OpenCloudOS-Kernel/mm/z3fold.c

1588 lines
41 KiB
C
Raw Normal View History

// SPDX-License-Identifier: GPL-2.0-only
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
/*
* z3fold.c
*
* Author: Vitaly Wool <vitaly.wool@konsulko.com>
* Copyright (C) 2016, Sony Mobile Communications Inc.
*
* This implementation is based on zbud written by Seth Jennings.
*
* z3fold is an special purpose allocator for storing compressed pages. It
* can store up to three compressed pages per page which improves the
* compression ratio of zbud while retaining its main concepts (e. g. always
* storing an integral number of objects per page) and simplicity.
* It still has simple and deterministic reclaim properties that make it
* preferable to a higher density approach (with no requirement on integral
* number of object per page) when reclaim is used.
*
* As in zbud, pages are divided into "chunks". The size of the chunks is
* fixed at compile time and is determined by NCHUNKS_ORDER below.
*
* z3fold doesn't export any API and is meant to be used via zpool API.
*/
#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
#include <linux/atomic.h>
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
#include <linux/sched.h>
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
#include <linux/cpumask.h>
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
#include <linux/list.h>
#include <linux/mm.h>
#include <linux/module.h>
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
#include <linux/page-flags.h>
#include <linux/migrate.h>
#include <linux/node.h>
#include <linux/compaction.h>
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
#include <linux/percpu.h>
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
#include <linux/mount.h>
#include <linux/pseudo_fs.h>
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
#include <linux/fs.h>
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
#include <linux/preempt.h>
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
#include <linux/workqueue.h>
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
#include <linux/slab.h>
#include <linux/spinlock.h>
#include <linux/zpool.h>
#include <linux/magic.h>
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
/*
* NCHUNKS_ORDER determines the internal allocation granularity, effectively
* adjusting internal fragmentation. It also determines the number of
* freelists maintained in each pool. NCHUNKS_ORDER of 6 means that the
* allocation granularity will be in chunks of size PAGE_SIZE/64. Some chunks
* in the beginning of an allocated page are occupied by z3fold header, so
* NCHUNKS will be calculated to 63 (or 62 in case CONFIG_DEBUG_SPINLOCK=y),
* which shows the max number of free chunks in z3fold page, also there will
* be 63, or 62, respectively, freelists per pool.
*/
#define NCHUNKS_ORDER 6
#define CHUNK_SHIFT (PAGE_SHIFT - NCHUNKS_ORDER)
#define CHUNK_SIZE (1 << CHUNK_SHIFT)
#define ZHDR_SIZE_ALIGNED round_up(sizeof(struct z3fold_header), CHUNK_SIZE)
#define ZHDR_CHUNKS (ZHDR_SIZE_ALIGNED >> CHUNK_SHIFT)
#define TOTAL_CHUNKS (PAGE_SIZE >> CHUNK_SHIFT)
#define NCHUNKS ((PAGE_SIZE - ZHDR_SIZE_ALIGNED) >> CHUNK_SHIFT)
#define BUDDY_MASK (0x3)
#define BUDDY_SHIFT 2
#define SLOTS_ALIGN (0x40)
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
/*****************
* Structures
*****************/
struct z3fold_pool;
struct z3fold_ops {
int (*evict)(struct z3fold_pool *pool, unsigned long handle);
};
enum buddy {
HEADLESS = 0,
FIRST,
MIDDLE,
LAST,
BUDDIES_MAX = LAST
};
struct z3fold_buddy_slots {
/*
* we are using BUDDY_MASK in handle_to_buddy etc. so there should
* be enough slots to hold all possible variants
*/
unsigned long slot[BUDDY_MASK + 1];
unsigned long pool; /* back link + flags */
};
#define HANDLE_FLAG_MASK (0x03)
/*
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
* struct z3fold_header - z3fold page metadata occupying first chunks of each
* z3fold page, except for HEADLESS pages
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
* @buddy: links the z3fold page into the relevant list in the
* pool
* @page_lock: per-page lock
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
* @refcount: reference count for the z3fold page
* @work: work_struct for page layout optimization
* @slots: pointer to the structure holding buddy slots
* @pool: pointer to the containing pool
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
* @cpu: CPU which this page "belongs" to
* @first_chunks: the size of the first buddy in chunks, 0 if free
* @middle_chunks: the size of the middle buddy in chunks, 0 if free
* @last_chunks: the size of the last buddy in chunks, 0 if free
* @first_num: the starting number (for the first handle)
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
* @mapped_count: the number of objects currently mapped
*/
struct z3fold_header {
struct list_head buddy;
spinlock_t page_lock;
struct kref refcount;
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
struct work_struct work;
struct z3fold_buddy_slots *slots;
struct z3fold_pool *pool;
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
short cpu;
unsigned short first_chunks;
unsigned short middle_chunks;
unsigned short last_chunks;
unsigned short start_middle;
unsigned short first_num:2;
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
unsigned short mapped_count:2;
};
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
/**
* struct z3fold_pool - stores metadata for each z3fold pool
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
* @name: pool name
* @lock: protects pool unbuddied/lru lists
* @stale_lock: protects pool stale page list
* @unbuddied: per-cpu array of lists tracking z3fold pages that contain 2-
* buddies; the list each z3fold page is added to depends on
* the size of its free region.
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
* @lru: list tracking the z3fold pages in LRU order by most recently
* added buddy.
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
* @stale: list of pages marked for freeing
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
* @pages_nr: number of z3fold pages in the pool.
* @c_handle: cache for z3fold_buddy_slots allocation
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
* @ops: pointer to a structure of user defined operations specified at
* pool creation time.
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
* @compact_wq: workqueue for page layout background optimization
* @release_wq: workqueue for safe page release
* @work: work_struct for safe page release
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
* @inode: inode for z3fold pseudo filesystem
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
*
* This structure is allocated at pool creation time and maintains metadata
* pertaining to a particular z3fold pool.
*/
struct z3fold_pool {
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
const char *name;
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
spinlock_t lock;
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
spinlock_t stale_lock;
struct list_head *unbuddied;
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
struct list_head lru;
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
struct list_head stale;
atomic64_t pages_nr;
struct kmem_cache *c_handle;
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
const struct z3fold_ops *ops;
struct zpool *zpool;
const struct zpool_ops *zpool_ops;
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
struct workqueue_struct *compact_wq;
struct workqueue_struct *release_wq;
struct work_struct work;
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
struct inode *inode;
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
};
/*
* Internal z3fold page flags
*/
enum z3fold_page_flags {
PAGE_HEADLESS = 0,
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
MIDDLE_CHUNK_MAPPED,
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
NEEDS_COMPACTING,
PAGE_STALE,
PAGE_CLAIMED, /* by either reclaim or free */
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
};
/*****************
* Helpers
*****************/
/* Converts an allocation size in bytes to size in z3fold chunks */
static int size_to_chunks(size_t size)
{
return (size + CHUNK_SIZE - 1) >> CHUNK_SHIFT;
}
#define for_each_unbuddied_list(_iter, _begin) \
for ((_iter) = (_begin); (_iter) < NCHUNKS; (_iter)++)
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
static void compact_page_work(struct work_struct *w);
static inline struct z3fold_buddy_slots *alloc_slots(struct z3fold_pool *pool,
gfp_t gfp)
{
struct z3fold_buddy_slots *slots;
slots = kmem_cache_alloc(pool->c_handle,
(gfp & ~(__GFP_HIGHMEM | __GFP_MOVABLE)));
if (slots) {
memset(slots->slot, 0, sizeof(slots->slot));
slots->pool = (unsigned long)pool;
}
return slots;
}
static inline struct z3fold_pool *slots_to_pool(struct z3fold_buddy_slots *s)
{
return (struct z3fold_pool *)(s->pool & ~HANDLE_FLAG_MASK);
}
static inline struct z3fold_buddy_slots *handle_to_slots(unsigned long handle)
{
return (struct z3fold_buddy_slots *)(handle & ~(SLOTS_ALIGN - 1));
}
static inline void free_handle(unsigned long handle)
{
struct z3fold_buddy_slots *slots;
int i;
bool is_free;
if (handle & (1 << PAGE_HEADLESS))
return;
WARN_ON(*(unsigned long *)handle == 0);
*(unsigned long *)handle = 0;
slots = handle_to_slots(handle);
is_free = true;
for (i = 0; i <= BUDDY_MASK; i++) {
if (slots->slot[i]) {
is_free = false;
break;
}
}
if (is_free) {
struct z3fold_pool *pool = slots_to_pool(slots);
kmem_cache_free(pool->c_handle, slots);
}
}
static int z3fold_init_fs_context(struct fs_context *fc)
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
{
return init_pseudo(fc, Z3FOLD_MAGIC) ? 0 : -ENOMEM;
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
}
static struct file_system_type z3fold_fs = {
.name = "z3fold",
.init_fs_context = z3fold_init_fs_context,
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
.kill_sb = kill_anon_super,
};
static struct vfsmount *z3fold_mnt;
static int z3fold_mount(void)
{
int ret = 0;
z3fold_mnt = kern_mount(&z3fold_fs);
if (IS_ERR(z3fold_mnt))
ret = PTR_ERR(z3fold_mnt);
return ret;
}
static void z3fold_unmount(void)
{
kern_unmount(z3fold_mnt);
}
static const struct address_space_operations z3fold_aops;
static int z3fold_register_migration(struct z3fold_pool *pool)
{
pool->inode = alloc_anon_inode(z3fold_mnt->mnt_sb);
if (IS_ERR(pool->inode)) {
pool->inode = NULL;
return 1;
}
pool->inode->i_mapping->private_data = pool;
pool->inode->i_mapping->a_ops = &z3fold_aops;
return 0;
}
static void z3fold_unregister_migration(struct z3fold_pool *pool)
{
if (pool->inode)
iput(pool->inode);
}
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
/* Initializes the z3fold header of a newly allocated z3fold page */
static struct z3fold_header *init_z3fold_page(struct page *page, bool headless,
struct z3fold_pool *pool, gfp_t gfp)
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
{
struct z3fold_header *zhdr = page_address(page);
struct z3fold_buddy_slots *slots;
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
INIT_LIST_HEAD(&page->lru);
clear_bit(PAGE_HEADLESS, &page->private);
clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
clear_bit(NEEDS_COMPACTING, &page->private);
clear_bit(PAGE_STALE, &page->private);
clear_bit(PAGE_CLAIMED, &page->private);
if (headless)
return zhdr;
slots = alloc_slots(pool, gfp);
if (!slots)
return NULL;
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
spin_lock_init(&zhdr->page_lock);
kref_init(&zhdr->refcount);
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
zhdr->first_chunks = 0;
zhdr->middle_chunks = 0;
zhdr->last_chunks = 0;
zhdr->first_num = 0;
zhdr->start_middle = 0;
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
zhdr->cpu = -1;
zhdr->slots = slots;
zhdr->pool = pool;
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
INIT_LIST_HEAD(&zhdr->buddy);
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
INIT_WORK(&zhdr->work, compact_page_work);
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
return zhdr;
}
/* Resets the struct page fields and frees the page */
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
static void free_z3fold_page(struct page *page, bool headless)
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
{
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
if (!headless) {
lock_page(page);
__ClearPageMovable(page);
unlock_page(page);
}
ClearPagePrivate(page);
__free_page(page);
}
/* Lock a z3fold page */
static inline void z3fold_page_lock(struct z3fold_header *zhdr)
{
spin_lock(&zhdr->page_lock);
}
/* Try to lock a z3fold page */
static inline int z3fold_page_trylock(struct z3fold_header *zhdr)
{
return spin_trylock(&zhdr->page_lock);
}
/* Unlock a z3fold page */
static inline void z3fold_page_unlock(struct z3fold_header *zhdr)
{
spin_unlock(&zhdr->page_lock);
}
/* Helper function to build the index */
static inline int __idx(struct z3fold_header *zhdr, enum buddy bud)
{
return (bud + zhdr->first_num) & BUDDY_MASK;
}
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
/*
* Encodes the handle of a particular buddy within a z3fold page
* Pool lock should be held as this function accesses first_num
*/
static unsigned long __encode_handle(struct z3fold_header *zhdr,
struct z3fold_buddy_slots *slots,
enum buddy bud)
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
{
unsigned long h = (unsigned long)zhdr;
int idx = 0;
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
/*
* For a headless page, its handle is its pointer with the extra
* PAGE_HEADLESS bit set
*/
if (bud == HEADLESS)
return h | (1 << PAGE_HEADLESS);
/* otherwise, return pointer to encoded handle */
idx = __idx(zhdr, bud);
h += idx;
if (bud == LAST)
h |= (zhdr->last_chunks << BUDDY_SHIFT);
slots->slot[idx] = h;
return (unsigned long)&slots->slot[idx];
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
}
static unsigned long encode_handle(struct z3fold_header *zhdr, enum buddy bud)
{
return __encode_handle(zhdr, zhdr->slots, bud);
}
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
/* Returns the z3fold page where a given handle is stored */
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
static inline struct z3fold_header *handle_to_z3fold_header(unsigned long h)
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
{
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
unsigned long addr = h;
if (!(addr & (1 << PAGE_HEADLESS)))
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
addr = *(unsigned long *)h;
return (struct z3fold_header *)(addr & PAGE_MASK);
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
}
/* only for LAST bud, returns zero otherwise */
static unsigned short handle_to_chunks(unsigned long handle)
{
unsigned long addr = *(unsigned long *)handle;
return (addr & ~PAGE_MASK) >> BUDDY_SHIFT;
}
/*
* (handle & BUDDY_MASK) < zhdr->first_num is possible in encode_handle
* but that doesn't matter. because the masking will result in the
* correct buddy number.
*/
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
static enum buddy handle_to_buddy(unsigned long handle)
{
struct z3fold_header *zhdr;
unsigned long addr;
WARN_ON(handle & (1 << PAGE_HEADLESS));
addr = *(unsigned long *)handle;
zhdr = (struct z3fold_header *)(addr & PAGE_MASK);
return (addr - zhdr->first_num) & BUDDY_MASK;
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
}
mm/z3fold.c: introduce helper functions Patch series "z3fold: support page migration", v2. This patchset implements page migration support and slightly better buddy search. To implement page migration support, z3fold has to move away from the current scheme of handle encoding. i. e. stop encoding page address in handles. Instead, a small per-page structure is created which will contain actual addresses for z3fold objects, while pointers to fields of that structure will be used as handles. Thus, it will be possible to change the underlying addresses to reflect page migration. To support migration itself, 3 callbacks will be implemented: 1: isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise 2: migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the system. Returns 0 on success or negative error code otherwise 3: putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). To make sure an isolated page doesn't get freed, its kref is incremented in z3fold_page_isolate() and decremented during post-migration compaction, if migration was successful, or by z3fold_page_putback() in the other case. Since the new handle encoding scheme implies slight memory consumption increase, better buddy search (which decreases memory consumption) is included in this patchset. This patch (of 4): Introduce a separate helper function for object allocation, as well as 2 smaller helpers to add a buddy to the list and to get a pointer to the pool from the z3fold header. No functional changes here. Link: http://lkml.kernel.org/r/20190417103633.a4bb770b5bf0fb7e43ce1666@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:43 +08:00
static inline struct z3fold_pool *zhdr_to_pool(struct z3fold_header *zhdr)
{
return zhdr->pool;
mm/z3fold.c: introduce helper functions Patch series "z3fold: support page migration", v2. This patchset implements page migration support and slightly better buddy search. To implement page migration support, z3fold has to move away from the current scheme of handle encoding. i. e. stop encoding page address in handles. Instead, a small per-page structure is created which will contain actual addresses for z3fold objects, while pointers to fields of that structure will be used as handles. Thus, it will be possible to change the underlying addresses to reflect page migration. To support migration itself, 3 callbacks will be implemented: 1: isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise 2: migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the system. Returns 0 on success or negative error code otherwise 3: putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). To make sure an isolated page doesn't get freed, its kref is incremented in z3fold_page_isolate() and decremented during post-migration compaction, if migration was successful, or by z3fold_page_putback() in the other case. Since the new handle encoding scheme implies slight memory consumption increase, better buddy search (which decreases memory consumption) is included in this patchset. This patch (of 4): Introduce a separate helper function for object allocation, as well as 2 smaller helpers to add a buddy to the list and to get a pointer to the pool from the z3fold header. No functional changes here. Link: http://lkml.kernel.org/r/20190417103633.a4bb770b5bf0fb7e43ce1666@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:43 +08:00
}
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
static void __release_z3fold_page(struct z3fold_header *zhdr, bool locked)
{
struct page *page = virt_to_page(zhdr);
mm/z3fold.c: introduce helper functions Patch series "z3fold: support page migration", v2. This patchset implements page migration support and slightly better buddy search. To implement page migration support, z3fold has to move away from the current scheme of handle encoding. i. e. stop encoding page address in handles. Instead, a small per-page structure is created which will contain actual addresses for z3fold objects, while pointers to fields of that structure will be used as handles. Thus, it will be possible to change the underlying addresses to reflect page migration. To support migration itself, 3 callbacks will be implemented: 1: isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise 2: migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the system. Returns 0 on success or negative error code otherwise 3: putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). To make sure an isolated page doesn't get freed, its kref is incremented in z3fold_page_isolate() and decremented during post-migration compaction, if migration was successful, or by z3fold_page_putback() in the other case. Since the new handle encoding scheme implies slight memory consumption increase, better buddy search (which decreases memory consumption) is included in this patchset. This patch (of 4): Introduce a separate helper function for object allocation, as well as 2 smaller helpers to add a buddy to the list and to get a pointer to the pool from the z3fold header. No functional changes here. Link: http://lkml.kernel.org/r/20190417103633.a4bb770b5bf0fb7e43ce1666@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:43 +08:00
struct z3fold_pool *pool = zhdr_to_pool(zhdr);
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
WARN_ON(!list_empty(&zhdr->buddy));
set_bit(PAGE_STALE, &page->private);
clear_bit(NEEDS_COMPACTING, &page->private);
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
spin_lock(&pool->lock);
if (!list_empty(&page->lru))
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
list_del_init(&page->lru);
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
spin_unlock(&pool->lock);
if (locked)
z3fold_page_unlock(zhdr);
spin_lock(&pool->stale_lock);
list_add(&zhdr->buddy, &pool->stale);
queue_work(pool->release_wq, &pool->work);
spin_unlock(&pool->stale_lock);
}
static void __attribute__((__unused__))
release_z3fold_page(struct kref *ref)
{
struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
refcount);
__release_z3fold_page(zhdr, false);
}
static void release_z3fold_page_locked(struct kref *ref)
{
struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
refcount);
WARN_ON(z3fold_page_trylock(zhdr));
__release_z3fold_page(zhdr, true);
}
static void release_z3fold_page_locked_list(struct kref *ref)
{
struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
refcount);
mm/z3fold.c: introduce helper functions Patch series "z3fold: support page migration", v2. This patchset implements page migration support and slightly better buddy search. To implement page migration support, z3fold has to move away from the current scheme of handle encoding. i. e. stop encoding page address in handles. Instead, a small per-page structure is created which will contain actual addresses for z3fold objects, while pointers to fields of that structure will be used as handles. Thus, it will be possible to change the underlying addresses to reflect page migration. To support migration itself, 3 callbacks will be implemented: 1: isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise 2: migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the system. Returns 0 on success or negative error code otherwise 3: putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). To make sure an isolated page doesn't get freed, its kref is incremented in z3fold_page_isolate() and decremented during post-migration compaction, if migration was successful, or by z3fold_page_putback() in the other case. Since the new handle encoding scheme implies slight memory consumption increase, better buddy search (which decreases memory consumption) is included in this patchset. This patch (of 4): Introduce a separate helper function for object allocation, as well as 2 smaller helpers to add a buddy to the list and to get a pointer to the pool from the z3fold header. No functional changes here. Link: http://lkml.kernel.org/r/20190417103633.a4bb770b5bf0fb7e43ce1666@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:43 +08:00
struct z3fold_pool *pool = zhdr_to_pool(zhdr);
spin_lock(&pool->lock);
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
list_del_init(&zhdr->buddy);
mm/z3fold.c: introduce helper functions Patch series "z3fold: support page migration", v2. This patchset implements page migration support and slightly better buddy search. To implement page migration support, z3fold has to move away from the current scheme of handle encoding. i. e. stop encoding page address in handles. Instead, a small per-page structure is created which will contain actual addresses for z3fold objects, while pointers to fields of that structure will be used as handles. Thus, it will be possible to change the underlying addresses to reflect page migration. To support migration itself, 3 callbacks will be implemented: 1: isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise 2: migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the system. Returns 0 on success or negative error code otherwise 3: putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). To make sure an isolated page doesn't get freed, its kref is incremented in z3fold_page_isolate() and decremented during post-migration compaction, if migration was successful, or by z3fold_page_putback() in the other case. Since the new handle encoding scheme implies slight memory consumption increase, better buddy search (which decreases memory consumption) is included in this patchset. This patch (of 4): Introduce a separate helper function for object allocation, as well as 2 smaller helpers to add a buddy to the list and to get a pointer to the pool from the z3fold header. No functional changes here. Link: http://lkml.kernel.org/r/20190417103633.a4bb770b5bf0fb7e43ce1666@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:43 +08:00
spin_unlock(&pool->lock);
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
WARN_ON(z3fold_page_trylock(zhdr));
__release_z3fold_page(zhdr, true);
}
static void free_pages_work(struct work_struct *w)
{
struct z3fold_pool *pool = container_of(w, struct z3fold_pool, work);
spin_lock(&pool->stale_lock);
while (!list_empty(&pool->stale)) {
struct z3fold_header *zhdr = list_first_entry(&pool->stale,
struct z3fold_header, buddy);
struct page *page = virt_to_page(zhdr);
list_del(&zhdr->buddy);
if (WARN_ON(!test_bit(PAGE_STALE, &page->private)))
continue;
spin_unlock(&pool->stale_lock);
cancel_work_sync(&zhdr->work);
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
free_z3fold_page(page, false);
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
cond_resched();
spin_lock(&pool->stale_lock);
}
spin_unlock(&pool->stale_lock);
}
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
/*
* Returns the number of free chunks in a z3fold page.
* NB: can't be used with HEADLESS pages.
*/
static int num_free_chunks(struct z3fold_header *zhdr)
{
int nfree;
/*
* If there is a middle object, pick up the bigger free space
* either before or after it. Otherwise just subtract the number
* of chunks occupied by the first and the last objects.
*/
if (zhdr->middle_chunks != 0) {
int nfree_before = zhdr->first_chunks ?
0 : zhdr->start_middle - ZHDR_CHUNKS;
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
int nfree_after = zhdr->last_chunks ?
0 : TOTAL_CHUNKS -
(zhdr->start_middle + zhdr->middle_chunks);
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
nfree = max(nfree_before, nfree_after);
} else
nfree = NCHUNKS - zhdr->first_chunks - zhdr->last_chunks;
return nfree;
}
mm/z3fold.c: introduce helper functions Patch series "z3fold: support page migration", v2. This patchset implements page migration support and slightly better buddy search. To implement page migration support, z3fold has to move away from the current scheme of handle encoding. i. e. stop encoding page address in handles. Instead, a small per-page structure is created which will contain actual addresses for z3fold objects, while pointers to fields of that structure will be used as handles. Thus, it will be possible to change the underlying addresses to reflect page migration. To support migration itself, 3 callbacks will be implemented: 1: isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise 2: migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the system. Returns 0 on success or negative error code otherwise 3: putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). To make sure an isolated page doesn't get freed, its kref is incremented in z3fold_page_isolate() and decremented during post-migration compaction, if migration was successful, or by z3fold_page_putback() in the other case. Since the new handle encoding scheme implies slight memory consumption increase, better buddy search (which decreases memory consumption) is included in this patchset. This patch (of 4): Introduce a separate helper function for object allocation, as well as 2 smaller helpers to add a buddy to the list and to get a pointer to the pool from the z3fold header. No functional changes here. Link: http://lkml.kernel.org/r/20190417103633.a4bb770b5bf0fb7e43ce1666@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:43 +08:00
/* Add to the appropriate unbuddied list */
static inline void add_to_unbuddied(struct z3fold_pool *pool,
struct z3fold_header *zhdr)
{
if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0 ||
zhdr->middle_chunks == 0) {
struct list_head *unbuddied = get_cpu_ptr(pool->unbuddied);
int freechunks = num_free_chunks(zhdr);
spin_lock(&pool->lock);
list_add(&zhdr->buddy, &unbuddied[freechunks]);
spin_unlock(&pool->lock);
zhdr->cpu = smp_processor_id();
put_cpu_ptr(pool->unbuddied);
}
}
static inline void *mchunk_memmove(struct z3fold_header *zhdr,
unsigned short dst_chunk)
{
void *beg = zhdr;
return memmove(beg + (dst_chunk << CHUNK_SHIFT),
beg + (zhdr->start_middle << CHUNK_SHIFT),
zhdr->middle_chunks << CHUNK_SHIFT);
}
#define BIG_CHUNK_GAP 3
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
/* Has to be called with lock held */
static int z3fold_compact_page(struct z3fold_header *zhdr)
{
struct page *page = virt_to_page(zhdr);
if (test_bit(MIDDLE_CHUNK_MAPPED, &page->private))
return 0; /* can't move middle chunk, it's used */
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
if (unlikely(PageIsolated(page)))
return 0;
if (zhdr->middle_chunks == 0)
return 0; /* nothing to compact */
if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
/* move to the beginning */
mchunk_memmove(zhdr, ZHDR_CHUNKS);
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
zhdr->first_chunks = zhdr->middle_chunks;
zhdr->middle_chunks = 0;
zhdr->start_middle = 0;
zhdr->first_num++;
return 1;
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
}
/*
* moving data is expensive, so let's only do that if
* there's substantial gain (at least BIG_CHUNK_GAP chunks)
*/
if (zhdr->first_chunks != 0 && zhdr->last_chunks == 0 &&
zhdr->start_middle - (zhdr->first_chunks + ZHDR_CHUNKS) >=
BIG_CHUNK_GAP) {
mchunk_memmove(zhdr, zhdr->first_chunks + ZHDR_CHUNKS);
zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
return 1;
} else if (zhdr->last_chunks != 0 && zhdr->first_chunks == 0 &&
TOTAL_CHUNKS - (zhdr->last_chunks + zhdr->start_middle
+ zhdr->middle_chunks) >=
BIG_CHUNK_GAP) {
unsigned short new_start = TOTAL_CHUNKS - zhdr->last_chunks -
zhdr->middle_chunks;
mchunk_memmove(zhdr, new_start);
zhdr->start_middle = new_start;
return 1;
}
return 0;
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
}
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
static void do_compact_page(struct z3fold_header *zhdr, bool locked)
{
mm/z3fold.c: introduce helper functions Patch series "z3fold: support page migration", v2. This patchset implements page migration support and slightly better buddy search. To implement page migration support, z3fold has to move away from the current scheme of handle encoding. i. e. stop encoding page address in handles. Instead, a small per-page structure is created which will contain actual addresses for z3fold objects, while pointers to fields of that structure will be used as handles. Thus, it will be possible to change the underlying addresses to reflect page migration. To support migration itself, 3 callbacks will be implemented: 1: isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise 2: migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the system. Returns 0 on success or negative error code otherwise 3: putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). To make sure an isolated page doesn't get freed, its kref is incremented in z3fold_page_isolate() and decremented during post-migration compaction, if migration was successful, or by z3fold_page_putback() in the other case. Since the new handle encoding scheme implies slight memory consumption increase, better buddy search (which decreases memory consumption) is included in this patchset. This patch (of 4): Introduce a separate helper function for object allocation, as well as 2 smaller helpers to add a buddy to the list and to get a pointer to the pool from the z3fold header. No functional changes here. Link: http://lkml.kernel.org/r/20190417103633.a4bb770b5bf0fb7e43ce1666@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:43 +08:00
struct z3fold_pool *pool = zhdr_to_pool(zhdr);
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
struct page *page;
page = virt_to_page(zhdr);
if (locked)
WARN_ON(z3fold_page_trylock(zhdr));
else
z3fold_page_lock(zhdr);
if (WARN_ON(!test_and_clear_bit(NEEDS_COMPACTING, &page->private))) {
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
z3fold_page_unlock(zhdr);
return;
}
spin_lock(&pool->lock);
list_del_init(&zhdr->buddy);
spin_unlock(&pool->lock);
if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
atomic64_dec(&pool->pages_nr);
return;
}
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
if (unlikely(PageIsolated(page) ||
test_bit(PAGE_CLAIMED, &page->private) ||
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
test_bit(PAGE_STALE, &page->private))) {
z3fold_page_unlock(zhdr);
return;
}
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
z3fold_compact_page(zhdr);
mm/z3fold.c: introduce helper functions Patch series "z3fold: support page migration", v2. This patchset implements page migration support and slightly better buddy search. To implement page migration support, z3fold has to move away from the current scheme of handle encoding. i. e. stop encoding page address in handles. Instead, a small per-page structure is created which will contain actual addresses for z3fold objects, while pointers to fields of that structure will be used as handles. Thus, it will be possible to change the underlying addresses to reflect page migration. To support migration itself, 3 callbacks will be implemented: 1: isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise 2: migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the system. Returns 0 on success or negative error code otherwise 3: putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). To make sure an isolated page doesn't get freed, its kref is incremented in z3fold_page_isolate() and decremented during post-migration compaction, if migration was successful, or by z3fold_page_putback() in the other case. Since the new handle encoding scheme implies slight memory consumption increase, better buddy search (which decreases memory consumption) is included in this patchset. This patch (of 4): Introduce a separate helper function for object allocation, as well as 2 smaller helpers to add a buddy to the list and to get a pointer to the pool from the z3fold header. No functional changes here. Link: http://lkml.kernel.org/r/20190417103633.a4bb770b5bf0fb7e43ce1666@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:43 +08:00
add_to_unbuddied(pool, zhdr);
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
z3fold_page_unlock(zhdr);
}
static void compact_page_work(struct work_struct *w)
{
struct z3fold_header *zhdr = container_of(w, struct z3fold_header,
work);
do_compact_page(zhdr, false);
}
mm/z3fold.c: introduce helper functions Patch series "z3fold: support page migration", v2. This patchset implements page migration support and slightly better buddy search. To implement page migration support, z3fold has to move away from the current scheme of handle encoding. i. e. stop encoding page address in handles. Instead, a small per-page structure is created which will contain actual addresses for z3fold objects, while pointers to fields of that structure will be used as handles. Thus, it will be possible to change the underlying addresses to reflect page migration. To support migration itself, 3 callbacks will be implemented: 1: isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise 2: migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the system. Returns 0 on success or negative error code otherwise 3: putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). To make sure an isolated page doesn't get freed, its kref is incremented in z3fold_page_isolate() and decremented during post-migration compaction, if migration was successful, or by z3fold_page_putback() in the other case. Since the new handle encoding scheme implies slight memory consumption increase, better buddy search (which decreases memory consumption) is included in this patchset. This patch (of 4): Introduce a separate helper function for object allocation, as well as 2 smaller helpers to add a buddy to the list and to get a pointer to the pool from the z3fold header. No functional changes here. Link: http://lkml.kernel.org/r/20190417103633.a4bb770b5bf0fb7e43ce1666@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:43 +08:00
/* returns _locked_ z3fold page header or NULL */
static inline struct z3fold_header *__z3fold_alloc(struct z3fold_pool *pool,
size_t size, bool can_sleep)
{
struct z3fold_header *zhdr = NULL;
struct page *page;
struct list_head *unbuddied;
int chunks = size_to_chunks(size), i;
lookup:
/* First, try to find an unbuddied z3fold page. */
unbuddied = get_cpu_ptr(pool->unbuddied);
for_each_unbuddied_list(i, chunks) {
struct list_head *l = &unbuddied[i];
zhdr = list_first_entry_or_null(READ_ONCE(l),
struct z3fold_header, buddy);
if (!zhdr)
continue;
/* Re-check under lock. */
spin_lock(&pool->lock);
l = &unbuddied[i];
if (unlikely(zhdr != list_first_entry(READ_ONCE(l),
struct z3fold_header, buddy)) ||
!z3fold_page_trylock(zhdr)) {
spin_unlock(&pool->lock);
zhdr = NULL;
put_cpu_ptr(pool->unbuddied);
if (can_sleep)
cond_resched();
goto lookup;
}
list_del_init(&zhdr->buddy);
zhdr->cpu = -1;
spin_unlock(&pool->lock);
page = virt_to_page(zhdr);
if (test_bit(NEEDS_COMPACTING, &page->private)) {
z3fold_page_unlock(zhdr);
zhdr = NULL;
put_cpu_ptr(pool->unbuddied);
if (can_sleep)
cond_resched();
goto lookup;
}
/*
* this page could not be removed from its unbuddied
* list while pool lock was held, and then we've taken
* page lock so kref_put could not be called before
* we got here, so it's safe to just call kref_get()
*/
kref_get(&zhdr->refcount);
break;
}
put_cpu_ptr(pool->unbuddied);
if (!zhdr) {
int cpu;
/* look for _exact_ match on other cpus' lists */
for_each_online_cpu(cpu) {
struct list_head *l;
unbuddied = per_cpu_ptr(pool->unbuddied, cpu);
spin_lock(&pool->lock);
l = &unbuddied[chunks];
zhdr = list_first_entry_or_null(READ_ONCE(l),
struct z3fold_header, buddy);
if (!zhdr || !z3fold_page_trylock(zhdr)) {
spin_unlock(&pool->lock);
zhdr = NULL;
continue;
}
list_del_init(&zhdr->buddy);
zhdr->cpu = -1;
spin_unlock(&pool->lock);
page = virt_to_page(zhdr);
if (test_bit(NEEDS_COMPACTING, &page->private)) {
z3fold_page_unlock(zhdr);
zhdr = NULL;
if (can_sleep)
cond_resched();
continue;
}
kref_get(&zhdr->refcount);
break;
}
}
mm/z3fold.c: introduce helper functions Patch series "z3fold: support page migration", v2. This patchset implements page migration support and slightly better buddy search. To implement page migration support, z3fold has to move away from the current scheme of handle encoding. i. e. stop encoding page address in handles. Instead, a small per-page structure is created which will contain actual addresses for z3fold objects, while pointers to fields of that structure will be used as handles. Thus, it will be possible to change the underlying addresses to reflect page migration. To support migration itself, 3 callbacks will be implemented: 1: isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise 2: migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the system. Returns 0 on success or negative error code otherwise 3: putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). To make sure an isolated page doesn't get freed, its kref is incremented in z3fold_page_isolate() and decremented during post-migration compaction, if migration was successful, or by z3fold_page_putback() in the other case. Since the new handle encoding scheme implies slight memory consumption increase, better buddy search (which decreases memory consumption) is included in this patchset. This patch (of 4): Introduce a separate helper function for object allocation, as well as 2 smaller helpers to add a buddy to the list and to get a pointer to the pool from the z3fold header. No functional changes here. Link: http://lkml.kernel.org/r/20190417103633.a4bb770b5bf0fb7e43ce1666@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:43 +08:00
return zhdr;
}
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
/*
* API Functions
*/
/**
* z3fold_create_pool() - create a new z3fold pool
* @name: pool name
* @gfp: gfp flags when allocating the z3fold pool structure
* @ops: user-defined operations for the z3fold pool
*
* Return: pointer to the new z3fold pool or NULL if the metadata allocation
* failed.
*/
static struct z3fold_pool *z3fold_create_pool(const char *name, gfp_t gfp,
const struct z3fold_ops *ops)
{
struct z3fold_pool *pool = NULL;
int i, cpu;
pool = kzalloc(sizeof(struct z3fold_pool), gfp);
if (!pool)
goto out;
pool->c_handle = kmem_cache_create("z3fold_handle",
sizeof(struct z3fold_buddy_slots),
SLOTS_ALIGN, 0, NULL);
if (!pool->c_handle)
goto out_c;
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
spin_lock_init(&pool->lock);
spin_lock_init(&pool->stale_lock);
pool->unbuddied = __alloc_percpu(sizeof(struct list_head)*NCHUNKS, 2);
if (!pool->unbuddied)
goto out_pool;
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
for_each_possible_cpu(cpu) {
struct list_head *unbuddied =
per_cpu_ptr(pool->unbuddied, cpu);
for_each_unbuddied_list(i, 0)
INIT_LIST_HEAD(&unbuddied[i]);
}
INIT_LIST_HEAD(&pool->lru);
INIT_LIST_HEAD(&pool->stale);
atomic64_set(&pool->pages_nr, 0);
pool->name = name;
pool->compact_wq = create_singlethread_workqueue(pool->name);
if (!pool->compact_wq)
goto out_unbuddied;
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
pool->release_wq = create_singlethread_workqueue(pool->name);
if (!pool->release_wq)
goto out_wq;
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
if (z3fold_register_migration(pool))
goto out_rwq;
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
INIT_WORK(&pool->work, free_pages_work);
pool->ops = ops;
return pool;
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
out_rwq:
destroy_workqueue(pool->release_wq);
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
out_wq:
destroy_workqueue(pool->compact_wq);
out_unbuddied:
free_percpu(pool->unbuddied);
out_pool:
kmem_cache_destroy(pool->c_handle);
out_c:
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
kfree(pool);
out:
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
return NULL;
}
/**
* z3fold_destroy_pool() - destroys an existing z3fold pool
* @pool: the z3fold pool to be destroyed
*
* The pool should be emptied before this function is called.
*/
static void z3fold_destroy_pool(struct z3fold_pool *pool)
{
kmem_cache_destroy(pool->c_handle);
/*
* We need to destroy pool->compact_wq before pool->release_wq,
* as any pending work on pool->compact_wq will call
* queue_work(pool->release_wq, &pool->work).
*
* There are still outstanding pages until both workqueues are drained,
* so we cannot unregister migration until then.
*/
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
destroy_workqueue(pool->compact_wq);
destroy_workqueue(pool->release_wq);
z3fold_unregister_migration(pool);
free_percpu(pool->unbuddied);
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
kfree(pool);
}
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
/**
* z3fold_alloc() - allocates a region of a given size
* @pool: z3fold pool from which to allocate
* @size: size in bytes of the desired allocation
* @gfp: gfp flags used if the pool needs to grow
* @handle: handle of the new allocation
*
* This function will attempt to find a free region in the pool large enough to
* satisfy the allocation request. A search of the unbuddied lists is
* performed first. If no suitable free region is found, then a new page is
* allocated and added to the pool to satisfy the request.
*
* gfp should not set __GFP_HIGHMEM as highmem pages cannot be used
* as z3fold pool pages.
*
* Return: 0 if success and handle is set, otherwise -EINVAL if the size or
* gfp arguments are invalid or -ENOMEM if the pool was unable to allocate
* a new page.
*/
static int z3fold_alloc(struct z3fold_pool *pool, size_t size, gfp_t gfp,
unsigned long *handle)
{
mm/z3fold.c: introduce helper functions Patch series "z3fold: support page migration", v2. This patchset implements page migration support and slightly better buddy search. To implement page migration support, z3fold has to move away from the current scheme of handle encoding. i. e. stop encoding page address in handles. Instead, a small per-page structure is created which will contain actual addresses for z3fold objects, while pointers to fields of that structure will be used as handles. Thus, it will be possible to change the underlying addresses to reflect page migration. To support migration itself, 3 callbacks will be implemented: 1: isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise 2: migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the system. Returns 0 on success or negative error code otherwise 3: putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). To make sure an isolated page doesn't get freed, its kref is incremented in z3fold_page_isolate() and decremented during post-migration compaction, if migration was successful, or by z3fold_page_putback() in the other case. Since the new handle encoding scheme implies slight memory consumption increase, better buddy search (which decreases memory consumption) is included in this patchset. This patch (of 4): Introduce a separate helper function for object allocation, as well as 2 smaller helpers to add a buddy to the list and to get a pointer to the pool from the z3fold header. No functional changes here. Link: http://lkml.kernel.org/r/20190417103633.a4bb770b5bf0fb7e43ce1666@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:43 +08:00
int chunks = size_to_chunks(size);
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
struct z3fold_header *zhdr = NULL;
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
struct page *page = NULL;
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
enum buddy bud;
bool can_sleep = gfpflags_allow_blocking(gfp);
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
if (!size)
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
return -EINVAL;
if (size > PAGE_SIZE)
return -ENOSPC;
if (size > PAGE_SIZE - ZHDR_SIZE_ALIGNED - CHUNK_SIZE)
bud = HEADLESS;
else {
mm/z3fold.c: introduce helper functions Patch series "z3fold: support page migration", v2. This patchset implements page migration support and slightly better buddy search. To implement page migration support, z3fold has to move away from the current scheme of handle encoding. i. e. stop encoding page address in handles. Instead, a small per-page structure is created which will contain actual addresses for z3fold objects, while pointers to fields of that structure will be used as handles. Thus, it will be possible to change the underlying addresses to reflect page migration. To support migration itself, 3 callbacks will be implemented: 1: isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise 2: migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the system. Returns 0 on success or negative error code otherwise 3: putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). To make sure an isolated page doesn't get freed, its kref is incremented in z3fold_page_isolate() and decremented during post-migration compaction, if migration was successful, or by z3fold_page_putback() in the other case. Since the new handle encoding scheme implies slight memory consumption increase, better buddy search (which decreases memory consumption) is included in this patchset. This patch (of 4): Introduce a separate helper function for object allocation, as well as 2 smaller helpers to add a buddy to the list and to get a pointer to the pool from the z3fold header. No functional changes here. Link: http://lkml.kernel.org/r/20190417103633.a4bb770b5bf0fb7e43ce1666@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:43 +08:00
retry:
zhdr = __z3fold_alloc(pool, size, can_sleep);
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
if (zhdr) {
if (zhdr->first_chunks == 0) {
if (zhdr->middle_chunks != 0 &&
chunks >= zhdr->start_middle)
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
bud = LAST;
else
bud = FIRST;
} else if (zhdr->last_chunks == 0)
bud = LAST;
else if (zhdr->middle_chunks == 0)
bud = MIDDLE;
else {
if (kref_put(&zhdr->refcount,
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
release_z3fold_page_locked))
atomic64_dec(&pool->pages_nr);
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
else
z3fold_page_unlock(zhdr);
pr_err("No free chunks in unbuddied\n");
WARN_ON(1);
mm/z3fold.c: introduce helper functions Patch series "z3fold: support page migration", v2. This patchset implements page migration support and slightly better buddy search. To implement page migration support, z3fold has to move away from the current scheme of handle encoding. i. e. stop encoding page address in handles. Instead, a small per-page structure is created which will contain actual addresses for z3fold objects, while pointers to fields of that structure will be used as handles. Thus, it will be possible to change the underlying addresses to reflect page migration. To support migration itself, 3 callbacks will be implemented: 1: isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise 2: migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the system. Returns 0 on success or negative error code otherwise 3: putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). To make sure an isolated page doesn't get freed, its kref is incremented in z3fold_page_isolate() and decremented during post-migration compaction, if migration was successful, or by z3fold_page_putback() in the other case. Since the new handle encoding scheme implies slight memory consumption increase, better buddy search (which decreases memory consumption) is included in this patchset. This patch (of 4): Introduce a separate helper function for object allocation, as well as 2 smaller helpers to add a buddy to the list and to get a pointer to the pool from the z3fold header. No functional changes here. Link: http://lkml.kernel.org/r/20190417103633.a4bb770b5bf0fb7e43ce1666@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:43 +08:00
goto retry;
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
}
mm/z3fold.c: introduce helper functions Patch series "z3fold: support page migration", v2. This patchset implements page migration support and slightly better buddy search. To implement page migration support, z3fold has to move away from the current scheme of handle encoding. i. e. stop encoding page address in handles. Instead, a small per-page structure is created which will contain actual addresses for z3fold objects, while pointers to fields of that structure will be used as handles. Thus, it will be possible to change the underlying addresses to reflect page migration. To support migration itself, 3 callbacks will be implemented: 1: isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise 2: migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the system. Returns 0 on success or negative error code otherwise 3: putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). To make sure an isolated page doesn't get freed, its kref is incremented in z3fold_page_isolate() and decremented during post-migration compaction, if migration was successful, or by z3fold_page_putback() in the other case. Since the new handle encoding scheme implies slight memory consumption increase, better buddy search (which decreases memory consumption) is included in this patchset. This patch (of 4): Introduce a separate helper function for object allocation, as well as 2 smaller helpers to add a buddy to the list and to get a pointer to the pool from the z3fold header. No functional changes here. Link: http://lkml.kernel.org/r/20190417103633.a4bb770b5bf0fb7e43ce1666@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:43 +08:00
page = virt_to_page(zhdr);
goto found;
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
}
bud = FIRST;
}
page = NULL;
if (can_sleep) {
spin_lock(&pool->stale_lock);
zhdr = list_first_entry_or_null(&pool->stale,
struct z3fold_header, buddy);
/*
* Before allocating a page, let's see if we can take one from
* the stale pages list. cancel_work_sync() can sleep so we
* limit this case to the contexts where we can sleep
*/
if (zhdr) {
list_del(&zhdr->buddy);
spin_unlock(&pool->stale_lock);
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
cancel_work_sync(&zhdr->work);
page = virt_to_page(zhdr);
} else {
spin_unlock(&pool->stale_lock);
}
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
}
if (!page)
page = alloc_page(gfp);
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
if (!page)
return -ENOMEM;
zhdr = init_z3fold_page(page, bud == HEADLESS, pool, gfp);
mm/z3fold.c: introduce helper functions Patch series "z3fold: support page migration", v2. This patchset implements page migration support and slightly better buddy search. To implement page migration support, z3fold has to move away from the current scheme of handle encoding. i. e. stop encoding page address in handles. Instead, a small per-page structure is created which will contain actual addresses for z3fold objects, while pointers to fields of that structure will be used as handles. Thus, it will be possible to change the underlying addresses to reflect page migration. To support migration itself, 3 callbacks will be implemented: 1: isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise 2: migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the system. Returns 0 on success or negative error code otherwise 3: putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). To make sure an isolated page doesn't get freed, its kref is incremented in z3fold_page_isolate() and decremented during post-migration compaction, if migration was successful, or by z3fold_page_putback() in the other case. Since the new handle encoding scheme implies slight memory consumption increase, better buddy search (which decreases memory consumption) is included in this patchset. This patch (of 4): Introduce a separate helper function for object allocation, as well as 2 smaller helpers to add a buddy to the list and to get a pointer to the pool from the z3fold header. No functional changes here. Link: http://lkml.kernel.org/r/20190417103633.a4bb770b5bf0fb7e43ce1666@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:43 +08:00
if (!zhdr) {
__free_page(page);
return -ENOMEM;
}
atomic64_inc(&pool->pages_nr);
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
if (bud == HEADLESS) {
set_bit(PAGE_HEADLESS, &page->private);
goto headless;
}
if (can_sleep) {
lock_page(page);
__SetPageMovable(page, pool->inode->i_mapping);
unlock_page(page);
} else {
if (trylock_page(page)) {
__SetPageMovable(page, pool->inode->i_mapping);
unlock_page(page);
}
}
z3fold_page_lock(zhdr);
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
found:
if (bud == FIRST)
zhdr->first_chunks = chunks;
else if (bud == LAST)
zhdr->last_chunks = chunks;
else {
zhdr->middle_chunks = chunks;
zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
}
mm/z3fold.c: introduce helper functions Patch series "z3fold: support page migration", v2. This patchset implements page migration support and slightly better buddy search. To implement page migration support, z3fold has to move away from the current scheme of handle encoding. i. e. stop encoding page address in handles. Instead, a small per-page structure is created which will contain actual addresses for z3fold objects, while pointers to fields of that structure will be used as handles. Thus, it will be possible to change the underlying addresses to reflect page migration. To support migration itself, 3 callbacks will be implemented: 1: isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise 2: migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the system. Returns 0 on success or negative error code otherwise 3: putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). To make sure an isolated page doesn't get freed, its kref is incremented in z3fold_page_isolate() and decremented during post-migration compaction, if migration was successful, or by z3fold_page_putback() in the other case. Since the new handle encoding scheme implies slight memory consumption increase, better buddy search (which decreases memory consumption) is included in this patchset. This patch (of 4): Introduce a separate helper function for object allocation, as well as 2 smaller helpers to add a buddy to the list and to get a pointer to the pool from the z3fold header. No functional changes here. Link: http://lkml.kernel.org/r/20190417103633.a4bb770b5bf0fb7e43ce1666@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:43 +08:00
add_to_unbuddied(pool, zhdr);
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
headless:
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
spin_lock(&pool->lock);
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
/* Add/move z3fold page to beginning of LRU */
if (!list_empty(&page->lru))
list_del(&page->lru);
list_add(&page->lru, &pool->lru);
*handle = encode_handle(zhdr, bud);
spin_unlock(&pool->lock);
if (bud != HEADLESS)
z3fold_page_unlock(zhdr);
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
return 0;
}
/**
* z3fold_free() - frees the allocation associated with the given handle
* @pool: pool in which the allocation resided
* @handle: handle associated with the allocation returned by z3fold_alloc()
*
* In the case that the z3fold page in which the allocation resides is under
* reclaim, as indicated by the PG_reclaim flag being set, this function
* only sets the first|last_chunks to 0. The page is actually freed
* once both buddies are evicted (see z3fold_reclaim_page() below).
*/
static void z3fold_free(struct z3fold_pool *pool, unsigned long handle)
{
struct z3fold_header *zhdr;
struct page *page;
enum buddy bud;
bool page_claimed;
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
zhdr = handle_to_z3fold_header(handle);
page = virt_to_page(zhdr);
page_claimed = test_and_set_bit(PAGE_CLAIMED, &page->private);
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
if (test_bit(PAGE_HEADLESS, &page->private)) {
/* if a headless page is under reclaim, just leave.
* NB: we use test_and_set_bit for a reason: if the bit
* has not been set before, we release this page
* immediately so we don't care about its value any more.
*/
if (!page_claimed) {
spin_lock(&pool->lock);
list_del(&page->lru);
spin_unlock(&pool->lock);
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
free_z3fold_page(page, true);
atomic64_dec(&pool->pages_nr);
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
}
return;
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
}
/* Non-headless case */
z3fold_page_lock(zhdr);
bud = handle_to_buddy(handle);
switch (bud) {
case FIRST:
zhdr->first_chunks = 0;
break;
case MIDDLE:
zhdr->middle_chunks = 0;
break;
case LAST:
zhdr->last_chunks = 0;
break;
default:
pr_err("%s: unknown bud %d\n", __func__, bud);
WARN_ON(1);
z3fold_page_unlock(zhdr);
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
return;
}
free_handle(handle);
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
if (kref_put(&zhdr->refcount, release_z3fold_page_locked_list)) {
atomic64_dec(&pool->pages_nr);
return;
}
if (page_claimed) {
/* the page has not been claimed by us */
z3fold_page_unlock(zhdr);
return;
}
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
if (unlikely(PageIsolated(page)) ||
test_and_set_bit(NEEDS_COMPACTING, &page->private)) {
z3fold_page_unlock(zhdr);
clear_bit(PAGE_CLAIMED, &page->private);
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
return;
}
if (zhdr->cpu < 0 || !cpu_online(zhdr->cpu)) {
spin_lock(&pool->lock);
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
list_del_init(&zhdr->buddy);
spin_unlock(&pool->lock);
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
zhdr->cpu = -1;
kref_get(&zhdr->refcount);
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
do_compact_page(zhdr, true);
clear_bit(PAGE_CLAIMED, &page->private);
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
return;
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
}
kref_get(&zhdr->refcount);
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
queue_work_on(zhdr->cpu, pool->compact_wq, &zhdr->work);
clear_bit(PAGE_CLAIMED, &page->private);
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
z3fold_page_unlock(zhdr);
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
}
/**
* z3fold_reclaim_page() - evicts allocations from a pool page and frees it
* @pool: pool from which a page will attempt to be evicted
* @retries: number of pages on the LRU list for which eviction will
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
* be attempted before failing
*
* z3fold reclaim is different from normal system reclaim in that it is done
* from the bottom, up. This is because only the bottom layer, z3fold, has
* information on how the allocations are organized within each z3fold page.
* This has the potential to create interesting locking situations between
* z3fold and the user, however.
*
* To avoid these, this is how z3fold_reclaim_page() should be called:
*
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
* The user detects a page should be reclaimed and calls z3fold_reclaim_page().
* z3fold_reclaim_page() will remove a z3fold page from the pool LRU list and
* call the user-defined eviction handler with the pool and handle as
* arguments.
*
* If the handle can not be evicted, the eviction handler should return
* non-zero. z3fold_reclaim_page() will add the z3fold page back to the
* appropriate list and try the next z3fold page on the LRU up to
* a user defined number of retries.
*
* If the handle is successfully evicted, the eviction handler should
* return 0 _and_ should have called z3fold_free() on the handle. z3fold_free()
* contains logic to delay freeing the page if the page is under reclaim,
* as indicated by the setting of the PG_reclaim flag on the underlying page.
*
* If all buddies in the z3fold page are successfully evicted, then the
* z3fold page can be freed.
*
* Returns: 0 if page is successfully freed, otherwise -EINVAL if there are
* no pages to evict or an eviction handler is not registered, -EAGAIN if
* the retry limit was hit.
*/
static int z3fold_reclaim_page(struct z3fold_pool *pool, unsigned int retries)
{
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
int i, ret = 0;
struct z3fold_header *zhdr = NULL;
struct page *page = NULL;
struct list_head *pos;
struct z3fold_buddy_slots slots;
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
unsigned long first_handle = 0, middle_handle = 0, last_handle = 0;
spin_lock(&pool->lock);
if (!pool->ops || !pool->ops->evict || retries == 0) {
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
spin_unlock(&pool->lock);
return -EINVAL;
}
for (i = 0; i < retries; i++) {
if (list_empty(&pool->lru)) {
spin_unlock(&pool->lock);
return -EINVAL;
}
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
list_for_each_prev(pos, &pool->lru) {
page = list_entry(pos, struct page, lru);
/* this bit could have been set by free, in which case
* we pass over to the next page in the pool.
*/
if (test_and_set_bit(PAGE_CLAIMED, &page->private)) {
page = NULL;
continue;
}
if (unlikely(PageIsolated(page))) {
clear_bit(PAGE_CLAIMED, &page->private);
page = NULL;
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
continue;
}
zhdr = page_address(page);
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
if (test_bit(PAGE_HEADLESS, &page->private))
break;
if (!z3fold_page_trylock(zhdr)) {
clear_bit(PAGE_CLAIMED, &page->private);
zhdr = NULL;
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
continue; /* can't evict at this point */
}
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
kref_get(&zhdr->refcount);
list_del_init(&zhdr->buddy);
zhdr->cpu = -1;
break;
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
}
if (!zhdr)
break;
list_del_init(&page->lru);
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
spin_unlock(&pool->lock);
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
if (!test_bit(PAGE_HEADLESS, &page->private)) {
/*
* We need encode the handles before unlocking, and
* use our local slots structure because z3fold_free
* can zero out zhdr->slots and we can't do much
* about that
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
*/
first_handle = 0;
last_handle = 0;
middle_handle = 0;
if (zhdr->first_chunks)
first_handle = __encode_handle(zhdr, &slots,
FIRST);
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
if (zhdr->middle_chunks)
middle_handle = __encode_handle(zhdr, &slots,
MIDDLE);
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
if (zhdr->last_chunks)
last_handle = __encode_handle(zhdr, &slots,
LAST);
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
/*
* it's safe to unlock here because we hold a
* reference to this page
*/
z3fold_page_unlock(zhdr);
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
} else {
first_handle = __encode_handle(zhdr, &slots, HEADLESS);
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
last_handle = middle_handle = 0;
}
/* Issue the eviction callback(s) */
if (middle_handle) {
ret = pool->ops->evict(pool, middle_handle);
if (ret)
goto next;
}
if (first_handle) {
ret = pool->ops->evict(pool, first_handle);
if (ret)
goto next;
}
if (last_handle) {
ret = pool->ops->evict(pool, last_handle);
if (ret)
goto next;
}
next:
if (test_bit(PAGE_HEADLESS, &page->private)) {
if (ret == 0) {
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
free_z3fold_page(page, true);
atomic64_dec(&pool->pages_nr);
return 0;
}
spin_lock(&pool->lock);
list_add(&page->lru, &pool->lru);
spin_unlock(&pool->lock);
clear_bit(PAGE_CLAIMED, &page->private);
} else {
z3fold_page_lock(zhdr);
if (kref_put(&zhdr->refcount,
release_z3fold_page_locked)) {
atomic64_dec(&pool->pages_nr);
return 0;
}
/*
* if we are here, the page is still not completely
* free. Take the global pool lock then to be able
* to add it back to the lru list
*/
spin_lock(&pool->lock);
list_add(&page->lru, &pool->lru);
spin_unlock(&pool->lock);
z3fold_page_unlock(zhdr);
clear_bit(PAGE_CLAIMED, &page->private);
}
/* We started off locked to we need to lock the pool back */
spin_lock(&pool->lock);
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
}
spin_unlock(&pool->lock);
return -EAGAIN;
}
/**
* z3fold_map() - maps the allocation associated with the given handle
* @pool: pool in which the allocation resides
* @handle: handle associated with the allocation to be mapped
*
* Extracts the buddy number from handle and constructs the pointer to the
* correct starting chunk within the page.
*
* Returns: a pointer to the mapped allocation
*/
static void *z3fold_map(struct z3fold_pool *pool, unsigned long handle)
{
struct z3fold_header *zhdr;
struct page *page;
void *addr;
enum buddy buddy;
zhdr = handle_to_z3fold_header(handle);
addr = zhdr;
page = virt_to_page(zhdr);
if (test_bit(PAGE_HEADLESS, &page->private))
goto out;
z3fold_page_lock(zhdr);
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
buddy = handle_to_buddy(handle);
switch (buddy) {
case FIRST:
addr += ZHDR_SIZE_ALIGNED;
break;
case MIDDLE:
addr += zhdr->start_middle << CHUNK_SHIFT;
set_bit(MIDDLE_CHUNK_MAPPED, &page->private);
break;
case LAST:
addr += PAGE_SIZE - (handle_to_chunks(handle) << CHUNK_SHIFT);
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
break;
default:
pr_err("unknown buddy id %d\n", buddy);
WARN_ON(1);
addr = NULL;
break;
}
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
if (addr)
zhdr->mapped_count++;
z3fold_page_unlock(zhdr);
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
out:
return addr;
}
/**
* z3fold_unmap() - unmaps the allocation associated with the given handle
* @pool: pool in which the allocation resides
* @handle: handle associated with the allocation to be unmapped
*/
static void z3fold_unmap(struct z3fold_pool *pool, unsigned long handle)
{
struct z3fold_header *zhdr;
struct page *page;
enum buddy buddy;
zhdr = handle_to_z3fold_header(handle);
page = virt_to_page(zhdr);
if (test_bit(PAGE_HEADLESS, &page->private))
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
return;
z3fold_page_lock(zhdr);
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
buddy = handle_to_buddy(handle);
if (buddy == MIDDLE)
clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
zhdr->mapped_count--;
z3fold_page_unlock(zhdr);
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
}
/**
* z3fold_get_pool_size() - gets the z3fold pool size in pages
* @pool: pool whose size is being queried
*
* Returns: size in pages of the given pool.
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
*/
static u64 z3fold_get_pool_size(struct z3fold_pool *pool)
{
return atomic64_read(&pool->pages_nr);
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
}
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
static bool z3fold_page_isolate(struct page *page, isolate_mode_t mode)
{
struct z3fold_header *zhdr;
struct z3fold_pool *pool;
VM_BUG_ON_PAGE(!PageMovable(page), page);
VM_BUG_ON_PAGE(PageIsolated(page), page);
if (test_bit(PAGE_HEADLESS, &page->private) ||
test_bit(PAGE_CLAIMED, &page->private))
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
return false;
zhdr = page_address(page);
z3fold_page_lock(zhdr);
if (test_bit(NEEDS_COMPACTING, &page->private) ||
test_bit(PAGE_STALE, &page->private))
goto out;
pool = zhdr_to_pool(zhdr);
if (zhdr->mapped_count == 0) {
kref_get(&zhdr->refcount);
if (!list_empty(&zhdr->buddy))
list_del_init(&zhdr->buddy);
spin_lock(&pool->lock);
if (!list_empty(&page->lru))
list_del(&page->lru);
spin_unlock(&pool->lock);
z3fold_page_unlock(zhdr);
return true;
}
out:
z3fold_page_unlock(zhdr);
return false;
}
static int z3fold_page_migrate(struct address_space *mapping, struct page *newpage,
struct page *page, enum migrate_mode mode)
{
struct z3fold_header *zhdr, *new_zhdr;
struct z3fold_pool *pool;
struct address_space *new_mapping;
VM_BUG_ON_PAGE(!PageMovable(page), page);
VM_BUG_ON_PAGE(!PageIsolated(page), page);
VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
zhdr = page_address(page);
pool = zhdr_to_pool(zhdr);
if (!z3fold_page_trylock(zhdr)) {
return -EAGAIN;
}
if (zhdr->mapped_count != 0) {
z3fold_page_unlock(zhdr);
return -EBUSY;
}
if (work_pending(&zhdr->work)) {
z3fold_page_unlock(zhdr);
return -EAGAIN;
}
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
new_zhdr = page_address(newpage);
memcpy(new_zhdr, zhdr, PAGE_SIZE);
newpage->private = page->private;
page->private = 0;
z3fold_page_unlock(zhdr);
spin_lock_init(&new_zhdr->page_lock);
INIT_WORK(&new_zhdr->work, compact_page_work);
/*
* z3fold_page_isolate() ensures that new_zhdr->buddy is empty,
* so we only have to reinitialize it.
*/
INIT_LIST_HEAD(&new_zhdr->buddy);
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
new_mapping = page_mapping(page);
__ClearPageMovable(page);
ClearPagePrivate(page);
get_page(newpage);
z3fold_page_lock(new_zhdr);
if (new_zhdr->first_chunks)
encode_handle(new_zhdr, FIRST);
if (new_zhdr->last_chunks)
encode_handle(new_zhdr, LAST);
if (new_zhdr->middle_chunks)
encode_handle(new_zhdr, MIDDLE);
set_bit(NEEDS_COMPACTING, &newpage->private);
new_zhdr->cpu = smp_processor_id();
spin_lock(&pool->lock);
list_add(&newpage->lru, &pool->lru);
spin_unlock(&pool->lock);
__SetPageMovable(newpage, new_mapping);
z3fold_page_unlock(new_zhdr);
queue_work_on(new_zhdr->cpu, pool->compact_wq, &new_zhdr->work);
page_mapcount_reset(page);
put_page(page);
return 0;
}
static void z3fold_page_putback(struct page *page)
{
struct z3fold_header *zhdr;
struct z3fold_pool *pool;
zhdr = page_address(page);
pool = zhdr_to_pool(zhdr);
z3fold_page_lock(zhdr);
if (!list_empty(&zhdr->buddy))
list_del_init(&zhdr->buddy);
INIT_LIST_HEAD(&page->lru);
if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
atomic64_dec(&pool->pages_nr);
return;
}
spin_lock(&pool->lock);
list_add(&page->lru, &pool->lru);
spin_unlock(&pool->lock);
z3fold_page_unlock(zhdr);
}
static const struct address_space_operations z3fold_aops = {
.isolate_page = z3fold_page_isolate,
.migratepage = z3fold_page_migrate,
.putback_page = z3fold_page_putback,
};
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
/*****************
* zpool
****************/
static int z3fold_zpool_evict(struct z3fold_pool *pool, unsigned long handle)
{
if (pool->zpool && pool->zpool_ops && pool->zpool_ops->evict)
return pool->zpool_ops->evict(pool->zpool, handle);
else
return -ENOENT;
}
static const struct z3fold_ops z3fold_zpool_ops = {
.evict = z3fold_zpool_evict
};
static void *z3fold_zpool_create(const char *name, gfp_t gfp,
const struct zpool_ops *zpool_ops,
struct zpool *zpool)
{
struct z3fold_pool *pool;
z3fold: use per-cpu unbuddied lists It's been noted that z3fold doesn't scale well when it's run in a large number of threads on many cores, which can be easily reproduced with fio 'randrw' test with --numjobs=32. E.g. the result for 1 cluster (4 cores) is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=496883KB/s, minb=15527KB/s, ... WRITE: io=246735MB, aggrb=500841KB/s, minb=15651KB/s, ... While for 8 cores (2 clusters) the result is: Run status group 0 (all jobs): READ: io=244785MB, aggrb=265942KB/s, minb=8310KB/s, ... WRITE: io=246735MB, aggrb=268060KB/s, minb=8376KB/s, ... The bottleneck here is the pool lock which many threads become waiting upon. To reduce that spin lock contention, z3fold can operate only on the lists local to the current CPU whenever possible. Due to the nature of z3fold unbuddied list handling (it only takes the first entry off the list on a hot path), if the z3fold pool is big enough and balanced well enough, limiting search to only local unbuddied list doesn't lead to a significant compression ratio degrade (2.57x vs 2.65x in our measurements). This patch also introduces two worker threads: one for async in-page object layout optimization and one for releasing freed pages. This is done to speed up z3fold_free() which is often on a hot path. The fio results for 8-core case are now the following: Run status group 0 (all jobs): READ: io=244785MB, aggrb=1568.3MB/s, minb=50182KB/s, ... WRITE: io=246735MB, aggrb=1580.8MB/s, minb=50582KB/s, ... So we're in for almost 6x performance increase. Link: http://lkml.kernel.org/r/20170806181443.f9b65018f8bde25ef990f9e8@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Dan Streetman <ddstreet@ieee.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-07 07:24:47 +08:00
pool = z3fold_create_pool(name, gfp,
zpool_ops ? &z3fold_zpool_ops : NULL);
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
if (pool) {
pool->zpool = zpool;
pool->zpool_ops = zpool_ops;
}
return pool;
}
static void z3fold_zpool_destroy(void *pool)
{
z3fold_destroy_pool(pool);
}
static int z3fold_zpool_malloc(void *pool, size_t size, gfp_t gfp,
unsigned long *handle)
{
return z3fold_alloc(pool, size, gfp, handle);
}
static void z3fold_zpool_free(void *pool, unsigned long handle)
{
z3fold_free(pool, handle);
}
static int z3fold_zpool_shrink(void *pool, unsigned int pages,
unsigned int *reclaimed)
{
unsigned int total = 0;
int ret = -EINVAL;
while (total < pages) {
ret = z3fold_reclaim_page(pool, 8);
if (ret < 0)
break;
total++;
}
if (reclaimed)
*reclaimed = total;
return ret;
}
static void *z3fold_zpool_map(void *pool, unsigned long handle,
enum zpool_mapmode mm)
{
return z3fold_map(pool, handle);
}
static void z3fold_zpool_unmap(void *pool, unsigned long handle)
{
z3fold_unmap(pool, handle);
}
static u64 z3fold_zpool_total_size(void *pool)
{
return z3fold_get_pool_size(pool) * PAGE_SIZE;
}
static struct zpool_driver z3fold_zpool_driver = {
.type = "z3fold",
.owner = THIS_MODULE,
.create = z3fold_zpool_create,
.destroy = z3fold_zpool_destroy,
.malloc = z3fold_zpool_malloc,
.free = z3fold_zpool_free,
.shrink = z3fold_zpool_shrink,
.map = z3fold_zpool_map,
.unmap = z3fold_zpool_unmap,
.total_size = z3fold_zpool_total_size,
};
MODULE_ALIAS("zpool-z3fold");
static int __init init_z3fold(void)
{
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
int ret;
/* Make sure the z3fold header is not larger than the page size */
BUILD_BUG_ON(ZHDR_SIZE_ALIGNED > PAGE_SIZE);
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
ret = z3fold_mount();
if (ret)
return ret;
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
zpool_register_driver(&z3fold_zpool_driver);
return 0;
}
static void __exit exit_z3fold(void)
{
mm/z3fold.c: support page migration Now that we are not using page address in handles directly, we can make z3fold pages movable to decrease the memory fragmentation z3fold may create over time. This patch starts advertising non-headless z3fold pages as movable and uses the existing kernel infrastructure to implement moving of such pages per memory management subsystem's request. It thus implements 3 required callbacks for page migration: * isolation callback: z3fold_page_isolate(): try to isolate the page by removing it from all lists. Pages scheduled for some activity and mapped pages will not be isolated. Return true if isolation was successful or false otherwise * migration callback: z3fold_page_migrate(): re-check critical conditions and migrate page contents to the new page provided by the memory subsystem. Returns 0 on success or negative error code otherwise * putback callback: z3fold_page_putback(): put back the page if z3fold_page_migrate() for it failed permanently (i. e. not with -EAGAIN code). [lkp@intel.com: z3fold_page_isolate() can be static] Link: http://lkml.kernel.org/r/20190419130924.GA161478@ivb42 Link: http://lkml.kernel.org/r/20190417103922.31253da5c366c4ebe0419cfc@gmail.com Signed-off-by: Vitaly Wool <vitaly.vul@sony.com> Signed-off-by: kbuild test robot <lkp@intel.com> Cc: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Krzysztof Kozlowski <k.kozlowski@samsung.com> Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com> Cc: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-14 08:22:52 +08:00
z3fold_unmount();
z3fold: the 3-fold allocator for compressed pages This patch introduces z3fold, a special purpose allocator for storing compressed pages. It is designed to store up to three compressed pages per physical page. It is a ZBUD derivative which allows for higher compression ratio keeping the simplicity and determinism of its predecessor. This patch comes as a follow-up to the discussions at the Embedded Linux Conference in San-Diego related to the talk [1]. The outcome of these discussions was that it would be good to have a compressed page allocator as stable and deterministic as zbud with with higher compression ratio. To keep the determinism and simplicity, z3fold, just like zbud, always stores an integral number of compressed pages per page, but it can store up to 3 pages unlike zbud which can store at most 2. Therefore the compression ratio goes to around 2.6x while zbud's one is around 1.7x. The patch is based on the latest linux.git tree. This version has been updated after testing on various simulators (e.g. ARM Versatile Express, MIPS Malta, x86_64/Haswell) and basing on comments from Dan Streetman [3]. [1] https://openiotelc2016.sched.org/event/6DAC/swapping-and-embedded-compression-relieves-the-pressure-vitaly-wool-softprise-consulting-ou [2] https://lkml.org/lkml/2016/4/21/799 [3] https://lkml.org/lkml/2016/5/4/852 Link: http://lkml.kernel.org/r/20160509151753.ec3f9fda3c9898d31ff52a32@gmail.com Signed-off-by: Vitaly Wool <vitalywool@gmail.com> Cc: Seth Jennings <sjenning@redhat.com> Cc: Dan Streetman <ddstreet@ieee.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>
2016-05-21 07:58:30 +08:00
zpool_unregister_driver(&z3fold_zpool_driver);
}
module_init(init_z3fold);
module_exit(exit_z3fold);
MODULE_LICENSE("GPL");
MODULE_AUTHOR("Vitaly Wool <vitalywool@gmail.com>");
MODULE_DESCRIPTION("3-Fold Allocator for Compressed Pages");