activity. This allows the display of the slabs most important to the performance of a certain load. -r Report option will report detailed statistics on Example (tbench load): slabinfo -AD ->Shows the most active slabs Name Objects Alloc Free %Fast skbuff_fclone_cache 33 111953835 111953835 99 99 :0000192 2666 5283688 5281047 99 99 :0001024 849 5247230 5246389 83 83 vm_area_struct 1349 119642 118355 91 22 :0004096 15 66753 66751 98 98 :0000064 2067 25297 23383 98 78 dentry 10259 28635 18464 91 45 :0000080 11004 18950 8089 98 98 :0000096 1703 12358 10784 99 98 :0000128 762 10582 9875 94 18 :0000512 184 9807 9647 95 81 :0002048 479 9669 9195 83 65 anon_vma 777 9461 9002 99 71 kmalloc-8 6492 9981 5624 99 97 :0000768 258 7174 6931 58 15 So the skbuff_fclone_cache is of highest importance for the tbench load. Pretty high load on the 192 sized slab. Look for the aliases slabinfo -a | grep 000192 :0000192 <- xfs_btree_cur filp kmalloc-192 uid_cache tw_sock_TCP request_sock_TCPv6 tw_sock_TCPv6 skbuff_head_cache xfs_ili Likely skbuff_head_cache. Looking into the statistics of the skbuff_fclone_cache is possible through slabinfo skbuff_fclone_cache ->-r option implied if cache name is mentioned .... Usual output ... Slab Perf Counter Alloc Free %Al %Fr -------------------------------------------------- Fastpath 111953360 111946981 99 99 Slowpath 1044 7423 0 0 Page Alloc 272 264 0 0 Add partial 25 325 0 0 Remove partial 86 264 0 0 RemoteObj/SlabFrozen 350 4832 0 0 Total 111954404 111954404 Flushes 49 Refill 0 Deactivate Full=325(92%) Empty=0(0%) ToHead=24(6%) ToTail=1(0%) Looks good because the fastpath is overwhelmingly taken. skbuff_head_cache: Slab Perf Counter Alloc Free %Al %Fr -------------------------------------------------- Fastpath 5297262 5259882 99 99 Slowpath 4477 39586 0 0 Page Alloc 937 824 0 0 Add partial 0 2515 0 0 Remove partial 1691 824 0 0 RemoteObj/SlabFrozen 2621 9684 0 0 Total 5301739 5299468 Deactivate Full=2620(100%) Empty=0(0%) ToHead=0(0%) ToTail=0(0%) Descriptions of the output: Total: The total number of allocation and frees that occurred for a slab Fastpath: The number of allocations/frees that used the fastpath. Slowpath: Other allocations Page Alloc: Number of calls to the page allocator as a result of slowpath processing Add Partial: Number of slabs added to the partial list through free or alloc (occurs during cpuslab flushes) Remove Partial: Number of slabs removed from the partial list as a result of allocations retrieving a partial slab or by a free freeing the last object of a slab. RemoteObj/Froz: How many times were remotely freed object encountered when a slab was about to be deactivated. Frozen: How many times was free able to skip list processing because the slab was in use as the cpuslab of another processor. Flushes: Number of times the cpuslab was flushed on request (kmem_cache_shrink, may result from races in __slab_alloc) Refill: Number of times we were able to refill the cpuslab from remotely freed objects for the same slab. Deactivate: Statistics how slabs were deactivated. Shows how they were put onto the partial list. In general fastpath is very good. Slowpath without partial list processing is also desirable. Any touching of partial list uses node specific locks which may potentially cause list lock contention. Signed-off-by: Christoph Lameter <clameter@sgi.com>
2008-02-08 09:47:41 +08:00
&free_slab_attr.attr,
&cpuslab_flush_attr.attr,
&deactivate_full_attr.attr,
&deactivate_empty_attr.attr,
&deactivate_to_head_attr.attr,
&deactivate_to_tail_attr.attr,
&deactivate_remote_frees_attr.attr,
&deactivate_bypass_attr.attr,
&order_fallback_attr.attr,
&cmpxchg_double_fail_attr.attr,
&cmpxchg_double_cpu_fail_attr.attr,
slub: per cpu cache for partial pages Allow filling out the rest of the kmem_cache_cpu cacheline with pointers to partial pages. The partial page list is used in slab_free() to avoid per node lock taking. In __slab_alloc() we can then take multiple partial pages off the per node partial list in one go reducing node lock pressure. We can also use the per cpu partial list in slab_alloc() to avoid scanning partial lists for pages with free objects. The main effect of a per cpu partial list is that the per node list_lock is taken for batches of partial pages instead of individual ones. Potential future enhancements: 1. The pickup from the partial list could be perhaps be done without disabling interrupts with some work. The free path already puts the page into the per cpu partial list without disabling interrupts. 2. __slab_free() may have some code paths that could use optimization. Performance: Before After ./hackbench 100 process 200000 Time: 1953.047 1564.614 ./hackbench 100 process 20000 Time: 207.176 156.940 ./hackbench 100 process 20000 Time: 204.468 156.940 ./hackbench 100 process 20000 Time: 204.879 158.772 ./hackbench 10 process 20000 Time: 20.153 15.853 ./hackbench 10 process 20000 Time: 20.153 15.986 ./hackbench 10 process 20000 Time: 19.363 16.111 ./hackbench 1 process 20000 Time: 2.518 2.307 ./hackbench 1 process 20000 Time: 2.258 2.339 ./hackbench 1 process 20000 Time: 2.864 2.163 Signed-off-by: Christoph Lameter <cl@linux.com> Signed-off-by: Pekka Enberg <penberg@kernel.org>
2011-08-10 05:12:27 +08:00
&cpu_partial_alloc_attr.attr,
&cpu_partial_free_attr.attr,
&cpu_partial_node_attr.attr,
&cpu_partial_drain_attr.attr,
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
#endif
#ifdef CONFIG_FAILSLAB
&failslab_attr.attr,
#endif
usercopy: Prepare for usercopy whitelisting This patch prepares the slab allocator to handle caches having annotations (useroffset and usersize) defining usercopy regions. This patch is modified from Brad Spengler/PaX Team's PAX_USERCOPY whitelisting code in the last public patch of grsecurity/PaX based on my understanding of the code. Changes or omissions from the original code are mine and don't reflect the original grsecurity/PaX code. Currently, hardened usercopy performs dynamic bounds checking on slab cache objects. This is good, but still leaves a lot of kernel memory available to be copied to/from userspace in the face of bugs. To further restrict what memory is available for copying, this creates a way to whitelist specific areas of a given slab cache object for copying to/from userspace, allowing much finer granularity of access control. Slab caches that are never exposed to userspace can declare no whitelist for their objects, thereby keeping them unavailable to userspace via dynamic copy operations. (Note, an implicit form of whitelisting is the use of constant sizes in usercopy operations and get_user()/put_user(); these bypass hardened usercopy checks since these sizes cannot change at runtime.) To support this whitelist annotation, usercopy region offset and size members are added to struct kmem_cache. The slab allocator receives a new function, kmem_cache_create_usercopy(), that creates a new cache with a usercopy region defined, suitable for declaring spans of fields within the objects that get copied to/from userspace. In this patch, the default kmem_cache_create() marks the entire allocation as whitelisted, leaving it semantically unchanged. Once all fine-grained whitelists have been added (in subsequent patches), this will be changed to a usersize of 0, making caches created with kmem_cache_create() not copyable to/from userspace. After the entire usercopy whitelist series is applied, less than 15% of the slab cache memory remains exposed to potential usercopy bugs after a fresh boot: Total Slab Memory: 48074720 Usercopyable Memory: 6367532 13.2% task_struct 0.2% 4480/1630720 RAW 0.3% 300/96000 RAWv6 2.1% 1408/64768 ext4_inode_cache 3.0% 269760/8740224 dentry 11.1% 585984/5273856 mm_struct 29.1% 54912/188448 kmalloc-8 100.0% 24576/24576 kmalloc-16 100.0% 28672/28672 kmalloc-32 100.0% 81920/81920 kmalloc-192 100.0% 96768/96768 kmalloc-128 100.0% 143360/143360 names_cache 100.0% 163840/163840 kmalloc-64 100.0% 167936/167936 kmalloc-256 100.0% 339968/339968 kmalloc-512 100.0% 350720/350720 kmalloc-96 100.0% 455616/455616 kmalloc-8192 100.0% 655360/655360 kmalloc-1024 100.0% 812032/812032 kmalloc-4096 100.0% 819200/819200 kmalloc-2048 100.0% 1310720/1310720 After some kernel build workloads, the percentage (mainly driven by dentry and inode caches expanding) drops under 10%: Total Slab Memory: 95516184 Usercopyable Memory: 8497452 8.8% task_struct 0.2% 4000/1456000 RAW 0.3% 300/96000 RAWv6 2.1% 1408/64768 ext4_inode_cache 3.0% 1217280/39439872 dentry 11.1% 1623200/14608800 mm_struct 29.1% 73216/251264 kmalloc-8 100.0% 24576/24576 kmalloc-16 100.0% 28672/28672 kmalloc-32 100.0% 94208/94208 kmalloc-192 100.0% 96768/96768 kmalloc-128 100.0% 143360/143360 names_cache 100.0% 163840/163840 kmalloc-64 100.0% 245760/245760 kmalloc-256 100.0% 339968/339968 kmalloc-512 100.0% 350720/350720 kmalloc-96 100.0% 563520/563520 kmalloc-8192 100.0% 655360/655360 kmalloc-1024 100.0% 794624/794624 kmalloc-4096 100.0% 819200/819200 kmalloc-2048 100.0% 1257472/1257472 Signed-off-by: David Windsor <dave@nullcore.net> [kees: adjust commit log, split out a few extra kmalloc hunks] [kees: add field names to function declarations] [kees: convert BUGs to WARNs and fail closed] [kees: add attack surface reduction analysis to commit log] Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: linux-mm@kvack.org Cc: linux-xfs@vger.kernel.org Signed-off-by: Kees Cook <keescook@chromium.org> Acked-by: Christoph Lameter <cl@linux.com>
2017-06-11 10:50:28 +08:00
&usersize_attr.attr,
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
NULL
};
static const struct attribute_group slab_attr_group = {
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
.attrs = slab_attrs,
};
static ssize_t slab_attr_show(struct kobject *kobj,
struct attribute *attr,
char *buf)
{
struct slab_attribute *attribute;
struct kmem_cache *s;
int err;
attribute = to_slab_attr(attr);
s = to_slab(kobj);
if (!attribute->show)
return -EIO;
err = attribute->show(s, buf);
return err;
}
static ssize_t slab_attr_store(struct kobject *kobj,
struct attribute *attr,
const char *buf, size_t len)
{
struct slab_attribute *attribute;
struct kmem_cache *s;
int err;
attribute = to_slab_attr(attr);
s = to_slab(kobj);
if (!attribute->store)
return -EIO;
err = attribute->store(s, buf, len);
return err;
}
slub: use sysfs'es release mechanism for kmem_cache debugobjects warning during netfilter exit: ------------[ cut here ]------------ WARNING: CPU: 6 PID: 4178 at lib/debugobjects.c:260 debug_print_object+0x8d/0xb0() ODEBUG: free active (active state 0) object type: timer_list hint: delayed_work_timer_fn+0x0/0x20 Modules linked in: CPU: 6 PID: 4178 Comm: kworker/u16:2 Tainted: G W 3.11.0-next-20130906-sasha #3984 Workqueue: netns cleanup_net Call Trace: dump_stack+0x52/0x87 warn_slowpath_common+0x8c/0xc0 warn_slowpath_fmt+0x46/0x50 debug_print_object+0x8d/0xb0 __debug_check_no_obj_freed+0xa5/0x220 debug_check_no_obj_freed+0x15/0x20 kmem_cache_free+0x197/0x340 kmem_cache_destroy+0x86/0xe0 nf_conntrack_cleanup_net_list+0x131/0x170 nf_conntrack_pernet_exit+0x5d/0x70 ops_exit_list+0x5e/0x70 cleanup_net+0xfb/0x1c0 process_one_work+0x338/0x550 worker_thread+0x215/0x350 kthread+0xe7/0xf0 ret_from_fork+0x7c/0xb0 Also during dcookie cleanup: WARNING: CPU: 12 PID: 9725 at lib/debugobjects.c:260 debug_print_object+0x8c/0xb0() ODEBUG: free active (active state 0) object type: timer_list hint: delayed_work_timer_fn+0x0/0x20 Modules linked in: CPU: 12 PID: 9725 Comm: trinity-c141 Not tainted 3.15.0-rc2-next-20140423-sasha-00018-gc4ff6c4 #408 Call Trace: dump_stack (lib/dump_stack.c:52) warn_slowpath_common (kernel/panic.c:430) warn_slowpath_fmt (kernel/panic.c:445) debug_print_object (lib/debugobjects.c:262) __debug_check_no_obj_freed (lib/debugobjects.c:697) debug_check_no_obj_freed (lib/debugobjects.c:726) kmem_cache_free (mm/slub.c:2689 mm/slub.c:2717) kmem_cache_destroy (mm/slab_common.c:363) dcookie_unregister (fs/dcookies.c:302 fs/dcookies.c:343) event_buffer_release (arch/x86/oprofile/../../../drivers/oprofile/event_buffer.c:153) __fput (fs/file_table.c:217) ____fput (fs/file_table.c:253) task_work_run (kernel/task_work.c:125 (discriminator 1)) do_notify_resume (include/linux/tracehook.h:196 arch/x86/kernel/signal.c:751) int_signal (arch/x86/kernel/entry_64.S:807) Sysfs has a release mechanism. Use that to release the kmem_cache structure if CONFIG_SYSFS is enabled. Only slub is changed - slab currently only supports /proc/slabinfo and not /sys/kernel/slab/*. We talked about adding that and someone was working on it. [akpm@linux-foundation.org: fix CONFIG_SYSFS=n build] [akpm@linux-foundation.org: fix CONFIG_SYSFS=n build even more] Signed-off-by: Christoph Lameter <cl@linux.com> Reported-by: Sasha Levin <sasha.levin@oracle.com> Tested-by: Sasha Levin <sasha.levin@oracle.com> Acked-by: Greg KH <greg@kroah.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Pekka Enberg <penberg@kernel.org> Cc: Russell King <rmk@arm.linux.org.uk> Cc: Bart Van Assche <bvanassche@acm.org> Cc: Al Viro <viro@ZenIV.linux.org.uk> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-05-07 03:50:08 +08:00
static void kmem_cache_release(struct kobject *k)
{
slab_kmem_cache_release(to_slab(k));
}
static const struct sysfs_ops slab_sysfs_ops = {
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
.show = slab_attr_show,
.store = slab_attr_store,
};
static struct kobj_type slab_ktype = {
.sysfs_ops = &slab_sysfs_ops,
slub: use sysfs'es release mechanism for kmem_cache debugobjects warning during netfilter exit: ------------[ cut here ]------------ WARNING: CPU: 6 PID: 4178 at lib/debugobjects.c:260 debug_print_object+0x8d/0xb0() ODEBUG: free active (active state 0) object type: timer_list hint: delayed_work_timer_fn+0x0/0x20 Modules linked in: CPU: 6 PID: 4178 Comm: kworker/u16:2 Tainted: G W 3.11.0-next-20130906-sasha #3984 Workqueue: netns cleanup_net Call Trace: dump_stack+0x52/0x87 warn_slowpath_common+0x8c/0xc0 warn_slowpath_fmt+0x46/0x50 debug_print_object+0x8d/0xb0 __debug_check_no_obj_freed+0xa5/0x220 debug_check_no_obj_freed+0x15/0x20 kmem_cache_free+0x197/0x340 kmem_cache_destroy+0x86/0xe0 nf_conntrack_cleanup_net_list+0x131/0x170 nf_conntrack_pernet_exit+0x5d/0x70 ops_exit_list+0x5e/0x70 cleanup_net+0xfb/0x1c0 process_one_work+0x338/0x550 worker_thread+0x215/0x350 kthread+0xe7/0xf0 ret_from_fork+0x7c/0xb0 Also during dcookie cleanup: WARNING: CPU: 12 PID: 9725 at lib/debugobjects.c:260 debug_print_object+0x8c/0xb0() ODEBUG: free active (active state 0) object type: timer_list hint: delayed_work_timer_fn+0x0/0x20 Modules linked in: CPU: 12 PID: 9725 Comm: trinity-c141 Not tainted 3.15.0-rc2-next-20140423-sasha-00018-gc4ff6c4 #408 Call Trace: dump_stack (lib/dump_stack.c:52) warn_slowpath_common (kernel/panic.c:430) warn_slowpath_fmt (kernel/panic.c:445) debug_print_object (lib/debugobjects.c:262) __debug_check_no_obj_freed (lib/debugobjects.c:697) debug_check_no_obj_freed (lib/debugobjects.c:726) kmem_cache_free (mm/slub.c:2689 mm/slub.c:2717) kmem_cache_destroy (mm/slab_common.c:363) dcookie_unregister (fs/dcookies.c:302 fs/dcookies.c:343) event_buffer_release (arch/x86/oprofile/../../../drivers/oprofile/event_buffer.c:153) __fput (fs/file_table.c:217) ____fput (fs/file_table.c:253) task_work_run (kernel/task_work.c:125 (discriminator 1)) do_notify_resume (include/linux/tracehook.h:196 arch/x86/kernel/signal.c:751) int_signal (arch/x86/kernel/entry_64.S:807) Sysfs has a release mechanism. Use that to release the kmem_cache structure if CONFIG_SYSFS is enabled. Only slub is changed - slab currently only supports /proc/slabinfo and not /sys/kernel/slab/*. We talked about adding that and someone was working on it. [akpm@linux-foundation.org: fix CONFIG_SYSFS=n build] [akpm@linux-foundation.org: fix CONFIG_SYSFS=n build even more] Signed-off-by: Christoph Lameter <cl@linux.com> Reported-by: Sasha Levin <sasha.levin@oracle.com> Tested-by: Sasha Levin <sasha.levin@oracle.com> Acked-by: Greg KH <greg@kroah.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Pekka Enberg <penberg@kernel.org> Cc: Russell King <rmk@arm.linux.org.uk> Cc: Bart Van Assche <bvanassche@acm.org> Cc: Al Viro <viro@ZenIV.linux.org.uk> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-05-07 03:50:08 +08:00
.release = kmem_cache_release,
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
};
static struct kset *slab_kset;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
2014-04-08 06:39:31 +08:00
static inline struct kset *cache_kset(struct kmem_cache *s)
{
return slab_kset;
}
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
#define ID_STR_LENGTH 64
/* Create a unique string id for a slab cache:
*
* Format :[flags-]size
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*/
static char *create_unique_id(struct kmem_cache *s)
{
char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
char *p = name;
BUG_ON(!name);
*p++ = ':';
/*
* First flags affecting slabcache operations. We will only
* get here for aliasable slabs so we do not need to support
* too many flags. The flags here must cover all flags that
* are matched during merging to guarantee that the id is
* unique.
*/
if (s->flags & SLAB_CACHE_DMA)
*p++ = 'd';
mm: add support for kmem caches in DMA32 zone Patch series "iommu/io-pgtable-arm-v7s: Use DMA32 zone for page tables", v6. This is a followup to the discussion in [1], [2]. IOMMUs using ARMv7 short-descriptor format require page tables (level 1 and 2) to be allocated within the first 4GB of RAM, even on 64-bit systems. For L1 tables that are bigger than a page, we can just use __get_free_pages with GFP_DMA32 (on arm64 systems only, arm would still use GFP_DMA). For L2 tables that only take 1KB, it would be a waste to allocate a full page, so we considered 3 approaches: 1. This series, adding support for GFP_DMA32 slab caches. 2. genalloc, which requires pre-allocating the maximum number of L2 page tables (4096, so 4MB of memory). 3. page_frag, which is not very memory-efficient as it is unable to reuse freed fragments until the whole page is freed. [3] This series is the most memory-efficient approach. stable@ note: We confirmed that this is a regression, and IOMMU errors happen on 4.19 and linux-next/master on MT8173 (elm, Acer Chromebook R13). The issue most likely starts from commit ad67f5a6545f ("arm64: replace ZONE_DMA with ZONE_DMA32"), i.e. 4.15, and presumably breaks a number of Mediatek platforms (and maybe others?). [1] https://lists.linuxfoundation.org/pipermail/iommu/2018-November/030876.html [2] https://lists.linuxfoundation.org/pipermail/iommu/2018-December/031696.html [3] https://patchwork.codeaurora.org/patch/671639/ This patch (of 3): IOMMUs using ARMv7 short-descriptor format require page tables to be allocated within the first 4GB of RAM, even on 64-bit systems. On arm64, this is done by passing GFP_DMA32 flag to memory allocation functions. For IOMMU L2 tables that only take 1KB, it would be a waste to allocate a full page using get_free_pages, so we considered 3 approaches: 1. This patch, adding support for GFP_DMA32 slab caches. 2. genalloc, which requires pre-allocating the maximum number of L2 page tables (4096, so 4MB of memory). 3. page_frag, which is not very memory-efficient as it is unable to reuse freed fragments until the whole page is freed. This change makes it possible to create a custom cache in DMA32 zone using kmem_cache_create, then allocate memory using kmem_cache_alloc. We do not create a DMA32 kmalloc cache array, as there are currently no users of kmalloc(..., GFP_DMA32). These calls will continue to trigger a warning, as we keep GFP_DMA32 in GFP_SLAB_BUG_MASK. This implies that calls to kmem_cache_*alloc on a SLAB_CACHE_DMA32 kmem_cache must _not_ use GFP_DMA32 (it is anyway redundant and unnecessary). Link: http://lkml.kernel.org/r/20181210011504.122604-2-drinkcat@chromium.org Signed-off-by: Nicolas Boichat <drinkcat@chromium.org> Acked-by: Vlastimil Babka <vbabka@suse.cz> Acked-by: Will Deacon <will.deacon@arm.com> Cc: Robin Murphy <robin.murphy@arm.com> Cc: Joerg Roedel <joro@8bytes.org> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Sasha Levin <Alexander.Levin@microsoft.com> Cc: Huaisheng Ye <yehs1@lenovo.com> Cc: Mike Rapoport <rppt@linux.vnet.ibm.com> Cc: Yong Wu <yong.wu@mediatek.com> Cc: Matthias Brugger <matthias.bgg@gmail.com> Cc: Tomasz Figa <tfiga@google.com> Cc: Yingjoe Chen <yingjoe.chen@mediatek.com> Cc: Christoph Hellwig <hch@infradead.org> Cc: Matthew Wilcox <willy@infradead.org> Cc: Hsin-Yi Wang <hsinyi@chromium.org> Cc: <stable@vger.kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-29 11:43:42 +08:00
if (s->flags & SLAB_CACHE_DMA32)
*p++ = 'D';
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (s->flags & SLAB_RECLAIM_ACCOUNT)
*p++ = 'a';
if (s->flags & SLAB_CONSISTENCY_CHECKS)
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*p++ = 'F';
if (s->flags & SLAB_ACCOUNT)
*p++ = 'A';
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (p != name + 1)
*p++ = '-';
p += sprintf(p, "%07u", s->size);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
BUG_ON(p > name + ID_STR_LENGTH - 1);
return name;
}
static int sysfs_slab_add(struct kmem_cache *s)
{
int err;
const char *name;
slub: make sysfs directories for memcg sub-caches optional SLUB creates a per-cache directory under /sys/kernel/slab which hosts a bunch of debug files. Usually, there aren't that many caches on a system and this doesn't really matter; however, if memcg is in use, each cache can have per-cgroup sub-caches. SLUB creates the same directories for these sub-caches under /sys/kernel/slab/$CACHE/cgroup. Unfortunately, because there can be a lot of cgroups, active or draining, the product of the numbers of caches, cgroups and files in each directory can reach a very high number - hundreds of thousands is commonplace. Millions and beyond aren't difficult to reach either. What's under /sys/kernel/slab is primarily for debugging and the information and control on the a root cache already cover its sub-caches. While having a separate directory for each sub-cache can be helpful for development, it doesn't make much sense to pay this amount of overhead by default. This patch introduces a boot parameter slub_memcg_sysfs which determines whether to create sysfs directories for per-memcg sub-caches. It also adds CONFIG_SLUB_MEMCG_SYSFS_ON which determines the boot parameter's default value and defaults to 0. [akpm@linux-foundation.org: kset_unregister(NULL) is legal] Link: http://lkml.kernel.org/r/20170204145203.GB26958@mtj.duckdns.org Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-23 07:41:39 +08:00
struct kset *kset = cache_kset(s);
int unmergeable = slab_unmergeable(s);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
slub: make sysfs directories for memcg sub-caches optional SLUB creates a per-cache directory under /sys/kernel/slab which hosts a bunch of debug files. Usually, there aren't that many caches on a system and this doesn't really matter; however, if memcg is in use, each cache can have per-cgroup sub-caches. SLUB creates the same directories for these sub-caches under /sys/kernel/slab/$CACHE/cgroup. Unfortunately, because there can be a lot of cgroups, active or draining, the product of the numbers of caches, cgroups and files in each directory can reach a very high number - hundreds of thousands is commonplace. Millions and beyond aren't difficult to reach either. What's under /sys/kernel/slab is primarily for debugging and the information and control on the a root cache already cover its sub-caches. While having a separate directory for each sub-cache can be helpful for development, it doesn't make much sense to pay this amount of overhead by default. This patch introduces a boot parameter slub_memcg_sysfs which determines whether to create sysfs directories for per-memcg sub-caches. It also adds CONFIG_SLUB_MEMCG_SYSFS_ON which determines the boot parameter's default value and defaults to 0. [akpm@linux-foundation.org: kset_unregister(NULL) is legal] Link: http://lkml.kernel.org/r/20170204145203.GB26958@mtj.duckdns.org Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-23 07:41:39 +08:00
if (!kset) {
kobject_init(&s->kobj, &slab_ktype);
return 0;
}
slub: fix sysfs duplicate filename creation when slub_debug=O When slub_debug=O is set. It is possible to clear debug flags for an "unmergeable" slab cache in kmem_cache_open(). It makes the "unmergeable" cache became "mergeable" in sysfs_slab_add(). These caches will generate their "unique IDs" by create_unique_id(), but it is possible to create identical unique IDs. In my experiment, sgpool-128, names_cache, biovec-256 generate the same ID ":Ft-0004096" and the kernel reports "sysfs: cannot create duplicate filename '/kernel/slab/:Ft-0004096'". To repeat my experiment, set disable_higher_order_debug=1, CONFIG_SLUB_DEBUG_ON=y in kernel-4.14. Fix this issue by setting unmergeable=1 if slub_debug=O and the the default slub_debug contains any no-merge flags. call path: kmem_cache_create() __kmem_cache_alias() -> we set SLAB_NEVER_MERGE flags here create_cache() __kmem_cache_create() kmem_cache_open() -> clear DEBUG_METADATA_FLAGS sysfs_slab_add() -> the slab cache is mergeable now sysfs: cannot create duplicate filename '/kernel/slab/:Ft-0004096' ------------[ cut here ]------------ WARNING: CPU: 0 PID: 1 at fs/sysfs/dir.c:31 sysfs_warn_dup+0x60/0x7c Modules linked in: CPU: 0 PID: 1 Comm: swapper/0 Tainted: G W 4.14.0-rc7ajb-00131-gd4c2e9f-dirty #123 Hardware name: linux,dummy-virt (DT) task: ffffffc07d4e0080 task.stack: ffffff8008008000 PC is at sysfs_warn_dup+0x60/0x7c LR is at sysfs_warn_dup+0x60/0x7c pc : lr : pstate: 60000145 Call trace: sysfs_warn_dup+0x60/0x7c sysfs_create_dir_ns+0x98/0xa0 kobject_add_internal+0xa0/0x294 kobject_init_and_add+0x90/0xb4 sysfs_slab_add+0x90/0x200 __kmem_cache_create+0x26c/0x438 kmem_cache_create+0x164/0x1f4 sg_pool_init+0x60/0x100 do_one_initcall+0x38/0x12c kernel_init_freeable+0x138/0x1d4 kernel_init+0x10/0xfc ret_from_fork+0x10/0x18 Link: http://lkml.kernel.org/r/1510365805-5155-1-git-send-email-miles.chen@mediatek.com Signed-off-by: Miles Chen <miles.chen@mediatek.com> Acked-by: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-16 09:32:25 +08:00
if (!unmergeable && disable_higher_order_debug &&
(slub_debug & DEBUG_METADATA_FLAGS))
unmergeable = 1;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (unmergeable) {
/*
* Slabcache can never be merged so we can use the name proper.
* This is typically the case for debug situations. In that
* case we can catch duplicate names easily.
*/
sysfs_remove_link(&slab_kset->kobj, s->name);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
name = s->name;
} else {
/*
* Create a unique name for the slab as a target
* for the symlinks.
*/
name = create_unique_id(s);
}
slub: make sysfs directories for memcg sub-caches optional SLUB creates a per-cache directory under /sys/kernel/slab which hosts a bunch of debug files. Usually, there aren't that many caches on a system and this doesn't really matter; however, if memcg is in use, each cache can have per-cgroup sub-caches. SLUB creates the same directories for these sub-caches under /sys/kernel/slab/$CACHE/cgroup. Unfortunately, because there can be a lot of cgroups, active or draining, the product of the numbers of caches, cgroups and files in each directory can reach a very high number - hundreds of thousands is commonplace. Millions and beyond aren't difficult to reach either. What's under /sys/kernel/slab is primarily for debugging and the information and control on the a root cache already cover its sub-caches. While having a separate directory for each sub-cache can be helpful for development, it doesn't make much sense to pay this amount of overhead by default. This patch introduces a boot parameter slub_memcg_sysfs which determines whether to create sysfs directories for per-memcg sub-caches. It also adds CONFIG_SLUB_MEMCG_SYSFS_ON which determines the boot parameter's default value and defaults to 0. [akpm@linux-foundation.org: kset_unregister(NULL) is legal] Link: http://lkml.kernel.org/r/20170204145203.GB26958@mtj.duckdns.org Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-23 07:41:39 +08:00
s->kobj.kset = kset;
err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
if (err)
goto out;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
err = sysfs_create_group(&s->kobj, &slab_attr_group);
if (err)
goto out_del_kobj;
2014-04-08 06:39:31 +08:00
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
if (!unmergeable) {
/* Setup first alias */
sysfs_slab_alias(s, s->name);
}
out:
if (!unmergeable)
kfree(name);
return err;
out_del_kobj:
kobject_del(&s->kobj);
goto out;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
slub: fix failure when we delete and create a slab cache In kernel 4.17 I removed some code from dm-bufio that did slab cache merging (commit 21bb13276768: "dm bufio: remove code that merges slab caches") - both slab and slub support merging caches with identical attributes, so dm-bufio now just calls kmem_cache_create and relies on implicit merging. This uncovered a bug in the slub subsystem - if we delete a cache and immediatelly create another cache with the same attributes, it fails because of duplicate filename in /sys/kernel/slab/. The slub subsystem offloads freeing the cache to a workqueue - and if we create the new cache before the workqueue runs, it complains because of duplicate filename in sysfs. This patch fixes the bug by moving the call of kobject_del from sysfs_slab_remove_workfn to shutdown_cache. kobject_del must be called while we hold slab_mutex - so that the sysfs entry is deleted before a cache with the same attributes could be created. Running device-mapper-test-suite with: dmtest run --suite thin-provisioning -n /commit_failure_causes_fallback/ triggered: Buffer I/O error on dev dm-0, logical block 1572848, async page read device-mapper: thin: 253:1: metadata operation 'dm_pool_alloc_data_block' failed: error = -5 device-mapper: thin: 253:1: aborting current metadata transaction sysfs: cannot create duplicate filename '/kernel/slab/:a-0000144' CPU: 2 PID: 1037 Comm: kworker/u48:1 Not tainted 4.17.0.snitm+ #25 Hardware name: Supermicro SYS-1029P-WTR/X11DDW-L, BIOS 2.0a 12/06/2017 Workqueue: dm-thin do_worker [dm_thin_pool] Call Trace: dump_stack+0x5a/0x73 sysfs_warn_dup+0x58/0x70 sysfs_create_dir_ns+0x77/0x80 kobject_add_internal+0xba/0x2e0 kobject_init_and_add+0x70/0xb0 sysfs_slab_add+0xb1/0x250 __kmem_cache_create+0x116/0x150 create_cache+0xd9/0x1f0 kmem_cache_create_usercopy+0x1c1/0x250 kmem_cache_create+0x18/0x20 dm_bufio_client_create+0x1ae/0x410 [dm_bufio] dm_block_manager_create+0x5e/0x90 [dm_persistent_data] __create_persistent_data_objects+0x38/0x940 [dm_thin_pool] dm_pool_abort_metadata+0x64/0x90 [dm_thin_pool] metadata_operation_failed+0x59/0x100 [dm_thin_pool] alloc_data_block.isra.53+0x86/0x180 [dm_thin_pool] process_cell+0x2a3/0x550 [dm_thin_pool] do_worker+0x28d/0x8f0 [dm_thin_pool] process_one_work+0x171/0x370 worker_thread+0x49/0x3f0 kthread+0xf8/0x130 ret_from_fork+0x35/0x40 kobject_add_internal failed for :a-0000144 with -EEXIST, don't try to register things with the same name in the same directory. kmem_cache_create(dm_bufio_buffer-16) failed with error -17 Link: http://lkml.kernel.org/r/alpine.LRH.2.02.1806151817130.6333@file01.intranet.prod.int.rdu2.redhat.com Signed-off-by: Mikulas Patocka <mpatocka@redhat.com> Reported-by: Mike Snitzer <snitzer@redhat.com> Tested-by: Mike Snitzer <snitzer@redhat.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: <stable@vger.kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-06-28 14:26:09 +08:00
void sysfs_slab_unlink(struct kmem_cache *s)
{
if (slab_state >= FULL)
kobject_del(&s->kobj);
}
void sysfs_slab_release(struct kmem_cache *s)
{
if (slab_state >= FULL)
kobject_put(&s->kobj);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
/*
* Need to buffer aliases during bootup until sysfs becomes
* available lest we lose that information.
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
*/
struct saved_alias {
struct kmem_cache *s;
const char *name;
struct saved_alias *next;
};
static struct saved_alias *alias_list;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
{
struct saved_alias *al;
if (slab_state == FULL) {
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
/*
* If we have a leftover link then remove it.
*/
sysfs_remove_link(&slab_kset->kobj, name);
return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
}
al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
if (!al)
return -ENOMEM;
al->s = s;
al->name = name;
al->next = alias_list;
alias_list = al;
return 0;
}
static int __init slab_sysfs_init(void)
{
struct kmem_cache *s;
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
int err;
mutex_lock(&slab_mutex);
slab_kset = kset_create_and_add("slab", NULL, kernel_kobj);
if (!slab_kset) {
mutex_unlock(&slab_mutex);
pr_err("Cannot register slab subsystem.\n");
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
return -ENOSYS;
}
slab_state = FULL;
list_for_each_entry(s, &slab_caches, list) {
err = sysfs_slab_add(s);
if (err)
pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
s->name);
}
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
while (alias_list) {
struct saved_alias *al = alias_list;
alias_list = alias_list->next;
err = sysfs_slab_alias(al->s, al->name);
if (err)
pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
al->name);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
kfree(al);
}
mutex_unlock(&slab_mutex);
SLUB core This is a new slab allocator which was motivated by the complexity of the existing code in mm/slab.c. It attempts to address a variety of concerns with the existing implementation. A. Management of object queues A particular concern was the complex management of the numerous object queues in SLAB. SLUB has no such queues. Instead we dedicate a slab for each allocating CPU and use objects from a slab directly instead of queueing them up. B. Storage overhead of object queues SLAB Object queues exist per node, per CPU. The alien cache queue even has a queue array that contain a queue for each processor on each node. For very large systems the number of queues and the number of objects that may be caught in those queues grows exponentially. On our systems with 1k nodes / processors we have several gigabytes just tied up for storing references to objects for those queues This does not include the objects that could be on those queues. One fears that the whole memory of the machine could one day be consumed by those queues. C. SLAB meta data overhead SLAB has overhead at the beginning of each slab. This means that data cannot be naturally aligned at the beginning of a slab block. SLUB keeps all meta data in the corresponding page_struct. Objects can be naturally aligned in the slab. F.e. a 128 byte object will be aligned at 128 byte boundaries and can fit tightly into a 4k page with no bytes left over. SLAB cannot do this. D. SLAB has a complex cache reaper SLUB does not need a cache reaper for UP systems. On SMP systems the per CPU slab may be pushed back into partial list but that operation is simple and does not require an iteration over a list of objects. SLAB expires per CPU, shared and alien object queues during cache reaping which may cause strange hold offs. E. SLAB has complex NUMA policy layer support SLUB pushes NUMA policy handling into the page allocator. This means that allocation is coarser (SLUB does interleave on a page level) but that situation was also present before 2.6.13. SLABs application of policies to individual slab objects allocated in SLAB is certainly a performance concern due to the frequent references to memory policies which may lead a sequence of objects to come from one node after another. SLUB will get a slab full of objects from one node and then will switch to the next. F. Reduction of the size of partial slab lists SLAB has per node partial lists. This means that over time a large number of partial slabs may accumulate on those lists. These can only be reused if allocator occur on specific nodes. SLUB has a global pool of partial slabs and will consume slabs from that pool to decrease fragmentation. G. Tunables SLAB has sophisticated tuning abilities for each slab cache. One can manipulate the queue sizes in detail. However, filling the queues still requires the uses of the spin lock to check out slabs. SLUB has a global parameter (min_slab_order) for tuning. Increasing the minimum slab order can decrease the locking overhead. The bigger the slab order the less motions of pages between per CPU and partial lists occur and the better SLUB will be scaling. G. Slab merging We often have slab caches with similar parameters. SLUB detects those on boot up and merges them into the corresponding general caches. This leads to more effective memory use. About 50% of all caches can be eliminated through slab merging. This will also decrease slab fragmentation because partial allocated slabs can be filled up again. Slab merging can be switched off by specifying slub_nomerge on boot up. Note that merging can expose heretofore unknown bugs in the kernel because corrupted objects may now be placed differently and corrupt differing neighboring objects. Enable sanity checks to find those. H. Diagnostics The current slab diagnostics are difficult to use and require a recompilation of the kernel. SLUB contains debugging code that is always available (but is kept out of the hot code paths). SLUB diagnostics can be enabled via the "slab_debug" option. Parameters can be specified to select a single or a group of slab caches for diagnostics. This means that the system is running with the usual performance and it is much more likely that race conditions can be reproduced. I. Resiliency If basic sanity checks are on then SLUB is capable of detecting common error conditions and recover as best as possible to allow the system to continue. J. Tracing Tracing can be enabled via the slab_debug=T,<slabcache> option during boot. SLUB will then protocol all actions on that slabcache and dump the object contents on free. K. On demand DMA cache creation. Generally DMA caches are not needed. If a kmalloc is used with __GFP_DMA then just create this single slabcache that is needed. For systems that have no ZONE_DMA requirement the support is completely eliminated. L. Performance increase Some benchmarks have shown speed improvements on kernbench in the range of 5-10%. The locking overhead of slub is based on the underlying base allocation size. If we can reliably allocate larger order pages then it is possible to increase slub performance much further. The anti-fragmentation patches may enable further performance increases. Tested on: i386 UP + SMP, x86_64 UP + SMP + NUMA emulation, IA64 NUMA + Simulator SLUB Boot options slub_nomerge Disable merging of slabs slub_min_order=x Require a minimum order for slab caches. This increases the managed chunk size and therefore reduces meta data and locking overhead. slub_min_objects=x Mininum objects per slab. Default is 8. slub_max_order=x Avoid generating slabs larger than order specified. slub_debug Enable all diagnostics for all caches slub_debug=<options> Enable selective options for all caches slub_debug=<o>,<cache> Enable selective options for a certain set of caches Available Debug options F Double Free checking, sanity and resiliency R Red zoning P Object / padding poisoning U Track last free / alloc T Trace all allocs / frees (only use for individual slabs). To use SLUB: Apply this patch and then select SLUB as the default slab allocator. [hugh@veritas.com: fix an oops-causing locking error] [akpm@linux-foundation.org: various stupid cleanups and small fixes] Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-07 05:49:36 +08:00
resiliency_test();
return 0;
}
__initcall(slab_sysfs_init);
#endif /* CONFIG_SYSFS */
/*
* The /proc/slabinfo ABI
*/
#ifdef CONFIG_SLUB_DEBUG
void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
{
unsigned long nr_slabs = 0;
unsigned long nr_objs = 0;
unsigned long nr_free = 0;
int node;
struct kmem_cache_node *n;
for_each_kmem_cache_node(s, node, n) {
nr_slabs += node_nr_slabs(n);
nr_objs += node_nr_objs(n);
nr_free += count_partial(n, count_free);
}
sinfo->active_objs = nr_objs - nr_free;
sinfo->num_objs = nr_objs;
sinfo->active_slabs = nr_slabs;
sinfo->num_slabs = nr_slabs;
sinfo->objects_per_slab = oo_objects(s->oo);
sinfo->cache_order = oo_order(s->oo);
}
void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
{
}
ssize_t slabinfo_write(struct file *file, const char __user *buffer,
size_t count, loff_t *ppos)
{
return -EIO;
}
#endif /* CONFIG_SLUB_DEBUG */