2019-05-19 20:08:55 +08:00
|
|
|
// SPDX-License-Identifier: GPL-2.0-only
|
2005-04-17 06:20:36 +08:00
|
|
|
/*
|
|
|
|
* linux/mm/vmalloc.c
|
|
|
|
*
|
|
|
|
* Copyright (C) 1993 Linus Torvalds
|
|
|
|
* Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
|
|
|
|
* SMP-safe vmalloc/vfree/ioremap, Tigran Aivazian <tigran@veritas.com>, May 2000
|
|
|
|
* Major rework to support vmap/vunmap, Christoph Hellwig, SGI, August 2002
|
2005-10-30 09:15:41 +08:00
|
|
|
* Numa awareness, Christoph Lameter, SGI, June 2005
|
2005-04-17 06:20:36 +08:00
|
|
|
*/
|
|
|
|
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
#include <linux/vmalloc.h>
|
2005-04-17 06:20:36 +08:00
|
|
|
#include <linux/mm.h>
|
|
|
|
#include <linux/module.h>
|
|
|
|
#include <linux/highmem.h>
|
2017-02-02 15:35:14 +08:00
|
|
|
#include <linux/sched/signal.h>
|
2005-04-17 06:20:36 +08:00
|
|
|
#include <linux/slab.h>
|
|
|
|
#include <linux/spinlock.h>
|
|
|
|
#include <linux/interrupt.h>
|
2008-10-06 07:50:47 +08:00
|
|
|
#include <linux/proc_fs.h>
|
2008-04-28 17:12:40 +08:00
|
|
|
#include <linux/seq_file.h>
|
2019-04-26 08:11:36 +08:00
|
|
|
#include <linux/set_memory.h>
|
2008-04-30 15:55:01 +08:00
|
|
|
#include <linux/debugobjects.h>
|
2008-04-28 17:12:42 +08:00
|
|
|
#include <linux/kallsyms.h>
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
#include <linux/list.h>
|
2016-04-04 21:46:42 +08:00
|
|
|
#include <linux/notifier.h>
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
#include <linux/rbtree.h>
|
|
|
|
#include <linux/radix-tree.h>
|
|
|
|
#include <linux/rcupdate.h>
|
2009-02-20 15:29:08 +08:00
|
|
|
#include <linux/pfn.h>
|
2009-06-11 20:23:19 +08:00
|
|
|
#include <linux/kmemleak.h>
|
2011-07-27 07:09:06 +08:00
|
|
|
#include <linux/atomic.h>
|
2014-04-08 06:37:26 +08:00
|
|
|
#include <linux/compiler.h>
|
2013-03-11 08:14:08 +08:00
|
|
|
#include <linux/llist.h>
|
mm: change __get_vm_area_node() to use fls_long()
ioremap() and its related interfaces are used to create I/O mappings to
memory-mapped I/O devices. The mapping sizes of the traditional I/O
devices are relatively small. Non-volatile memory (NVM), however, has
many GB and is going to have TB soon. It is not very efficient to create
large I/O mappings with 4KB.
This patchset extends the ioremap() interfaces to transparently create I/O
mappings with huge pages whenever possible. ioremap() continues to use
4KB mappings when a huge page does not fit into a requested range. There
is no change necessary to the drivers using ioremap(). A requested
physical address must be aligned by a huge page size (1GB or 2MB on x86)
for using huge page mapping, though. The kernel huge I/O mapping will
improve performance of NVM and other devices with large memory, and reduce
the time to create their mappings as well.
On x86, MTRRs can override PAT memory types with a 4KB granularity. When
using a huge page, MTRRs can override the memory type of the huge page,
which may lead a performance penalty. The processor can also behave in an
undefined manner if a huge page is mapped to a memory range that MTRRs
have mapped with multiple different memory types. Therefore, the mapping
code falls back to use a smaller page size toward 4KB when a mapping range
is covered by non-WB type of MTRRs. The WB type of MTRRs has no affect on
the PAT memory types.
The patchset introduces HAVE_ARCH_HUGE_VMAP, which indicates that the arch
supports huge KVA mappings for ioremap(). User may specify a new kernel
option "nohugeiomap" to disable the huge I/O mapping capability of
ioremap() when necessary.
Patch 1-4 change common files to support huge I/O mappings. There is no
change in the functinalities unless HAVE_ARCH_HUGE_VMAP is defined on the
architecture of the system.
Patch 5-6 implement the HAVE_ARCH_HUGE_VMAP funcs on x86, and set
HAVE_ARCH_HUGE_VMAP on x86.
This patch (of 6):
__get_vm_area_node() takes unsigned long size, which is a 64-bit value on
a 64-bit kernel. However, fls(size) simply ignores the upper 32-bit.
Change to use fls_long() to handle the size properly.
Signed-off-by: Toshi Kani <toshi.kani@hp.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Robert Elliott <Elliott@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-04-15 06:47:17 +08:00
|
|
|
#include <linux/bitops.h>
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
#include <linux/rbtree_augmented.h>
|
2014-04-08 06:37:26 +08:00
|
|
|
|
2016-12-25 03:46:01 +08:00
|
|
|
#include <linux/uaccess.h>
|
2005-04-17 06:20:36 +08:00
|
|
|
#include <asm/tlbflush.h>
|
2009-09-22 03:22:34 +08:00
|
|
|
#include <asm/shmparam.h>
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2015-11-07 08:28:43 +08:00
|
|
|
#include "internal.h"
|
|
|
|
|
2019-11-29 15:17:25 +08:00
|
|
|
bool is_vmalloc_addr(const void *x)
|
|
|
|
{
|
|
|
|
unsigned long addr = (unsigned long)x;
|
|
|
|
|
|
|
|
return addr >= VMALLOC_START && addr < VMALLOC_END;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(is_vmalloc_addr);
|
|
|
|
|
2013-03-11 08:14:08 +08:00
|
|
|
struct vfree_deferred {
|
|
|
|
struct llist_head list;
|
|
|
|
struct work_struct wq;
|
|
|
|
};
|
|
|
|
static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
|
|
|
|
|
|
|
|
static void __vunmap(const void *, int);
|
|
|
|
|
|
|
|
static void free_work(struct work_struct *w)
|
|
|
|
{
|
|
|
|
struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
|
2017-09-07 07:24:26 +08:00
|
|
|
struct llist_node *t, *llnode;
|
|
|
|
|
|
|
|
llist_for_each_safe(llnode, t, llist_del_all(&p->list))
|
|
|
|
__vunmap((void *)llnode, 1);
|
2013-03-11 08:14:08 +08:00
|
|
|
}
|
|
|
|
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
/*** Page table manipulation functions ***/
|
2006-09-26 14:31:02 +08:00
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end)
|
|
|
|
{
|
|
|
|
pte_t *pte;
|
|
|
|
|
|
|
|
pte = pte_offset_kernel(pmd, addr);
|
|
|
|
do {
|
|
|
|
pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
|
|
|
|
WARN_ON(!pte_none(ptent) && !pte_present(ptent));
|
|
|
|
} while (pte++, addr += PAGE_SIZE, addr != end);
|
|
|
|
}
|
|
|
|
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
|
|
|
pmd_t *pmd;
|
|
|
|
unsigned long next;
|
|
|
|
|
|
|
|
pmd = pmd_offset(pud, addr);
|
|
|
|
do {
|
|
|
|
next = pmd_addr_end(addr, end);
|
2015-04-15 06:47:26 +08:00
|
|
|
if (pmd_clear_huge(pmd))
|
|
|
|
continue;
|
2005-04-17 06:20:36 +08:00
|
|
|
if (pmd_none_or_clear_bad(pmd))
|
|
|
|
continue;
|
|
|
|
vunmap_pte_range(pmd, addr, next);
|
|
|
|
} while (pmd++, addr = next, addr != end);
|
|
|
|
}
|
|
|
|
|
2017-03-09 22:24:07 +08:00
|
|
|
static void vunmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
|
|
|
pud_t *pud;
|
|
|
|
unsigned long next;
|
|
|
|
|
2017-03-09 22:24:07 +08:00
|
|
|
pud = pud_offset(p4d, addr);
|
2005-04-17 06:20:36 +08:00
|
|
|
do {
|
|
|
|
next = pud_addr_end(addr, end);
|
2015-04-15 06:47:26 +08:00
|
|
|
if (pud_clear_huge(pud))
|
|
|
|
continue;
|
2005-04-17 06:20:36 +08:00
|
|
|
if (pud_none_or_clear_bad(pud))
|
|
|
|
continue;
|
|
|
|
vunmap_pmd_range(pud, addr, next);
|
|
|
|
} while (pud++, addr = next, addr != end);
|
|
|
|
}
|
|
|
|
|
2017-03-09 22:24:07 +08:00
|
|
|
static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end)
|
|
|
|
{
|
|
|
|
p4d_t *p4d;
|
|
|
|
unsigned long next;
|
|
|
|
|
|
|
|
p4d = p4d_offset(pgd, addr);
|
|
|
|
do {
|
|
|
|
next = p4d_addr_end(addr, end);
|
|
|
|
if (p4d_clear_huge(p4d))
|
|
|
|
continue;
|
|
|
|
if (p4d_none_or_clear_bad(p4d))
|
|
|
|
continue;
|
|
|
|
vunmap_pud_range(p4d, addr, next);
|
|
|
|
} while (p4d++, addr = next, addr != end);
|
|
|
|
}
|
|
|
|
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
static void vunmap_page_range(unsigned long addr, unsigned long end)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
|
|
|
pgd_t *pgd;
|
|
|
|
unsigned long next;
|
|
|
|
|
|
|
|
BUG_ON(addr >= end);
|
|
|
|
pgd = pgd_offset_k(addr);
|
|
|
|
do {
|
|
|
|
next = pgd_addr_end(addr, end);
|
|
|
|
if (pgd_none_or_clear_bad(pgd))
|
|
|
|
continue;
|
2017-03-09 22:24:07 +08:00
|
|
|
vunmap_p4d_range(pgd, addr, next);
|
2005-04-17 06:20:36 +08:00
|
|
|
} while (pgd++, addr = next, addr != end);
|
|
|
|
}
|
|
|
|
|
|
|
|
static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
unsigned long end, pgprot_t prot, struct page **pages, int *nr)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
|
|
|
pte_t *pte;
|
|
|
|
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
/*
|
|
|
|
* nr is a running index into the array which helps higher level
|
|
|
|
* callers keep track of where we're up to.
|
|
|
|
*/
|
|
|
|
|
[PATCH] mm: init_mm without ptlock
First step in pushing down the page_table_lock. init_mm.page_table_lock has
been used throughout the architectures (usually for ioremap): not to serialize
kernel address space allocation (that's usually vmlist_lock), but because
pud_alloc,pmd_alloc,pte_alloc_kernel expect caller holds it.
Reverse that: don't lock or unlock init_mm.page_table_lock in any of the
architectures; instead rely on pud_alloc,pmd_alloc,pte_alloc_kernel to take
and drop it when allocating a new one, to check lest a racing task already
did. Similarly no page_table_lock in vmalloc's map_vm_area.
Some temporary ugliness in __pud_alloc and __pmd_alloc: since they also handle
user mms, which are converted only by a later patch, for now they have to lock
differently according to whether or not it's init_mm.
If sources get muddled, there's a danger that an arch source taking
init_mm.page_table_lock will be mixed with common source also taking it (or
neither take it). So break the rules and make another change, which should
break the build for such a mismatch: remove the redundant mm arg from
pte_alloc_kernel (ppc64 scrapped its distinct ioremap_mm in 2.6.13).
Exceptions: arm26 used pte_alloc_kernel on user mm, now pte_alloc_map; ia64
used pte_alloc_map on init_mm, now pte_alloc_kernel; parisc had bad args to
pmd_alloc and pte_alloc_kernel in unused USE_HPPA_IOREMAP code; ppc64
map_io_page forgot to unlock on failure; ppc mmu_mapin_ram and ppc64 im_free
took page_table_lock for no good reason.
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-10-30 09:16:21 +08:00
|
|
|
pte = pte_alloc_kernel(pmd, addr);
|
2005-04-17 06:20:36 +08:00
|
|
|
if (!pte)
|
|
|
|
return -ENOMEM;
|
|
|
|
do {
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
struct page *page = pages[*nr];
|
|
|
|
|
|
|
|
if (WARN_ON(!pte_none(*pte)))
|
|
|
|
return -EBUSY;
|
|
|
|
if (WARN_ON(!page))
|
2005-04-17 06:20:36 +08:00
|
|
|
return -ENOMEM;
|
|
|
|
set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
(*nr)++;
|
2005-04-17 06:20:36 +08:00
|
|
|
} while (pte++, addr += PAGE_SIZE, addr != end);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
static int vmap_pmd_range(pud_t *pud, unsigned long addr,
|
|
|
|
unsigned long end, pgprot_t prot, struct page **pages, int *nr)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
|
|
|
pmd_t *pmd;
|
|
|
|
unsigned long next;
|
|
|
|
|
|
|
|
pmd = pmd_alloc(&init_mm, pud, addr);
|
|
|
|
if (!pmd)
|
|
|
|
return -ENOMEM;
|
|
|
|
do {
|
|
|
|
next = pmd_addr_end(addr, end);
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
if (vmap_pte_range(pmd, addr, next, prot, pages, nr))
|
2005-04-17 06:20:36 +08:00
|
|
|
return -ENOMEM;
|
|
|
|
} while (pmd++, addr = next, addr != end);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2017-03-09 22:24:07 +08:00
|
|
|
static int vmap_pud_range(p4d_t *p4d, unsigned long addr,
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
unsigned long end, pgprot_t prot, struct page **pages, int *nr)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
|
|
|
pud_t *pud;
|
|
|
|
unsigned long next;
|
|
|
|
|
2017-03-09 22:24:07 +08:00
|
|
|
pud = pud_alloc(&init_mm, p4d, addr);
|
2005-04-17 06:20:36 +08:00
|
|
|
if (!pud)
|
|
|
|
return -ENOMEM;
|
|
|
|
do {
|
|
|
|
next = pud_addr_end(addr, end);
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
if (vmap_pmd_range(pud, addr, next, prot, pages, nr))
|
2005-04-17 06:20:36 +08:00
|
|
|
return -ENOMEM;
|
|
|
|
} while (pud++, addr = next, addr != end);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2017-03-09 22:24:07 +08:00
|
|
|
static int vmap_p4d_range(pgd_t *pgd, unsigned long addr,
|
|
|
|
unsigned long end, pgprot_t prot, struct page **pages, int *nr)
|
|
|
|
{
|
|
|
|
p4d_t *p4d;
|
|
|
|
unsigned long next;
|
|
|
|
|
|
|
|
p4d = p4d_alloc(&init_mm, pgd, addr);
|
|
|
|
if (!p4d)
|
|
|
|
return -ENOMEM;
|
|
|
|
do {
|
|
|
|
next = p4d_addr_end(addr, end);
|
|
|
|
if (vmap_pud_range(p4d, addr, next, prot, pages, nr))
|
|
|
|
return -ENOMEM;
|
|
|
|
} while (p4d++, addr = next, addr != end);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
/*
|
|
|
|
* Set up page tables in kva (addr, end). The ptes shall have prot "prot", and
|
|
|
|
* will have pfns corresponding to the "pages" array.
|
|
|
|
*
|
|
|
|
* Ie. pte at addr+N*PAGE_SIZE shall point to pfn corresponding to pages[N]
|
|
|
|
*/
|
2009-02-20 15:29:08 +08:00
|
|
|
static int vmap_page_range_noflush(unsigned long start, unsigned long end,
|
|
|
|
pgprot_t prot, struct page **pages)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
|
|
|
pgd_t *pgd;
|
|
|
|
unsigned long next;
|
2009-01-05 04:00:46 +08:00
|
|
|
unsigned long addr = start;
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
int err = 0;
|
|
|
|
int nr = 0;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
BUG_ON(addr >= end);
|
|
|
|
pgd = pgd_offset_k(addr);
|
|
|
|
do {
|
|
|
|
next = pgd_addr_end(addr, end);
|
2017-03-09 22:24:07 +08:00
|
|
|
err = vmap_p4d_range(pgd, addr, next, prot, pages, &nr);
|
2005-04-17 06:20:36 +08:00
|
|
|
if (err)
|
2009-09-22 08:01:47 +08:00
|
|
|
return err;
|
2005-04-17 06:20:36 +08:00
|
|
|
} while (pgd++, addr = next, addr != end);
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
|
|
|
|
return nr;
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
2009-02-20 15:29:08 +08:00
|
|
|
static int vmap_page_range(unsigned long start, unsigned long end,
|
|
|
|
pgprot_t prot, struct page **pages)
|
|
|
|
{
|
|
|
|
int ret;
|
|
|
|
|
|
|
|
ret = vmap_page_range_noflush(start, end, prot, pages);
|
|
|
|
flush_cache_vmap(start, end);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2009-09-23 07:45:49 +08:00
|
|
|
int is_vmalloc_or_module_addr(const void *x)
|
2008-10-15 23:35:12 +08:00
|
|
|
{
|
|
|
|
/*
|
2008-11-07 01:11:07 +08:00
|
|
|
* ARM, x86-64 and sparc64 put modules in a special place,
|
2008-10-15 23:35:12 +08:00
|
|
|
* and fall back on vmalloc() if that fails. Others
|
|
|
|
* just put it in the vmalloc space.
|
|
|
|
*/
|
|
|
|
#if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
|
|
|
|
unsigned long addr = (unsigned long)x;
|
|
|
|
if (addr >= MODULES_VADDR && addr < MODULES_END)
|
|
|
|
return 1;
|
|
|
|
#endif
|
|
|
|
return is_vmalloc_addr(x);
|
|
|
|
}
|
|
|
|
|
2008-02-05 14:28:31 +08:00
|
|
|
/*
|
2014-01-28 09:06:53 +08:00
|
|
|
* Walk a vmap address to the struct page it maps.
|
2008-02-05 14:28:31 +08:00
|
|
|
*/
|
2014-01-28 09:06:53 +08:00
|
|
|
struct page *vmalloc_to_page(const void *vmalloc_addr)
|
2008-02-05 14:28:31 +08:00
|
|
|
{
|
|
|
|
unsigned long addr = (unsigned long) vmalloc_addr;
|
2014-01-28 09:06:53 +08:00
|
|
|
struct page *page = NULL;
|
2008-02-05 14:28:31 +08:00
|
|
|
pgd_t *pgd = pgd_offset_k(addr);
|
2017-03-09 22:24:07 +08:00
|
|
|
p4d_t *p4d;
|
|
|
|
pud_t *pud;
|
|
|
|
pmd_t *pmd;
|
|
|
|
pte_t *ptep, pte;
|
2008-02-05 14:28:31 +08:00
|
|
|
|
2008-06-19 19:28:11 +08:00
|
|
|
/*
|
|
|
|
* XXX we might need to change this if we add VIRTUAL_BUG_ON for
|
|
|
|
* architectures that do not vmalloc module space
|
|
|
|
*/
|
2008-10-15 23:35:12 +08:00
|
|
|
VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
|
2008-06-12 19:56:40 +08:00
|
|
|
|
2017-03-09 22:24:07 +08:00
|
|
|
if (pgd_none(*pgd))
|
|
|
|
return NULL;
|
|
|
|
p4d = p4d_offset(pgd, addr);
|
|
|
|
if (p4d_none(*p4d))
|
|
|
|
return NULL;
|
|
|
|
pud = pud_offset(p4d, addr);
|
mm/vmalloc.c: huge-vmap: fail gracefully on unexpected huge vmap mappings
Existing code that uses vmalloc_to_page() may assume that any address
for which is_vmalloc_addr() returns true may be passed into
vmalloc_to_page() to retrieve the associated struct page.
This is not un unreasonable assumption to make, but on architectures
that have CONFIG_HAVE_ARCH_HUGE_VMAP=y, it no longer holds, and we need
to ensure that vmalloc_to_page() does not go off into the weeds trying
to dereference huge PUDs or PMDs as table entries.
Given that vmalloc() and vmap() themselves never create huge mappings or
deal with compound pages at all, there is no correct answer in this
case, so return NULL instead, and issue a warning.
When reading /proc/kcore on arm64, you will hit an oops as soon as you
hit the huge mappings used for the various segments that make up the
mapping of vmlinux. With this patch applied, you will no longer hit the
oops, but the kcore contents willl be incorrect (these regions will be
zeroed out)
We are fixing this for kcore specifically, so it avoids vread() for
those regions. At least one other problematic user exists, i.e.,
/dev/kmem, but that is currently broken on arm64 for other reasons.
Link: http://lkml.kernel.org/r/20170609082226.26152-1-ard.biesheuvel@linaro.org
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Acked-by: Mark Rutland <mark.rutland@arm.com>
Reviewed-by: Laura Abbott <labbott@redhat.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: zhong jiang <zhongjiang@huawei.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-24 06:08:41 +08:00
|
|
|
|
|
|
|
/*
|
|
|
|
* Don't dereference bad PUD or PMD (below) entries. This will also
|
|
|
|
* identify huge mappings, which we may encounter on architectures
|
|
|
|
* that define CONFIG_HAVE_ARCH_HUGE_VMAP=y. Such regions will be
|
|
|
|
* identified as vmalloc addresses by is_vmalloc_addr(), but are
|
|
|
|
* not [unambiguously] associated with a struct page, so there is
|
|
|
|
* no correct value to return for them.
|
|
|
|
*/
|
|
|
|
WARN_ON_ONCE(pud_bad(*pud));
|
|
|
|
if (pud_none(*pud) || pud_bad(*pud))
|
2017-03-09 22:24:07 +08:00
|
|
|
return NULL;
|
|
|
|
pmd = pmd_offset(pud, addr);
|
mm/vmalloc.c: huge-vmap: fail gracefully on unexpected huge vmap mappings
Existing code that uses vmalloc_to_page() may assume that any address
for which is_vmalloc_addr() returns true may be passed into
vmalloc_to_page() to retrieve the associated struct page.
This is not un unreasonable assumption to make, but on architectures
that have CONFIG_HAVE_ARCH_HUGE_VMAP=y, it no longer holds, and we need
to ensure that vmalloc_to_page() does not go off into the weeds trying
to dereference huge PUDs or PMDs as table entries.
Given that vmalloc() and vmap() themselves never create huge mappings or
deal with compound pages at all, there is no correct answer in this
case, so return NULL instead, and issue a warning.
When reading /proc/kcore on arm64, you will hit an oops as soon as you
hit the huge mappings used for the various segments that make up the
mapping of vmlinux. With this patch applied, you will no longer hit the
oops, but the kcore contents willl be incorrect (these regions will be
zeroed out)
We are fixing this for kcore specifically, so it avoids vread() for
those regions. At least one other problematic user exists, i.e.,
/dev/kmem, but that is currently broken on arm64 for other reasons.
Link: http://lkml.kernel.org/r/20170609082226.26152-1-ard.biesheuvel@linaro.org
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Acked-by: Mark Rutland <mark.rutland@arm.com>
Reviewed-by: Laura Abbott <labbott@redhat.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: zhong jiang <zhongjiang@huawei.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-24 06:08:41 +08:00
|
|
|
WARN_ON_ONCE(pmd_bad(*pmd));
|
|
|
|
if (pmd_none(*pmd) || pmd_bad(*pmd))
|
2017-03-09 22:24:07 +08:00
|
|
|
return NULL;
|
|
|
|
|
|
|
|
ptep = pte_offset_map(pmd, addr);
|
|
|
|
pte = *ptep;
|
|
|
|
if (pte_present(pte))
|
|
|
|
page = pte_page(pte);
|
|
|
|
pte_unmap(ptep);
|
2014-01-28 09:06:53 +08:00
|
|
|
return page;
|
2008-02-05 14:28:31 +08:00
|
|
|
}
|
2014-01-28 09:06:53 +08:00
|
|
|
EXPORT_SYMBOL(vmalloc_to_page);
|
2008-02-05 14:28:31 +08:00
|
|
|
|
|
|
|
/*
|
2014-01-28 09:06:53 +08:00
|
|
|
* Map a vmalloc()-space virtual address to the physical page frame number.
|
2008-02-05 14:28:31 +08:00
|
|
|
*/
|
2014-01-28 09:06:53 +08:00
|
|
|
unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
|
2008-02-05 14:28:31 +08:00
|
|
|
{
|
2014-01-28 09:06:53 +08:00
|
|
|
return page_to_pfn(vmalloc_to_page(vmalloc_addr));
|
2008-02-05 14:28:31 +08:00
|
|
|
}
|
2014-01-28 09:06:53 +08:00
|
|
|
EXPORT_SYMBOL(vmalloc_to_pfn);
|
2008-02-05 14:28:31 +08:00
|
|
|
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
|
|
|
|
/*** Global kva allocator ***/
|
|
|
|
|
2019-05-18 05:31:34 +08:00
|
|
|
#define DEBUG_AUGMENT_PROPAGATE_CHECK 0
|
2019-05-18 05:31:37 +08:00
|
|
|
#define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
|
2019-05-18 05:31:34 +08:00
|
|
|
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
|
|
|
|
static DEFINE_SPINLOCK(vmap_area_lock);
|
mm/vmalloc: rework vmap_area_lock
With the new allocation approach introduced in the 5.2 kernel, it
becomes possible to get rid of one global spinlock. By doing that we
can further improve the KVA from the performance point of view.
Basically we can have two independent locks, one for allocation part and
another one for deallocation, because of two different entities: "free
data structures" and "busy data structures".
As a result, allocation/deallocation operations can still interfere
between each other in case of running simultaneously on different CPUs,
it means there is still dependency, but with two locks it becomes lower.
Summarizing:
- it reduces the high lock contention
- it allows to perform operations on "free" and "busy"
trees in parallel on different CPUs. Please note it
does not solve scalability issue.
Test results:
In order to evaluate this patch, we can run "vmalloc test driver" to see
how many CPU cycles it takes to complete all test cases running
sequentially. All online CPUs run it so it will cause a high lock
contention.
HiKey 960, ARM64, 8xCPUs, big.LITTLE:
<snip>
sudo ./test_vmalloc.sh sequential_test_order=1
<snip>
<default>
[ 390.950557] All test took CPU0=457126382 cycles
[ 391.046690] All test took CPU1=454763452 cycles
[ 391.128586] All test took CPU2=454539334 cycles
[ 391.222669] All test took CPU3=455649517 cycles
[ 391.313946] All test took CPU4=388272196 cycles
[ 391.410425] All test took CPU5=384036264 cycles
[ 391.492219] All test took CPU6=387432964 cycles
[ 391.578433] All test took CPU7=387201996 cycles
<default>
<patched>
[ 304.721224] All test took CPU0=391521310 cycles
[ 304.821219] All test took CPU1=393533002 cycles
[ 304.917120] All test took CPU2=392243032 cycles
[ 305.008986] All test took CPU3=392353853 cycles
[ 305.108944] All test took CPU4=297630721 cycles
[ 305.196406] All test took CPU5=297548736 cycles
[ 305.288602] All test took CPU6=297092392 cycles
[ 305.381088] All test took CPU7=297293597 cycles
<patched>
~14%-23% patched variant is better.
Link: http://lkml.kernel.org/r/20191022155800.20468-1-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Acked-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:47 +08:00
|
|
|
static DEFINE_SPINLOCK(free_vmap_area_lock);
|
2013-04-30 06:07:37 +08:00
|
|
|
/* Export for kexec only */
|
|
|
|
LIST_HEAD(vmap_area_list);
|
2016-05-21 07:57:38 +08:00
|
|
|
static LLIST_HEAD(vmap_purge_list);
|
mm: vmap area cache
Provide a free area cache for the vmalloc virtual address allocator, based
on the algorithm used by the user virtual memory allocator.
This reduces the number of rbtree operations and linear traversals over
the vmap extents in order to find a free area, by starting off at the last
point that a free area was found.
The free area cache is reset if areas are freed behind it, or if we are
searching for a smaller area or alignment than last time. So allocation
patterns are not changed (verified by corner-case and random test cases in
userspace testing).
This solves a regression caused by lazy vunmap TLB purging introduced in
db64fe02 (mm: rewrite vmap layer). That patch will leave extents in the
vmap allocator after they are vunmapped, and until a significant number
accumulate that can be flushed in a single batch. So in a workload that
vmalloc/vfree frequently, a chain of extents will build up from
VMALLOC_START address, which have to be iterated over each time (giving an
O(n) type of behaviour).
After this patch, the search will start from where it left off, giving
closer to an amortized O(1).
This is verified to solve regressions reported Steven in GFS2, and Avi in
KVM.
Hugh's update:
: I tried out the recent mmotm, and on one machine was fortunate to hit
: the BUG_ON(first->va_start < addr) which seems to have been stalling
: your vmap area cache patch ever since May.
: I can get you addresses etc, I did dump a few out; but once I stared
: at them, it was easier just to look at the code: and I cannot see how
: you would be so sure that first->va_start < addr, once you've done
: that addr = ALIGN(max(...), align) above, if align is over 0x1000
: (align was 0x8000 or 0x4000 in the cases I hit: ioremaps like Steve).
: I originally got around it by just changing the
: if (first->va_start < addr) {
: to
: while (first->va_start < addr) {
: without thinking about it any further; but that seemed unsatisfactory,
: why would we want to loop here when we've got another very similar
: loop just below it?
: I am never going to admit how long I've spent trying to grasp your
: "while (n)" rbtree loop just above this, the one with the peculiar
: if (!first && tmp->va_start < addr + size)
: in. That's unfamiliar to me, I'm guessing it's designed to save a
: subsequent rb_next() in a few circumstances (at risk of then setting
: a wrong cached_hole_size?); but they did appear few to me, and I didn't
: feel I could sign off something with that in when I don't grasp it,
: and it seems responsible for extra code and mistaken BUG_ON below it.
: I've reverted to the familiar rbtree loop that find_vma() does (but
: with va_end >= addr as you had, to respect the additional guard page):
: and then (given that cached_hole_size starts out 0) I don't see the
: need for any complications below it. If you do want to keep that loop
: as you had it, please add a comment to explain what it's trying to do,
: and where addr is relative to first when you emerge from it.
: Aren't your tests "size <= cached_hole_size" and
: "addr + size > first->va_start" forgetting the guard page we want
: before the next area? I've changed those.
: I have not changed your many "addr + size - 1 < addr" overflow tests,
: but have since come to wonder, shouldn't they be "addr + size < addr"
: tests - won't the vend checks go wrong if addr + size is 0?
: I have added a few comments - Wolfgang Wander's 2.6.13 description of
: 1363c3cd8603a913a27e2995dccbd70d5312d8e6 Avoiding mmap fragmentation
: helped me a lot, perhaps a pointer to that would be good too. And I found
: it easier to understand when I renamed cached_start slightly and moved the
: overflow label down.
: This patch would go after your mm-vmap-area-cache.patch in mmotm.
: Trivially, nobody is going to get that BUG_ON with this patch, and it
: appears to work fine on my machines; but I have not given it anything like
: the testing you did on your original, and may have broken all the
: performance you were aiming for. Please take a look and test it out
: integrate with yours if you're satisfied - thanks.
[akpm@linux-foundation.org: add locking comment]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Reported-and-tested-by: Steven Whitehouse <swhiteho@redhat.com>
Reported-and-tested-by: Avi Kivity <avi@redhat.com>
Tested-by: "Barry J. Marson" <bmarson@redhat.com>
Cc: Prarit Bhargava <prarit@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-03-23 07:30:36 +08:00
|
|
|
static struct rb_root vmap_area_root = RB_ROOT;
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
static bool vmap_initialized __read_mostly;
|
mm: vmap area cache
Provide a free area cache for the vmalloc virtual address allocator, based
on the algorithm used by the user virtual memory allocator.
This reduces the number of rbtree operations and linear traversals over
the vmap extents in order to find a free area, by starting off at the last
point that a free area was found.
The free area cache is reset if areas are freed behind it, or if we are
searching for a smaller area or alignment than last time. So allocation
patterns are not changed (verified by corner-case and random test cases in
userspace testing).
This solves a regression caused by lazy vunmap TLB purging introduced in
db64fe02 (mm: rewrite vmap layer). That patch will leave extents in the
vmap allocator after they are vunmapped, and until a significant number
accumulate that can be flushed in a single batch. So in a workload that
vmalloc/vfree frequently, a chain of extents will build up from
VMALLOC_START address, which have to be iterated over each time (giving an
O(n) type of behaviour).
After this patch, the search will start from where it left off, giving
closer to an amortized O(1).
This is verified to solve regressions reported Steven in GFS2, and Avi in
KVM.
Hugh's update:
: I tried out the recent mmotm, and on one machine was fortunate to hit
: the BUG_ON(first->va_start < addr) which seems to have been stalling
: your vmap area cache patch ever since May.
: I can get you addresses etc, I did dump a few out; but once I stared
: at them, it was easier just to look at the code: and I cannot see how
: you would be so sure that first->va_start < addr, once you've done
: that addr = ALIGN(max(...), align) above, if align is over 0x1000
: (align was 0x8000 or 0x4000 in the cases I hit: ioremaps like Steve).
: I originally got around it by just changing the
: if (first->va_start < addr) {
: to
: while (first->va_start < addr) {
: without thinking about it any further; but that seemed unsatisfactory,
: why would we want to loop here when we've got another very similar
: loop just below it?
: I am never going to admit how long I've spent trying to grasp your
: "while (n)" rbtree loop just above this, the one with the peculiar
: if (!first && tmp->va_start < addr + size)
: in. That's unfamiliar to me, I'm guessing it's designed to save a
: subsequent rb_next() in a few circumstances (at risk of then setting
: a wrong cached_hole_size?); but they did appear few to me, and I didn't
: feel I could sign off something with that in when I don't grasp it,
: and it seems responsible for extra code and mistaken BUG_ON below it.
: I've reverted to the familiar rbtree loop that find_vma() does (but
: with va_end >= addr as you had, to respect the additional guard page):
: and then (given that cached_hole_size starts out 0) I don't see the
: need for any complications below it. If you do want to keep that loop
: as you had it, please add a comment to explain what it's trying to do,
: and where addr is relative to first when you emerge from it.
: Aren't your tests "size <= cached_hole_size" and
: "addr + size > first->va_start" forgetting the guard page we want
: before the next area? I've changed those.
: I have not changed your many "addr + size - 1 < addr" overflow tests,
: but have since come to wonder, shouldn't they be "addr + size < addr"
: tests - won't the vend checks go wrong if addr + size is 0?
: I have added a few comments - Wolfgang Wander's 2.6.13 description of
: 1363c3cd8603a913a27e2995dccbd70d5312d8e6 Avoiding mmap fragmentation
: helped me a lot, perhaps a pointer to that would be good too. And I found
: it easier to understand when I renamed cached_start slightly and moved the
: overflow label down.
: This patch would go after your mm-vmap-area-cache.patch in mmotm.
: Trivially, nobody is going to get that BUG_ON with this patch, and it
: appears to work fine on my machines; but I have not given it anything like
: the testing you did on your original, and may have broken all the
: performance you were aiming for. Please take a look and test it out
: integrate with yours if you're satisfied - thanks.
[akpm@linux-foundation.org: add locking comment]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Reported-and-tested-by: Steven Whitehouse <swhiteho@redhat.com>
Reported-and-tested-by: Avi Kivity <avi@redhat.com>
Tested-by: "Barry J. Marson" <bmarson@redhat.com>
Cc: Prarit Bhargava <prarit@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-03-23 07:30:36 +08:00
|
|
|
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
/*
|
|
|
|
* This kmem_cache is used for vmap_area objects. Instead of
|
|
|
|
* allocating from slab we reuse an object from this cache to
|
|
|
|
* make things faster. Especially in "no edge" splitting of
|
|
|
|
* free block.
|
|
|
|
*/
|
|
|
|
static struct kmem_cache *vmap_area_cachep;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* This linked list is used in pair with free_vmap_area_root.
|
|
|
|
* It gives O(1) access to prev/next to perform fast coalescing.
|
|
|
|
*/
|
|
|
|
static LIST_HEAD(free_vmap_area_list);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* This augment red-black tree represents the free vmap space.
|
|
|
|
* All vmap_area objects in this tree are sorted by va->va_start
|
|
|
|
* address. It is used for allocation and merging when a vmap
|
|
|
|
* object is released.
|
|
|
|
*
|
|
|
|
* Each vmap_area node contains a maximum available free block
|
|
|
|
* of its sub-tree, right or left. Therefore it is possible to
|
|
|
|
* find a lowest match of free area.
|
|
|
|
*/
|
|
|
|
static struct rb_root free_vmap_area_root = RB_ROOT;
|
|
|
|
|
2019-07-12 11:58:57 +08:00
|
|
|
/*
|
|
|
|
* Preload a CPU with one object for "no edge" split case. The
|
|
|
|
* aim is to get rid of allocations from the atomic context, thus
|
|
|
|
* to use more permissive allocation masks.
|
|
|
|
*/
|
|
|
|
static DEFINE_PER_CPU(struct vmap_area *, ne_fit_preload_node);
|
|
|
|
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
static __always_inline unsigned long
|
|
|
|
va_size(struct vmap_area *va)
|
|
|
|
{
|
|
|
|
return (va->va_end - va->va_start);
|
|
|
|
}
|
|
|
|
|
|
|
|
static __always_inline unsigned long
|
|
|
|
get_subtree_max_size(struct rb_node *node)
|
|
|
|
{
|
|
|
|
struct vmap_area *va;
|
|
|
|
|
|
|
|
va = rb_entry_safe(node, struct vmap_area, rb_node);
|
|
|
|
return va ? va->subtree_max_size : 0;
|
|
|
|
}
|
mm: vmap area cache
Provide a free area cache for the vmalloc virtual address allocator, based
on the algorithm used by the user virtual memory allocator.
This reduces the number of rbtree operations and linear traversals over
the vmap extents in order to find a free area, by starting off at the last
point that a free area was found.
The free area cache is reset if areas are freed behind it, or if we are
searching for a smaller area or alignment than last time. So allocation
patterns are not changed (verified by corner-case and random test cases in
userspace testing).
This solves a regression caused by lazy vunmap TLB purging introduced in
db64fe02 (mm: rewrite vmap layer). That patch will leave extents in the
vmap allocator after they are vunmapped, and until a significant number
accumulate that can be flushed in a single batch. So in a workload that
vmalloc/vfree frequently, a chain of extents will build up from
VMALLOC_START address, which have to be iterated over each time (giving an
O(n) type of behaviour).
After this patch, the search will start from where it left off, giving
closer to an amortized O(1).
This is verified to solve regressions reported Steven in GFS2, and Avi in
KVM.
Hugh's update:
: I tried out the recent mmotm, and on one machine was fortunate to hit
: the BUG_ON(first->va_start < addr) which seems to have been stalling
: your vmap area cache patch ever since May.
: I can get you addresses etc, I did dump a few out; but once I stared
: at them, it was easier just to look at the code: and I cannot see how
: you would be so sure that first->va_start < addr, once you've done
: that addr = ALIGN(max(...), align) above, if align is over 0x1000
: (align was 0x8000 or 0x4000 in the cases I hit: ioremaps like Steve).
: I originally got around it by just changing the
: if (first->va_start < addr) {
: to
: while (first->va_start < addr) {
: without thinking about it any further; but that seemed unsatisfactory,
: why would we want to loop here when we've got another very similar
: loop just below it?
: I am never going to admit how long I've spent trying to grasp your
: "while (n)" rbtree loop just above this, the one with the peculiar
: if (!first && tmp->va_start < addr + size)
: in. That's unfamiliar to me, I'm guessing it's designed to save a
: subsequent rb_next() in a few circumstances (at risk of then setting
: a wrong cached_hole_size?); but they did appear few to me, and I didn't
: feel I could sign off something with that in when I don't grasp it,
: and it seems responsible for extra code and mistaken BUG_ON below it.
: I've reverted to the familiar rbtree loop that find_vma() does (but
: with va_end >= addr as you had, to respect the additional guard page):
: and then (given that cached_hole_size starts out 0) I don't see the
: need for any complications below it. If you do want to keep that loop
: as you had it, please add a comment to explain what it's trying to do,
: and where addr is relative to first when you emerge from it.
: Aren't your tests "size <= cached_hole_size" and
: "addr + size > first->va_start" forgetting the guard page we want
: before the next area? I've changed those.
: I have not changed your many "addr + size - 1 < addr" overflow tests,
: but have since come to wonder, shouldn't they be "addr + size < addr"
: tests - won't the vend checks go wrong if addr + size is 0?
: I have added a few comments - Wolfgang Wander's 2.6.13 description of
: 1363c3cd8603a913a27e2995dccbd70d5312d8e6 Avoiding mmap fragmentation
: helped me a lot, perhaps a pointer to that would be good too. And I found
: it easier to understand when I renamed cached_start slightly and moved the
: overflow label down.
: This patch would go after your mm-vmap-area-cache.patch in mmotm.
: Trivially, nobody is going to get that BUG_ON with this patch, and it
: appears to work fine on my machines; but I have not given it anything like
: the testing you did on your original, and may have broken all the
: performance you were aiming for. Please take a look and test it out
: integrate with yours if you're satisfied - thanks.
[akpm@linux-foundation.org: add locking comment]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Reported-and-tested-by: Steven Whitehouse <swhiteho@redhat.com>
Reported-and-tested-by: Avi Kivity <avi@redhat.com>
Tested-by: "Barry J. Marson" <bmarson@redhat.com>
Cc: Prarit Bhargava <prarit@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-03-23 07:30:36 +08:00
|
|
|
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
/*
|
|
|
|
* Gets called when remove the node and rotate.
|
|
|
|
*/
|
|
|
|
static __always_inline unsigned long
|
|
|
|
compute_subtree_max_size(struct vmap_area *va)
|
|
|
|
{
|
|
|
|
return max3(va_size(va),
|
|
|
|
get_subtree_max_size(va->rb_node.rb_left),
|
|
|
|
get_subtree_max_size(va->rb_node.rb_right));
|
|
|
|
}
|
|
|
|
|
2019-09-26 07:46:07 +08:00
|
|
|
RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb,
|
|
|
|
struct vmap_area, rb_node, unsigned long, subtree_max_size, va_size)
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
|
|
|
|
static void purge_vmap_area_lazy(void);
|
|
|
|
static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
|
|
|
|
static unsigned long lazy_max_pages(void);
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
|
2019-07-12 12:00:13 +08:00
|
|
|
static atomic_long_t nr_vmalloc_pages;
|
|
|
|
|
|
|
|
unsigned long vmalloc_nr_pages(void)
|
|
|
|
{
|
|
|
|
return atomic_long_read(&nr_vmalloc_pages);
|
|
|
|
}
|
|
|
|
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
static struct vmap_area *__find_vmap_area(unsigned long addr)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
struct rb_node *n = vmap_area_root.rb_node;
|
|
|
|
|
|
|
|
while (n) {
|
|
|
|
struct vmap_area *va;
|
|
|
|
|
|
|
|
va = rb_entry(n, struct vmap_area, rb_node);
|
|
|
|
if (addr < va->va_start)
|
|
|
|
n = n->rb_left;
|
2013-07-04 06:02:17 +08:00
|
|
|
else if (addr >= va->va_end)
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
n = n->rb_right;
|
|
|
|
else
|
|
|
|
return va;
|
|
|
|
}
|
|
|
|
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
/*
|
|
|
|
* This function returns back addresses of parent node
|
|
|
|
* and its left or right link for further processing.
|
|
|
|
*/
|
|
|
|
static __always_inline struct rb_node **
|
|
|
|
find_va_links(struct vmap_area *va,
|
|
|
|
struct rb_root *root, struct rb_node *from,
|
|
|
|
struct rb_node **parent)
|
|
|
|
{
|
|
|
|
struct vmap_area *tmp_va;
|
|
|
|
struct rb_node **link;
|
|
|
|
|
|
|
|
if (root) {
|
|
|
|
link = &root->rb_node;
|
|
|
|
if (unlikely(!*link)) {
|
|
|
|
*parent = NULL;
|
|
|
|
return link;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
link = &from;
|
|
|
|
}
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
/*
|
|
|
|
* Go to the bottom of the tree. When we hit the last point
|
|
|
|
* we end up with parent rb_node and correct direction, i name
|
|
|
|
* it link, where the new va->rb_node will be attached to.
|
|
|
|
*/
|
|
|
|
do {
|
|
|
|
tmp_va = rb_entry(*link, struct vmap_area, rb_node);
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
/*
|
|
|
|
* During the traversal we also do some sanity check.
|
|
|
|
* Trigger the BUG() if there are sides(left/right)
|
|
|
|
* or full overlaps.
|
|
|
|
*/
|
|
|
|
if (va->va_start < tmp_va->va_end &&
|
|
|
|
va->va_end <= tmp_va->va_start)
|
|
|
|
link = &(*link)->rb_left;
|
|
|
|
else if (va->va_end > tmp_va->va_start &&
|
|
|
|
va->va_start >= tmp_va->va_end)
|
|
|
|
link = &(*link)->rb_right;
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
else
|
|
|
|
BUG();
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
} while (*link);
|
|
|
|
|
|
|
|
*parent = &tmp_va->rb_node;
|
|
|
|
return link;
|
|
|
|
}
|
|
|
|
|
|
|
|
static __always_inline struct list_head *
|
|
|
|
get_va_next_sibling(struct rb_node *parent, struct rb_node **link)
|
|
|
|
{
|
|
|
|
struct list_head *list;
|
|
|
|
|
|
|
|
if (unlikely(!parent))
|
|
|
|
/*
|
|
|
|
* The red-black tree where we try to find VA neighbors
|
|
|
|
* before merging or inserting is empty, i.e. it means
|
|
|
|
* there is no free vmap space. Normally it does not
|
|
|
|
* happen but we handle this case anyway.
|
|
|
|
*/
|
|
|
|
return NULL;
|
|
|
|
|
|
|
|
list = &rb_entry(parent, struct vmap_area, rb_node)->list;
|
|
|
|
return (&parent->rb_right == link ? list->next : list);
|
|
|
|
}
|
|
|
|
|
|
|
|
static __always_inline void
|
|
|
|
link_va(struct vmap_area *va, struct rb_root *root,
|
|
|
|
struct rb_node *parent, struct rb_node **link, struct list_head *head)
|
|
|
|
{
|
|
|
|
/*
|
|
|
|
* VA is still not in the list, but we can
|
|
|
|
* identify its future previous list_head node.
|
|
|
|
*/
|
|
|
|
if (likely(parent)) {
|
|
|
|
head = &rb_entry(parent, struct vmap_area, rb_node)->list;
|
|
|
|
if (&parent->rb_right != link)
|
|
|
|
head = head->prev;
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
}
|
|
|
|
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
/* Insert to the rb-tree */
|
|
|
|
rb_link_node(&va->rb_node, parent, link);
|
|
|
|
if (root == &free_vmap_area_root) {
|
|
|
|
/*
|
|
|
|
* Some explanation here. Just perform simple insertion
|
|
|
|
* to the tree. We do not set va->subtree_max_size to
|
|
|
|
* its current size before calling rb_insert_augmented().
|
|
|
|
* It is because of we populate the tree from the bottom
|
|
|
|
* to parent levels when the node _is_ in the tree.
|
|
|
|
*
|
|
|
|
* Therefore we set subtree_max_size to zero after insertion,
|
|
|
|
* to let __augment_tree_propagate_from() puts everything to
|
|
|
|
* the correct order later on.
|
|
|
|
*/
|
|
|
|
rb_insert_augmented(&va->rb_node,
|
|
|
|
root, &free_vmap_area_rb_augment_cb);
|
|
|
|
va->subtree_max_size = 0;
|
|
|
|
} else {
|
|
|
|
rb_insert_color(&va->rb_node, root);
|
|
|
|
}
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
/* Address-sort this list */
|
|
|
|
list_add(&va->list, head);
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
}
|
|
|
|
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
static __always_inline void
|
|
|
|
unlink_va(struct vmap_area *va, struct rb_root *root)
|
|
|
|
{
|
2019-07-12 11:59:03 +08:00
|
|
|
if (WARN_ON(RB_EMPTY_NODE(&va->rb_node)))
|
|
|
|
return;
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
|
2019-07-12 11:59:03 +08:00
|
|
|
if (root == &free_vmap_area_root)
|
|
|
|
rb_erase_augmented(&va->rb_node,
|
|
|
|
root, &free_vmap_area_rb_augment_cb);
|
|
|
|
else
|
|
|
|
rb_erase(&va->rb_node, root);
|
|
|
|
|
|
|
|
list_del(&va->list);
|
|
|
|
RB_CLEAR_NODE(&va->rb_node);
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
}
|
|
|
|
|
2019-05-18 05:31:34 +08:00
|
|
|
#if DEBUG_AUGMENT_PROPAGATE_CHECK
|
|
|
|
static void
|
|
|
|
augment_tree_propagate_check(struct rb_node *n)
|
|
|
|
{
|
|
|
|
struct vmap_area *va;
|
|
|
|
struct rb_node *node;
|
|
|
|
unsigned long size;
|
|
|
|
bool found = false;
|
|
|
|
|
|
|
|
if (n == NULL)
|
|
|
|
return;
|
|
|
|
|
|
|
|
va = rb_entry(n, struct vmap_area, rb_node);
|
|
|
|
size = va->subtree_max_size;
|
|
|
|
node = n;
|
|
|
|
|
|
|
|
while (node) {
|
|
|
|
va = rb_entry(node, struct vmap_area, rb_node);
|
|
|
|
|
|
|
|
if (get_subtree_max_size(node->rb_left) == size) {
|
|
|
|
node = node->rb_left;
|
|
|
|
} else {
|
|
|
|
if (va_size(va) == size) {
|
|
|
|
found = true;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
node = node->rb_right;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!found) {
|
|
|
|
va = rb_entry(n, struct vmap_area, rb_node);
|
|
|
|
pr_emerg("tree is corrupted: %lu, %lu\n",
|
|
|
|
va_size(va), va->subtree_max_size);
|
|
|
|
}
|
|
|
|
|
|
|
|
augment_tree_propagate_check(n->rb_left);
|
|
|
|
augment_tree_propagate_check(n->rb_right);
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
/*
|
|
|
|
* This function populates subtree_max_size from bottom to upper
|
|
|
|
* levels starting from VA point. The propagation must be done
|
|
|
|
* when VA size is modified by changing its va_start/va_end. Or
|
|
|
|
* in case of newly inserting of VA to the tree.
|
|
|
|
*
|
|
|
|
* It means that __augment_tree_propagate_from() must be called:
|
|
|
|
* - After VA has been inserted to the tree(free path);
|
|
|
|
* - After VA has been shrunk(allocation path);
|
|
|
|
* - After VA has been increased(merging path).
|
|
|
|
*
|
|
|
|
* Please note that, it does not mean that upper parent nodes
|
|
|
|
* and their subtree_max_size are recalculated all the time up
|
|
|
|
* to the root node.
|
|
|
|
*
|
|
|
|
* 4--8
|
|
|
|
* /\
|
|
|
|
* / \
|
|
|
|
* / \
|
|
|
|
* 2--2 8--8
|
|
|
|
*
|
|
|
|
* For example if we modify the node 4, shrinking it to 2, then
|
|
|
|
* no any modification is required. If we shrink the node 2 to 1
|
|
|
|
* its subtree_max_size is updated only, and set to 1. If we shrink
|
|
|
|
* the node 8 to 6, then its subtree_max_size is set to 6 and parent
|
|
|
|
* node becomes 4--6.
|
|
|
|
*/
|
|
|
|
static __always_inline void
|
|
|
|
augment_tree_propagate_from(struct vmap_area *va)
|
|
|
|
{
|
|
|
|
struct rb_node *node = &va->rb_node;
|
|
|
|
unsigned long new_va_sub_max_size;
|
|
|
|
|
|
|
|
while (node) {
|
|
|
|
va = rb_entry(node, struct vmap_area, rb_node);
|
|
|
|
new_va_sub_max_size = compute_subtree_max_size(va);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* If the newly calculated maximum available size of the
|
|
|
|
* subtree is equal to the current one, then it means that
|
|
|
|
* the tree is propagated correctly. So we have to stop at
|
|
|
|
* this point to save cycles.
|
|
|
|
*/
|
|
|
|
if (va->subtree_max_size == new_va_sub_max_size)
|
|
|
|
break;
|
|
|
|
|
|
|
|
va->subtree_max_size = new_va_sub_max_size;
|
|
|
|
node = rb_parent(&va->rb_node);
|
|
|
|
}
|
2019-05-18 05:31:34 +08:00
|
|
|
|
|
|
|
#if DEBUG_AUGMENT_PROPAGATE_CHECK
|
|
|
|
augment_tree_propagate_check(free_vmap_area_root.rb_node);
|
|
|
|
#endif
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
static void
|
|
|
|
insert_vmap_area(struct vmap_area *va,
|
|
|
|
struct rb_root *root, struct list_head *head)
|
|
|
|
{
|
|
|
|
struct rb_node **link;
|
|
|
|
struct rb_node *parent;
|
|
|
|
|
|
|
|
link = find_va_links(va, root, NULL, &parent);
|
|
|
|
link_va(va, root, parent, link, head);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void
|
|
|
|
insert_vmap_area_augment(struct vmap_area *va,
|
|
|
|
struct rb_node *from, struct rb_root *root,
|
|
|
|
struct list_head *head)
|
|
|
|
{
|
|
|
|
struct rb_node **link;
|
|
|
|
struct rb_node *parent;
|
|
|
|
|
|
|
|
if (from)
|
|
|
|
link = find_va_links(va, NULL, from, &parent);
|
|
|
|
else
|
|
|
|
link = find_va_links(va, root, NULL, &parent);
|
|
|
|
|
|
|
|
link_va(va, root, parent, link, head);
|
|
|
|
augment_tree_propagate_from(va);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Merge de-allocated chunk of VA memory with previous
|
|
|
|
* and next free blocks. If coalesce is not done a new
|
|
|
|
* free area is inserted. If VA has been merged, it is
|
|
|
|
* freed.
|
|
|
|
*/
|
kasan: support backing vmalloc space with real shadow memory
Patch series "kasan: support backing vmalloc space with real shadow
memory", v11.
Currently, vmalloc space is backed by the early shadow page. This means
that kasan is incompatible with VMAP_STACK.
This series provides a mechanism to back vmalloc space with real,
dynamically allocated memory. I have only wired up x86, because that's
the only currently supported arch I can work with easily, but it's very
easy to wire up other architectures, and it appears that there is some
work-in-progress code to do this on arm64 and s390.
This has been discussed before in the context of VMAP_STACK:
- https://bugzilla.kernel.org/show_bug.cgi?id=202009
- https://lkml.org/lkml/2018/7/22/198
- https://lkml.org/lkml/2019/7/19/822
In terms of implementation details:
Most mappings in vmalloc space are small, requiring less than a full
page of shadow space. Allocating a full shadow page per mapping would
therefore be wasteful. Furthermore, to ensure that different mappings
use different shadow pages, mappings would have to be aligned to
KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE.
Instead, share backing space across multiple mappings. Allocate a
backing page when a mapping in vmalloc space uses a particular page of
the shadow region. This page can be shared by other vmalloc mappings
later on.
We hook in to the vmap infrastructure to lazily clean up unused shadow
memory.
Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that:
- Turning on KASAN, inline instrumentation, without vmalloc, introuduces
a 4.1x-4.2x slowdown in vmalloc operations.
- Turning this on introduces the following slowdowns over KASAN:
* ~1.76x slower single-threaded (test_vmalloc.sh performance)
* ~2.18x slower when both cpus are performing operations
simultaneously (test_vmalloc.sh sequential_test_order=1)
This is unfortunate but given that this is a debug feature only, not the
end of the world. The benchmarks are also a stress-test for the vmalloc
subsystem: they're not indicative of an overall 2x slowdown!
This patch (of 4):
Hook into vmalloc and vmap, and dynamically allocate real shadow memory
to back the mappings.
Most mappings in vmalloc space are small, requiring less than a full
page of shadow space. Allocating a full shadow page per mapping would
therefore be wasteful. Furthermore, to ensure that different mappings
use different shadow pages, mappings would have to be aligned to
KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE.
Instead, share backing space across multiple mappings. Allocate a
backing page when a mapping in vmalloc space uses a particular page of
the shadow region. This page can be shared by other vmalloc mappings
later on.
We hook in to the vmap infrastructure to lazily clean up unused shadow
memory.
To avoid the difficulties around swapping mappings around, this code
expects that the part of the shadow region that covers the vmalloc space
will not be covered by the early shadow page, but will be left unmapped.
This will require changes in arch-specific code.
This allows KASAN with VMAP_STACK, and may be helpful for architectures
that do not have a separate module space (e.g. powerpc64, which I am
currently working on). It also allows relaxing the module alignment
back to PAGE_SIZE.
Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that:
- Turning on KASAN, inline instrumentation, without vmalloc, introuduces
a 4.1x-4.2x slowdown in vmalloc operations.
- Turning this on introduces the following slowdowns over KASAN:
* ~1.76x slower single-threaded (test_vmalloc.sh performance)
* ~2.18x slower when both cpus are performing operations
simultaneously (test_vmalloc.sh sequential_test_order=3D1)
This is unfortunate but given that this is a debug feature only, not the
end of the world.
The full benchmark results are:
Performance
No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN
fix_size_alloc_test 662004 11404956 17.23 19144610 28.92 1.68
full_fit_alloc_test 710950 12029752 16.92 13184651 18.55 1.10
long_busy_list_alloc_test 9431875 43990172 4.66 82970178 8.80 1.89
random_size_alloc_test 5033626 23061762 4.58 47158834 9.37 2.04
fix_align_alloc_test 1252514 15276910 12.20 31266116 24.96 2.05
random_size_align_alloc_te 1648501 14578321 8.84 25560052 15.51 1.75
align_shift_alloc_test 147 830 5.65 5692 38.72 6.86
pcpu_alloc_test 80732 125520 1.55 140864 1.74 1.12
Total Cycles 119240774314 763211341128 6.40 1390338696894 11.66 1.82
Sequential, 2 cpus
No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN
fix_size_alloc_test 1423150 14276550 10.03 27733022 19.49 1.94
full_fit_alloc_test 1754219 14722640 8.39 15030786 8.57 1.02
long_busy_list_alloc_test 11451858 52154973 4.55 107016027 9.34 2.05
random_size_alloc_test 5989020 26735276 4.46 68885923 11.50 2.58
fix_align_alloc_test 2050976 20166900 9.83 50491675 24.62 2.50
random_size_align_alloc_te 2858229 17971700 6.29 38730225 13.55 2.16
align_shift_alloc_test 405 6428 15.87 26253 64.82 4.08
pcpu_alloc_test 127183 151464 1.19 216263 1.70 1.43
Total Cycles 54181269392 308723699764 5.70 650772566394 12.01 2.11
fix_size_alloc_test 1420404 14289308 10.06 27790035 19.56 1.94
full_fit_alloc_test 1736145 14806234 8.53 15274301 8.80 1.03
long_busy_list_alloc_test 11404638 52270785 4.58 107550254 9.43 2.06
random_size_alloc_test 6017006 26650625 4.43 68696127 11.42 2.58
fix_align_alloc_test 2045504 20280985 9.91 50414862 24.65 2.49
random_size_align_alloc_te 2845338 17931018 6.30 38510276 13.53 2.15
align_shift_alloc_test 472 3760 7.97 9656 20.46 2.57
pcpu_alloc_test 118643 132732 1.12 146504 1.23 1.10
Total Cycles 54040011688 309102805492 5.72 651325675652 12.05 2.11
[dja@axtens.net: fixups]
Link: http://lkml.kernel.org/r/20191120052719.7201-1-dja@axtens.net
Link: https://bugzilla.kernel.org/show_bug.cgi?id=3D202009
Link: http://lkml.kernel.org/r/20191031093909.9228-2-dja@axtens.net
Signed-off-by: Mark Rutland <mark.rutland@arm.com> [shadow rework]
Signed-off-by: Daniel Axtens <dja@axtens.net>
Co-developed-by: Mark Rutland <mark.rutland@arm.com>
Acked-by: Vasily Gorbik <gor@linux.ibm.com>
Reviewed-by: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Christophe Leroy <christophe.leroy@c-s.fr>
Cc: Qian Cai <cai@lca.pw>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:50 +08:00
|
|
|
static __always_inline struct vmap_area *
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
merge_or_add_vmap_area(struct vmap_area *va,
|
|
|
|
struct rb_root *root, struct list_head *head)
|
|
|
|
{
|
|
|
|
struct vmap_area *sibling;
|
|
|
|
struct list_head *next;
|
|
|
|
struct rb_node **link;
|
|
|
|
struct rb_node *parent;
|
|
|
|
bool merged = false;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Find a place in the tree where VA potentially will be
|
|
|
|
* inserted, unless it is merged with its sibling/siblings.
|
|
|
|
*/
|
|
|
|
link = find_va_links(va, root, NULL, &parent);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Get next node of VA to check if merging can be done.
|
|
|
|
*/
|
|
|
|
next = get_va_next_sibling(parent, link);
|
|
|
|
if (unlikely(next == NULL))
|
|
|
|
goto insert;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* start end
|
|
|
|
* | |
|
|
|
|
* |<------VA------>|<-----Next----->|
|
|
|
|
* | |
|
|
|
|
* start end
|
|
|
|
*/
|
|
|
|
if (next != head) {
|
|
|
|
sibling = list_entry(next, struct vmap_area, list);
|
|
|
|
if (sibling->va_start == va->va_end) {
|
|
|
|
sibling->va_start = va->va_start;
|
|
|
|
|
|
|
|
/* Check and update the tree if needed. */
|
|
|
|
augment_tree_propagate_from(sibling);
|
|
|
|
|
|
|
|
/* Free vmap_area object. */
|
|
|
|
kmem_cache_free(vmap_area_cachep, va);
|
|
|
|
|
|
|
|
/* Point to the new merged area. */
|
|
|
|
va = sibling;
|
|
|
|
merged = true;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* start end
|
|
|
|
* | |
|
|
|
|
* |<-----Prev----->|<------VA------>|
|
|
|
|
* | |
|
|
|
|
* start end
|
|
|
|
*/
|
|
|
|
if (next->prev != head) {
|
|
|
|
sibling = list_entry(next->prev, struct vmap_area, list);
|
|
|
|
if (sibling->va_end == va->va_start) {
|
|
|
|
sibling->va_end = va->va_end;
|
|
|
|
|
|
|
|
/* Check and update the tree if needed. */
|
|
|
|
augment_tree_propagate_from(sibling);
|
|
|
|
|
2019-07-12 11:59:00 +08:00
|
|
|
if (merged)
|
|
|
|
unlink_va(va, root);
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
|
|
|
|
/* Free vmap_area object. */
|
|
|
|
kmem_cache_free(vmap_area_cachep, va);
|
kasan: support backing vmalloc space with real shadow memory
Patch series "kasan: support backing vmalloc space with real shadow
memory", v11.
Currently, vmalloc space is backed by the early shadow page. This means
that kasan is incompatible with VMAP_STACK.
This series provides a mechanism to back vmalloc space with real,
dynamically allocated memory. I have only wired up x86, because that's
the only currently supported arch I can work with easily, but it's very
easy to wire up other architectures, and it appears that there is some
work-in-progress code to do this on arm64 and s390.
This has been discussed before in the context of VMAP_STACK:
- https://bugzilla.kernel.org/show_bug.cgi?id=202009
- https://lkml.org/lkml/2018/7/22/198
- https://lkml.org/lkml/2019/7/19/822
In terms of implementation details:
Most mappings in vmalloc space are small, requiring less than a full
page of shadow space. Allocating a full shadow page per mapping would
therefore be wasteful. Furthermore, to ensure that different mappings
use different shadow pages, mappings would have to be aligned to
KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE.
Instead, share backing space across multiple mappings. Allocate a
backing page when a mapping in vmalloc space uses a particular page of
the shadow region. This page can be shared by other vmalloc mappings
later on.
We hook in to the vmap infrastructure to lazily clean up unused shadow
memory.
Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that:
- Turning on KASAN, inline instrumentation, without vmalloc, introuduces
a 4.1x-4.2x slowdown in vmalloc operations.
- Turning this on introduces the following slowdowns over KASAN:
* ~1.76x slower single-threaded (test_vmalloc.sh performance)
* ~2.18x slower when both cpus are performing operations
simultaneously (test_vmalloc.sh sequential_test_order=1)
This is unfortunate but given that this is a debug feature only, not the
end of the world. The benchmarks are also a stress-test for the vmalloc
subsystem: they're not indicative of an overall 2x slowdown!
This patch (of 4):
Hook into vmalloc and vmap, and dynamically allocate real shadow memory
to back the mappings.
Most mappings in vmalloc space are small, requiring less than a full
page of shadow space. Allocating a full shadow page per mapping would
therefore be wasteful. Furthermore, to ensure that different mappings
use different shadow pages, mappings would have to be aligned to
KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE.
Instead, share backing space across multiple mappings. Allocate a
backing page when a mapping in vmalloc space uses a particular page of
the shadow region. This page can be shared by other vmalloc mappings
later on.
We hook in to the vmap infrastructure to lazily clean up unused shadow
memory.
To avoid the difficulties around swapping mappings around, this code
expects that the part of the shadow region that covers the vmalloc space
will not be covered by the early shadow page, but will be left unmapped.
This will require changes in arch-specific code.
This allows KASAN with VMAP_STACK, and may be helpful for architectures
that do not have a separate module space (e.g. powerpc64, which I am
currently working on). It also allows relaxing the module alignment
back to PAGE_SIZE.
Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that:
- Turning on KASAN, inline instrumentation, without vmalloc, introuduces
a 4.1x-4.2x slowdown in vmalloc operations.
- Turning this on introduces the following slowdowns over KASAN:
* ~1.76x slower single-threaded (test_vmalloc.sh performance)
* ~2.18x slower when both cpus are performing operations
simultaneously (test_vmalloc.sh sequential_test_order=3D1)
This is unfortunate but given that this is a debug feature only, not the
end of the world.
The full benchmark results are:
Performance
No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN
fix_size_alloc_test 662004 11404956 17.23 19144610 28.92 1.68
full_fit_alloc_test 710950 12029752 16.92 13184651 18.55 1.10
long_busy_list_alloc_test 9431875 43990172 4.66 82970178 8.80 1.89
random_size_alloc_test 5033626 23061762 4.58 47158834 9.37 2.04
fix_align_alloc_test 1252514 15276910 12.20 31266116 24.96 2.05
random_size_align_alloc_te 1648501 14578321 8.84 25560052 15.51 1.75
align_shift_alloc_test 147 830 5.65 5692 38.72 6.86
pcpu_alloc_test 80732 125520 1.55 140864 1.74 1.12
Total Cycles 119240774314 763211341128 6.40 1390338696894 11.66 1.82
Sequential, 2 cpus
No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN
fix_size_alloc_test 1423150 14276550 10.03 27733022 19.49 1.94
full_fit_alloc_test 1754219 14722640 8.39 15030786 8.57 1.02
long_busy_list_alloc_test 11451858 52154973 4.55 107016027 9.34 2.05
random_size_alloc_test 5989020 26735276 4.46 68885923 11.50 2.58
fix_align_alloc_test 2050976 20166900 9.83 50491675 24.62 2.50
random_size_align_alloc_te 2858229 17971700 6.29 38730225 13.55 2.16
align_shift_alloc_test 405 6428 15.87 26253 64.82 4.08
pcpu_alloc_test 127183 151464 1.19 216263 1.70 1.43
Total Cycles 54181269392 308723699764 5.70 650772566394 12.01 2.11
fix_size_alloc_test 1420404 14289308 10.06 27790035 19.56 1.94
full_fit_alloc_test 1736145 14806234 8.53 15274301 8.80 1.03
long_busy_list_alloc_test 11404638 52270785 4.58 107550254 9.43 2.06
random_size_alloc_test 6017006 26650625 4.43 68696127 11.42 2.58
fix_align_alloc_test 2045504 20280985 9.91 50414862 24.65 2.49
random_size_align_alloc_te 2845338 17931018 6.30 38510276 13.53 2.15
align_shift_alloc_test 472 3760 7.97 9656 20.46 2.57
pcpu_alloc_test 118643 132732 1.12 146504 1.23 1.10
Total Cycles 54040011688 309102805492 5.72 651325675652 12.05 2.11
[dja@axtens.net: fixups]
Link: http://lkml.kernel.org/r/20191120052719.7201-1-dja@axtens.net
Link: https://bugzilla.kernel.org/show_bug.cgi?id=3D202009
Link: http://lkml.kernel.org/r/20191031093909.9228-2-dja@axtens.net
Signed-off-by: Mark Rutland <mark.rutland@arm.com> [shadow rework]
Signed-off-by: Daniel Axtens <dja@axtens.net>
Co-developed-by: Mark Rutland <mark.rutland@arm.com>
Acked-by: Vasily Gorbik <gor@linux.ibm.com>
Reviewed-by: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Christophe Leroy <christophe.leroy@c-s.fr>
Cc: Qian Cai <cai@lca.pw>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:50 +08:00
|
|
|
|
|
|
|
/* Point to the new merged area. */
|
|
|
|
va = sibling;
|
|
|
|
merged = true;
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
insert:
|
|
|
|
if (!merged) {
|
|
|
|
link_va(va, root, parent, link, head);
|
|
|
|
augment_tree_propagate_from(va);
|
|
|
|
}
|
kasan: support backing vmalloc space with real shadow memory
Patch series "kasan: support backing vmalloc space with real shadow
memory", v11.
Currently, vmalloc space is backed by the early shadow page. This means
that kasan is incompatible with VMAP_STACK.
This series provides a mechanism to back vmalloc space with real,
dynamically allocated memory. I have only wired up x86, because that's
the only currently supported arch I can work with easily, but it's very
easy to wire up other architectures, and it appears that there is some
work-in-progress code to do this on arm64 and s390.
This has been discussed before in the context of VMAP_STACK:
- https://bugzilla.kernel.org/show_bug.cgi?id=202009
- https://lkml.org/lkml/2018/7/22/198
- https://lkml.org/lkml/2019/7/19/822
In terms of implementation details:
Most mappings in vmalloc space are small, requiring less than a full
page of shadow space. Allocating a full shadow page per mapping would
therefore be wasteful. Furthermore, to ensure that different mappings
use different shadow pages, mappings would have to be aligned to
KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE.
Instead, share backing space across multiple mappings. Allocate a
backing page when a mapping in vmalloc space uses a particular page of
the shadow region. This page can be shared by other vmalloc mappings
later on.
We hook in to the vmap infrastructure to lazily clean up unused shadow
memory.
Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that:
- Turning on KASAN, inline instrumentation, without vmalloc, introuduces
a 4.1x-4.2x slowdown in vmalloc operations.
- Turning this on introduces the following slowdowns over KASAN:
* ~1.76x slower single-threaded (test_vmalloc.sh performance)
* ~2.18x slower when both cpus are performing operations
simultaneously (test_vmalloc.sh sequential_test_order=1)
This is unfortunate but given that this is a debug feature only, not the
end of the world. The benchmarks are also a stress-test for the vmalloc
subsystem: they're not indicative of an overall 2x slowdown!
This patch (of 4):
Hook into vmalloc and vmap, and dynamically allocate real shadow memory
to back the mappings.
Most mappings in vmalloc space are small, requiring less than a full
page of shadow space. Allocating a full shadow page per mapping would
therefore be wasteful. Furthermore, to ensure that different mappings
use different shadow pages, mappings would have to be aligned to
KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE.
Instead, share backing space across multiple mappings. Allocate a
backing page when a mapping in vmalloc space uses a particular page of
the shadow region. This page can be shared by other vmalloc mappings
later on.
We hook in to the vmap infrastructure to lazily clean up unused shadow
memory.
To avoid the difficulties around swapping mappings around, this code
expects that the part of the shadow region that covers the vmalloc space
will not be covered by the early shadow page, but will be left unmapped.
This will require changes in arch-specific code.
This allows KASAN with VMAP_STACK, and may be helpful for architectures
that do not have a separate module space (e.g. powerpc64, which I am
currently working on). It also allows relaxing the module alignment
back to PAGE_SIZE.
Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that:
- Turning on KASAN, inline instrumentation, without vmalloc, introuduces
a 4.1x-4.2x slowdown in vmalloc operations.
- Turning this on introduces the following slowdowns over KASAN:
* ~1.76x slower single-threaded (test_vmalloc.sh performance)
* ~2.18x slower when both cpus are performing operations
simultaneously (test_vmalloc.sh sequential_test_order=3D1)
This is unfortunate but given that this is a debug feature only, not the
end of the world.
The full benchmark results are:
Performance
No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN
fix_size_alloc_test 662004 11404956 17.23 19144610 28.92 1.68
full_fit_alloc_test 710950 12029752 16.92 13184651 18.55 1.10
long_busy_list_alloc_test 9431875 43990172 4.66 82970178 8.80 1.89
random_size_alloc_test 5033626 23061762 4.58 47158834 9.37 2.04
fix_align_alloc_test 1252514 15276910 12.20 31266116 24.96 2.05
random_size_align_alloc_te 1648501 14578321 8.84 25560052 15.51 1.75
align_shift_alloc_test 147 830 5.65 5692 38.72 6.86
pcpu_alloc_test 80732 125520 1.55 140864 1.74 1.12
Total Cycles 119240774314 763211341128 6.40 1390338696894 11.66 1.82
Sequential, 2 cpus
No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN
fix_size_alloc_test 1423150 14276550 10.03 27733022 19.49 1.94
full_fit_alloc_test 1754219 14722640 8.39 15030786 8.57 1.02
long_busy_list_alloc_test 11451858 52154973 4.55 107016027 9.34 2.05
random_size_alloc_test 5989020 26735276 4.46 68885923 11.50 2.58
fix_align_alloc_test 2050976 20166900 9.83 50491675 24.62 2.50
random_size_align_alloc_te 2858229 17971700 6.29 38730225 13.55 2.16
align_shift_alloc_test 405 6428 15.87 26253 64.82 4.08
pcpu_alloc_test 127183 151464 1.19 216263 1.70 1.43
Total Cycles 54181269392 308723699764 5.70 650772566394 12.01 2.11
fix_size_alloc_test 1420404 14289308 10.06 27790035 19.56 1.94
full_fit_alloc_test 1736145 14806234 8.53 15274301 8.80 1.03
long_busy_list_alloc_test 11404638 52270785 4.58 107550254 9.43 2.06
random_size_alloc_test 6017006 26650625 4.43 68696127 11.42 2.58
fix_align_alloc_test 2045504 20280985 9.91 50414862 24.65 2.49
random_size_align_alloc_te 2845338 17931018 6.30 38510276 13.53 2.15
align_shift_alloc_test 472 3760 7.97 9656 20.46 2.57
pcpu_alloc_test 118643 132732 1.12 146504 1.23 1.10
Total Cycles 54040011688 309102805492 5.72 651325675652 12.05 2.11
[dja@axtens.net: fixups]
Link: http://lkml.kernel.org/r/20191120052719.7201-1-dja@axtens.net
Link: https://bugzilla.kernel.org/show_bug.cgi?id=3D202009
Link: http://lkml.kernel.org/r/20191031093909.9228-2-dja@axtens.net
Signed-off-by: Mark Rutland <mark.rutland@arm.com> [shadow rework]
Signed-off-by: Daniel Axtens <dja@axtens.net>
Co-developed-by: Mark Rutland <mark.rutland@arm.com>
Acked-by: Vasily Gorbik <gor@linux.ibm.com>
Reviewed-by: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Christophe Leroy <christophe.leroy@c-s.fr>
Cc: Qian Cai <cai@lca.pw>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:50 +08:00
|
|
|
|
|
|
|
return va;
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
static __always_inline bool
|
|
|
|
is_within_this_va(struct vmap_area *va, unsigned long size,
|
|
|
|
unsigned long align, unsigned long vstart)
|
|
|
|
{
|
|
|
|
unsigned long nva_start_addr;
|
|
|
|
|
|
|
|
if (va->va_start > vstart)
|
|
|
|
nva_start_addr = ALIGN(va->va_start, align);
|
|
|
|
else
|
|
|
|
nva_start_addr = ALIGN(vstart, align);
|
|
|
|
|
|
|
|
/* Can be overflowed due to big size or alignment. */
|
|
|
|
if (nva_start_addr + size < nva_start_addr ||
|
|
|
|
nva_start_addr < vstart)
|
|
|
|
return false;
|
|
|
|
|
|
|
|
return (nva_start_addr + size <= va->va_end);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Find the first free block(lowest start address) in the tree,
|
|
|
|
* that will accomplish the request corresponding to passing
|
|
|
|
* parameters.
|
|
|
|
*/
|
|
|
|
static __always_inline struct vmap_area *
|
|
|
|
find_vmap_lowest_match(unsigned long size,
|
|
|
|
unsigned long align, unsigned long vstart)
|
|
|
|
{
|
|
|
|
struct vmap_area *va;
|
|
|
|
struct rb_node *node;
|
|
|
|
unsigned long length;
|
|
|
|
|
|
|
|
/* Start from the root. */
|
|
|
|
node = free_vmap_area_root.rb_node;
|
|
|
|
|
|
|
|
/* Adjust the search size for alignment overhead. */
|
|
|
|
length = size + align - 1;
|
|
|
|
|
|
|
|
while (node) {
|
|
|
|
va = rb_entry(node, struct vmap_area, rb_node);
|
|
|
|
|
|
|
|
if (get_subtree_max_size(node->rb_left) >= length &&
|
|
|
|
vstart < va->va_start) {
|
|
|
|
node = node->rb_left;
|
|
|
|
} else {
|
|
|
|
if (is_within_this_va(va, size, align, vstart))
|
|
|
|
return va;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Does not make sense to go deeper towards the right
|
|
|
|
* sub-tree if it does not have a free block that is
|
|
|
|
* equal or bigger to the requested search length.
|
|
|
|
*/
|
|
|
|
if (get_subtree_max_size(node->rb_right) >= length) {
|
|
|
|
node = node->rb_right;
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
2019-06-01 13:30:03 +08:00
|
|
|
* OK. We roll back and find the first right sub-tree,
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
* that will satisfy the search criteria. It can happen
|
|
|
|
* only once due to "vstart" restriction.
|
|
|
|
*/
|
|
|
|
while ((node = rb_parent(node))) {
|
|
|
|
va = rb_entry(node, struct vmap_area, rb_node);
|
|
|
|
if (is_within_this_va(va, size, align, vstart))
|
|
|
|
return va;
|
|
|
|
|
|
|
|
if (get_subtree_max_size(node->rb_right) >= length &&
|
|
|
|
vstart <= va->va_start) {
|
|
|
|
node = node->rb_right;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
2019-05-18 05:31:37 +08:00
|
|
|
#if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
|
|
|
|
#include <linux/random.h>
|
|
|
|
|
|
|
|
static struct vmap_area *
|
|
|
|
find_vmap_lowest_linear_match(unsigned long size,
|
|
|
|
unsigned long align, unsigned long vstart)
|
|
|
|
{
|
|
|
|
struct vmap_area *va;
|
|
|
|
|
|
|
|
list_for_each_entry(va, &free_vmap_area_list, list) {
|
|
|
|
if (!is_within_this_va(va, size, align, vstart))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
return va;
|
|
|
|
}
|
|
|
|
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void
|
|
|
|
find_vmap_lowest_match_check(unsigned long size)
|
|
|
|
{
|
|
|
|
struct vmap_area *va_1, *va_2;
|
|
|
|
unsigned long vstart;
|
|
|
|
unsigned int rnd;
|
|
|
|
|
|
|
|
get_random_bytes(&rnd, sizeof(rnd));
|
|
|
|
vstart = VMALLOC_START + rnd;
|
|
|
|
|
|
|
|
va_1 = find_vmap_lowest_match(size, 1, vstart);
|
|
|
|
va_2 = find_vmap_lowest_linear_match(size, 1, vstart);
|
|
|
|
|
|
|
|
if (va_1 != va_2)
|
|
|
|
pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
|
|
|
|
va_1, va_2, vstart);
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
enum fit_type {
|
|
|
|
NOTHING_FIT = 0,
|
|
|
|
FL_FIT_TYPE = 1, /* full fit */
|
|
|
|
LE_FIT_TYPE = 2, /* left edge fit */
|
|
|
|
RE_FIT_TYPE = 3, /* right edge fit */
|
|
|
|
NE_FIT_TYPE = 4 /* no edge fit */
|
|
|
|
};
|
|
|
|
|
|
|
|
static __always_inline enum fit_type
|
|
|
|
classify_va_fit_type(struct vmap_area *va,
|
|
|
|
unsigned long nva_start_addr, unsigned long size)
|
|
|
|
{
|
|
|
|
enum fit_type type;
|
|
|
|
|
|
|
|
/* Check if it is within VA. */
|
|
|
|
if (nva_start_addr < va->va_start ||
|
|
|
|
nva_start_addr + size > va->va_end)
|
|
|
|
return NOTHING_FIT;
|
|
|
|
|
|
|
|
/* Now classify. */
|
|
|
|
if (va->va_start == nva_start_addr) {
|
|
|
|
if (va->va_end == nva_start_addr + size)
|
|
|
|
type = FL_FIT_TYPE;
|
|
|
|
else
|
|
|
|
type = LE_FIT_TYPE;
|
|
|
|
} else if (va->va_end == nva_start_addr + size) {
|
|
|
|
type = RE_FIT_TYPE;
|
|
|
|
} else {
|
|
|
|
type = NE_FIT_TYPE;
|
|
|
|
}
|
|
|
|
|
|
|
|
return type;
|
|
|
|
}
|
|
|
|
|
|
|
|
static __always_inline int
|
|
|
|
adjust_va_to_fit_type(struct vmap_area *va,
|
|
|
|
unsigned long nva_start_addr, unsigned long size,
|
|
|
|
enum fit_type type)
|
|
|
|
{
|
2019-06-29 03:07:09 +08:00
|
|
|
struct vmap_area *lva = NULL;
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
|
|
|
|
if (type == FL_FIT_TYPE) {
|
|
|
|
/*
|
|
|
|
* No need to split VA, it fully fits.
|
|
|
|
*
|
|
|
|
* | |
|
|
|
|
* V NVA V
|
|
|
|
* |---------------|
|
|
|
|
*/
|
|
|
|
unlink_va(va, &free_vmap_area_root);
|
|
|
|
kmem_cache_free(vmap_area_cachep, va);
|
|
|
|
} else if (type == LE_FIT_TYPE) {
|
|
|
|
/*
|
|
|
|
* Split left edge of fit VA.
|
|
|
|
*
|
|
|
|
* | |
|
|
|
|
* V NVA V R
|
|
|
|
* |-------|-------|
|
|
|
|
*/
|
|
|
|
va->va_start += size;
|
|
|
|
} else if (type == RE_FIT_TYPE) {
|
|
|
|
/*
|
|
|
|
* Split right edge of fit VA.
|
|
|
|
*
|
|
|
|
* | |
|
|
|
|
* L V NVA V
|
|
|
|
* |-------|-------|
|
|
|
|
*/
|
|
|
|
va->va_end = nva_start_addr;
|
|
|
|
} else if (type == NE_FIT_TYPE) {
|
|
|
|
/*
|
|
|
|
* Split no edge of fit VA.
|
|
|
|
*
|
|
|
|
* | |
|
|
|
|
* L V NVA V R
|
|
|
|
* |---|-------|---|
|
|
|
|
*/
|
2019-07-12 11:58:57 +08:00
|
|
|
lva = __this_cpu_xchg(ne_fit_preload_node, NULL);
|
|
|
|
if (unlikely(!lva)) {
|
|
|
|
/*
|
|
|
|
* For percpu allocator we do not do any pre-allocation
|
|
|
|
* and leave it as it is. The reason is it most likely
|
|
|
|
* never ends up with NE_FIT_TYPE splitting. In case of
|
|
|
|
* percpu allocations offsets and sizes are aligned to
|
|
|
|
* fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
|
|
|
|
* are its main fitting cases.
|
|
|
|
*
|
|
|
|
* There are a few exceptions though, as an example it is
|
|
|
|
* a first allocation (early boot up) when we have "one"
|
|
|
|
* big free space that has to be split.
|
2019-12-01 09:54:40 +08:00
|
|
|
*
|
|
|
|
* Also we can hit this path in case of regular "vmap"
|
|
|
|
* allocations, if "this" current CPU was not preloaded.
|
|
|
|
* See the comment in alloc_vmap_area() why. If so, then
|
|
|
|
* GFP_NOWAIT is used instead to get an extra object for
|
|
|
|
* split purpose. That is rare and most time does not
|
|
|
|
* occur.
|
|
|
|
*
|
|
|
|
* What happens if an allocation gets failed. Basically,
|
|
|
|
* an "overflow" path is triggered to purge lazily freed
|
|
|
|
* areas to free some memory, then, the "retry" path is
|
|
|
|
* triggered to repeat one more time. See more details
|
|
|
|
* in alloc_vmap_area() function.
|
2019-07-12 11:58:57 +08:00
|
|
|
*/
|
|
|
|
lva = kmem_cache_alloc(vmap_area_cachep, GFP_NOWAIT);
|
|
|
|
if (!lva)
|
|
|
|
return -1;
|
|
|
|
}
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
|
|
|
|
/*
|
|
|
|
* Build the remainder.
|
|
|
|
*/
|
|
|
|
lva->va_start = va->va_start;
|
|
|
|
lva->va_end = nva_start_addr;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Shrink this VA to remaining size.
|
|
|
|
*/
|
|
|
|
va->va_start = nva_start_addr + size;
|
|
|
|
} else {
|
|
|
|
return -1;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (type != FL_FIT_TYPE) {
|
|
|
|
augment_tree_propagate_from(va);
|
|
|
|
|
2019-06-29 03:07:09 +08:00
|
|
|
if (lva) /* type == NE_FIT_TYPE */
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
insert_vmap_area_augment(lva, &va->rb_node,
|
|
|
|
&free_vmap_area_root, &free_vmap_area_list);
|
|
|
|
}
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Returns a start address of the newly allocated area, if success.
|
|
|
|
* Otherwise a vend is returned that indicates failure.
|
|
|
|
*/
|
|
|
|
static __always_inline unsigned long
|
|
|
|
__alloc_vmap_area(unsigned long size, unsigned long align,
|
2019-07-12 11:58:53 +08:00
|
|
|
unsigned long vstart, unsigned long vend)
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
{
|
|
|
|
unsigned long nva_start_addr;
|
|
|
|
struct vmap_area *va;
|
|
|
|
enum fit_type type;
|
|
|
|
int ret;
|
|
|
|
|
|
|
|
va = find_vmap_lowest_match(size, align, vstart);
|
|
|
|
if (unlikely(!va))
|
|
|
|
return vend;
|
|
|
|
|
|
|
|
if (va->va_start > vstart)
|
|
|
|
nva_start_addr = ALIGN(va->va_start, align);
|
|
|
|
else
|
|
|
|
nva_start_addr = ALIGN(vstart, align);
|
|
|
|
|
|
|
|
/* Check the "vend" restriction. */
|
|
|
|
if (nva_start_addr + size > vend)
|
|
|
|
return vend;
|
|
|
|
|
|
|
|
/* Classify what we have found. */
|
|
|
|
type = classify_va_fit_type(va, nva_start_addr, size);
|
|
|
|
if (WARN_ON_ONCE(type == NOTHING_FIT))
|
|
|
|
return vend;
|
|
|
|
|
|
|
|
/* Update the free vmap_area. */
|
|
|
|
ret = adjust_va_to_fit_type(va, nva_start_addr, size, type);
|
|
|
|
if (ret)
|
|
|
|
return vend;
|
|
|
|
|
2019-05-18 05:31:37 +08:00
|
|
|
#if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
|
|
|
|
find_vmap_lowest_match_check(size);
|
|
|
|
#endif
|
|
|
|
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
return nva_start_addr;
|
|
|
|
}
|
2016-04-04 21:46:42 +08:00
|
|
|
|
2019-12-18 12:51:38 +08:00
|
|
|
/*
|
|
|
|
* Free a region of KVA allocated by alloc_vmap_area
|
|
|
|
*/
|
|
|
|
static void free_vmap_area(struct vmap_area *va)
|
|
|
|
{
|
|
|
|
/*
|
|
|
|
* Remove from the busy tree/list.
|
|
|
|
*/
|
|
|
|
spin_lock(&vmap_area_lock);
|
|
|
|
unlink_va(va, &vmap_area_root);
|
|
|
|
spin_unlock(&vmap_area_lock);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Insert/Merge it back to the free tree/list.
|
|
|
|
*/
|
|
|
|
spin_lock(&free_vmap_area_lock);
|
|
|
|
merge_or_add_vmap_area(va, &free_vmap_area_root, &free_vmap_area_list);
|
|
|
|
spin_unlock(&free_vmap_area_lock);
|
|
|
|
}
|
|
|
|
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
/*
|
|
|
|
* Allocate a region of KVA of the specified size and alignment, within the
|
|
|
|
* vstart and vend.
|
|
|
|
*/
|
|
|
|
static struct vmap_area *alloc_vmap_area(unsigned long size,
|
|
|
|
unsigned long align,
|
|
|
|
unsigned long vstart, unsigned long vend,
|
|
|
|
int node, gfp_t gfp_mask)
|
|
|
|
{
|
2019-07-12 11:58:57 +08:00
|
|
|
struct vmap_area *va, *pva;
|
2005-04-17 06:20:36 +08:00
|
|
|
unsigned long addr;
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
int purged = 0;
|
2019-12-18 12:51:38 +08:00
|
|
|
int ret;
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
|
2009-02-28 06:03:03 +08:00
|
|
|
BUG_ON(!size);
|
2015-11-06 10:46:51 +08:00
|
|
|
BUG_ON(offset_in_page(size));
|
mm: vmap area cache
Provide a free area cache for the vmalloc virtual address allocator, based
on the algorithm used by the user virtual memory allocator.
This reduces the number of rbtree operations and linear traversals over
the vmap extents in order to find a free area, by starting off at the last
point that a free area was found.
The free area cache is reset if areas are freed behind it, or if we are
searching for a smaller area or alignment than last time. So allocation
patterns are not changed (verified by corner-case and random test cases in
userspace testing).
This solves a regression caused by lazy vunmap TLB purging introduced in
db64fe02 (mm: rewrite vmap layer). That patch will leave extents in the
vmap allocator after they are vunmapped, and until a significant number
accumulate that can be flushed in a single batch. So in a workload that
vmalloc/vfree frequently, a chain of extents will build up from
VMALLOC_START address, which have to be iterated over each time (giving an
O(n) type of behaviour).
After this patch, the search will start from where it left off, giving
closer to an amortized O(1).
This is verified to solve regressions reported Steven in GFS2, and Avi in
KVM.
Hugh's update:
: I tried out the recent mmotm, and on one machine was fortunate to hit
: the BUG_ON(first->va_start < addr) which seems to have been stalling
: your vmap area cache patch ever since May.
: I can get you addresses etc, I did dump a few out; but once I stared
: at them, it was easier just to look at the code: and I cannot see how
: you would be so sure that first->va_start < addr, once you've done
: that addr = ALIGN(max(...), align) above, if align is over 0x1000
: (align was 0x8000 or 0x4000 in the cases I hit: ioremaps like Steve).
: I originally got around it by just changing the
: if (first->va_start < addr) {
: to
: while (first->va_start < addr) {
: without thinking about it any further; but that seemed unsatisfactory,
: why would we want to loop here when we've got another very similar
: loop just below it?
: I am never going to admit how long I've spent trying to grasp your
: "while (n)" rbtree loop just above this, the one with the peculiar
: if (!first && tmp->va_start < addr + size)
: in. That's unfamiliar to me, I'm guessing it's designed to save a
: subsequent rb_next() in a few circumstances (at risk of then setting
: a wrong cached_hole_size?); but they did appear few to me, and I didn't
: feel I could sign off something with that in when I don't grasp it,
: and it seems responsible for extra code and mistaken BUG_ON below it.
: I've reverted to the familiar rbtree loop that find_vma() does (but
: with va_end >= addr as you had, to respect the additional guard page):
: and then (given that cached_hole_size starts out 0) I don't see the
: need for any complications below it. If you do want to keep that loop
: as you had it, please add a comment to explain what it's trying to do,
: and where addr is relative to first when you emerge from it.
: Aren't your tests "size <= cached_hole_size" and
: "addr + size > first->va_start" forgetting the guard page we want
: before the next area? I've changed those.
: I have not changed your many "addr + size - 1 < addr" overflow tests,
: but have since come to wonder, shouldn't they be "addr + size < addr"
: tests - won't the vend checks go wrong if addr + size is 0?
: I have added a few comments - Wolfgang Wander's 2.6.13 description of
: 1363c3cd8603a913a27e2995dccbd70d5312d8e6 Avoiding mmap fragmentation
: helped me a lot, perhaps a pointer to that would be good too. And I found
: it easier to understand when I renamed cached_start slightly and moved the
: overflow label down.
: This patch would go after your mm-vmap-area-cache.patch in mmotm.
: Trivially, nobody is going to get that BUG_ON with this patch, and it
: appears to work fine on my machines; but I have not given it anything like
: the testing you did on your original, and may have broken all the
: performance you were aiming for. Please take a look and test it out
: integrate with yours if you're satisfied - thanks.
[akpm@linux-foundation.org: add locking comment]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Reported-and-tested-by: Steven Whitehouse <swhiteho@redhat.com>
Reported-and-tested-by: Avi Kivity <avi@redhat.com>
Tested-by: "Barry J. Marson" <bmarson@redhat.com>
Cc: Prarit Bhargava <prarit@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-03-23 07:30:36 +08:00
|
|
|
BUG_ON(!is_power_of_2(align));
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
if (unlikely(!vmap_initialized))
|
|
|
|
return ERR_PTR(-EBUSY);
|
|
|
|
|
2016-12-13 08:44:20 +08:00
|
|
|
might_sleep();
|
2019-12-01 09:54:37 +08:00
|
|
|
gfp_mask = gfp_mask & GFP_RECLAIM_MASK;
|
2016-04-04 21:46:42 +08:00
|
|
|
|
2019-12-01 09:54:37 +08:00
|
|
|
va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
if (unlikely(!va))
|
|
|
|
return ERR_PTR(-ENOMEM);
|
|
|
|
|
2013-11-13 07:07:45 +08:00
|
|
|
/*
|
|
|
|
* Only scan the relevant parts containing pointers to other objects
|
|
|
|
* to avoid false negatives.
|
|
|
|
*/
|
2019-12-01 09:54:37 +08:00
|
|
|
kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask);
|
2013-11-13 07:07:45 +08:00
|
|
|
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
retry:
|
2019-07-12 11:58:57 +08:00
|
|
|
/*
|
2019-12-01 09:54:33 +08:00
|
|
|
* Preload this CPU with one extra vmap_area object. It is used
|
|
|
|
* when fit type of free area is NE_FIT_TYPE. Please note, it
|
|
|
|
* does not guarantee that an allocation occurs on a CPU that
|
|
|
|
* is preloaded, instead we minimize the case when it is not.
|
|
|
|
* It can happen because of cpu migration, because there is a
|
|
|
|
* race until the below spinlock is taken.
|
2019-07-12 11:58:57 +08:00
|
|
|
*
|
|
|
|
* The preload is done in non-atomic context, thus it allows us
|
|
|
|
* to use more permissive allocation masks to be more stable under
|
2019-12-01 09:54:33 +08:00
|
|
|
* low memory condition and high memory pressure. In rare case,
|
|
|
|
* if not preloaded, GFP_NOWAIT is used.
|
2019-07-12 11:58:57 +08:00
|
|
|
*
|
2019-12-01 09:54:33 +08:00
|
|
|
* Set "pva" to NULL here, because of "retry" path.
|
2019-07-12 11:58:57 +08:00
|
|
|
*/
|
2019-12-01 09:54:33 +08:00
|
|
|
pva = NULL;
|
2019-07-12 11:58:57 +08:00
|
|
|
|
2019-12-01 09:54:33 +08:00
|
|
|
if (!this_cpu_read(ne_fit_preload_node))
|
|
|
|
/*
|
|
|
|
* Even if it fails we do not really care about that.
|
|
|
|
* Just proceed as it is. If needed "overflow" path
|
|
|
|
* will refill the cache we allocate from.
|
|
|
|
*/
|
2019-12-01 09:54:37 +08:00
|
|
|
pva = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
|
2019-07-12 11:58:57 +08:00
|
|
|
|
mm/vmalloc: rework vmap_area_lock
With the new allocation approach introduced in the 5.2 kernel, it
becomes possible to get rid of one global spinlock. By doing that we
can further improve the KVA from the performance point of view.
Basically we can have two independent locks, one for allocation part and
another one for deallocation, because of two different entities: "free
data structures" and "busy data structures".
As a result, allocation/deallocation operations can still interfere
between each other in case of running simultaneously on different CPUs,
it means there is still dependency, but with two locks it becomes lower.
Summarizing:
- it reduces the high lock contention
- it allows to perform operations on "free" and "busy"
trees in parallel on different CPUs. Please note it
does not solve scalability issue.
Test results:
In order to evaluate this patch, we can run "vmalloc test driver" to see
how many CPU cycles it takes to complete all test cases running
sequentially. All online CPUs run it so it will cause a high lock
contention.
HiKey 960, ARM64, 8xCPUs, big.LITTLE:
<snip>
sudo ./test_vmalloc.sh sequential_test_order=1
<snip>
<default>
[ 390.950557] All test took CPU0=457126382 cycles
[ 391.046690] All test took CPU1=454763452 cycles
[ 391.128586] All test took CPU2=454539334 cycles
[ 391.222669] All test took CPU3=455649517 cycles
[ 391.313946] All test took CPU4=388272196 cycles
[ 391.410425] All test took CPU5=384036264 cycles
[ 391.492219] All test took CPU6=387432964 cycles
[ 391.578433] All test took CPU7=387201996 cycles
<default>
<patched>
[ 304.721224] All test took CPU0=391521310 cycles
[ 304.821219] All test took CPU1=393533002 cycles
[ 304.917120] All test took CPU2=392243032 cycles
[ 305.008986] All test took CPU3=392353853 cycles
[ 305.108944] All test took CPU4=297630721 cycles
[ 305.196406] All test took CPU5=297548736 cycles
[ 305.288602] All test took CPU6=297092392 cycles
[ 305.381088] All test took CPU7=297293597 cycles
<patched>
~14%-23% patched variant is better.
Link: http://lkml.kernel.org/r/20191022155800.20468-1-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Acked-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:47 +08:00
|
|
|
spin_lock(&free_vmap_area_lock);
|
2019-12-01 09:54:33 +08:00
|
|
|
|
|
|
|
if (pva && __this_cpu_cmpxchg(ne_fit_preload_node, NULL, pva))
|
|
|
|
kmem_cache_free(vmap_area_cachep, pva);
|
mm: vmap area cache
Provide a free area cache for the vmalloc virtual address allocator, based
on the algorithm used by the user virtual memory allocator.
This reduces the number of rbtree operations and linear traversals over
the vmap extents in order to find a free area, by starting off at the last
point that a free area was found.
The free area cache is reset if areas are freed behind it, or if we are
searching for a smaller area or alignment than last time. So allocation
patterns are not changed (verified by corner-case and random test cases in
userspace testing).
This solves a regression caused by lazy vunmap TLB purging introduced in
db64fe02 (mm: rewrite vmap layer). That patch will leave extents in the
vmap allocator after they are vunmapped, and until a significant number
accumulate that can be flushed in a single batch. So in a workload that
vmalloc/vfree frequently, a chain of extents will build up from
VMALLOC_START address, which have to be iterated over each time (giving an
O(n) type of behaviour).
After this patch, the search will start from where it left off, giving
closer to an amortized O(1).
This is verified to solve regressions reported Steven in GFS2, and Avi in
KVM.
Hugh's update:
: I tried out the recent mmotm, and on one machine was fortunate to hit
: the BUG_ON(first->va_start < addr) which seems to have been stalling
: your vmap area cache patch ever since May.
: I can get you addresses etc, I did dump a few out; but once I stared
: at them, it was easier just to look at the code: and I cannot see how
: you would be so sure that first->va_start < addr, once you've done
: that addr = ALIGN(max(...), align) above, if align is over 0x1000
: (align was 0x8000 or 0x4000 in the cases I hit: ioremaps like Steve).
: I originally got around it by just changing the
: if (first->va_start < addr) {
: to
: while (first->va_start < addr) {
: without thinking about it any further; but that seemed unsatisfactory,
: why would we want to loop here when we've got another very similar
: loop just below it?
: I am never going to admit how long I've spent trying to grasp your
: "while (n)" rbtree loop just above this, the one with the peculiar
: if (!first && tmp->va_start < addr + size)
: in. That's unfamiliar to me, I'm guessing it's designed to save a
: subsequent rb_next() in a few circumstances (at risk of then setting
: a wrong cached_hole_size?); but they did appear few to me, and I didn't
: feel I could sign off something with that in when I don't grasp it,
: and it seems responsible for extra code and mistaken BUG_ON below it.
: I've reverted to the familiar rbtree loop that find_vma() does (but
: with va_end >= addr as you had, to respect the additional guard page):
: and then (given that cached_hole_size starts out 0) I don't see the
: need for any complications below it. If you do want to keep that loop
: as you had it, please add a comment to explain what it's trying to do,
: and where addr is relative to first when you emerge from it.
: Aren't your tests "size <= cached_hole_size" and
: "addr + size > first->va_start" forgetting the guard page we want
: before the next area? I've changed those.
: I have not changed your many "addr + size - 1 < addr" overflow tests,
: but have since come to wonder, shouldn't they be "addr + size < addr"
: tests - won't the vend checks go wrong if addr + size is 0?
: I have added a few comments - Wolfgang Wander's 2.6.13 description of
: 1363c3cd8603a913a27e2995dccbd70d5312d8e6 Avoiding mmap fragmentation
: helped me a lot, perhaps a pointer to that would be good too. And I found
: it easier to understand when I renamed cached_start slightly and moved the
: overflow label down.
: This patch would go after your mm-vmap-area-cache.patch in mmotm.
: Trivially, nobody is going to get that BUG_ON with this patch, and it
: appears to work fine on my machines; but I have not given it anything like
: the testing you did on your original, and may have broken all the
: performance you were aiming for. Please take a look and test it out
: integrate with yours if you're satisfied - thanks.
[akpm@linux-foundation.org: add locking comment]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Reported-and-tested-by: Steven Whitehouse <swhiteho@redhat.com>
Reported-and-tested-by: Avi Kivity <avi@redhat.com>
Tested-by: "Barry J. Marson" <bmarson@redhat.com>
Cc: Prarit Bhargava <prarit@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-03-23 07:30:36 +08:00
|
|
|
|
2019-03-06 07:45:59 +08:00
|
|
|
/*
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
* If an allocation fails, the "vend" address is
|
|
|
|
* returned. Therefore trigger the overflow path.
|
2019-03-06 07:45:59 +08:00
|
|
|
*/
|
2019-07-12 11:58:53 +08:00
|
|
|
addr = __alloc_vmap_area(size, align, vstart, vend);
|
mm/vmalloc: rework vmap_area_lock
With the new allocation approach introduced in the 5.2 kernel, it
becomes possible to get rid of one global spinlock. By doing that we
can further improve the KVA from the performance point of view.
Basically we can have two independent locks, one for allocation part and
another one for deallocation, because of two different entities: "free
data structures" and "busy data structures".
As a result, allocation/deallocation operations can still interfere
between each other in case of running simultaneously on different CPUs,
it means there is still dependency, but with two locks it becomes lower.
Summarizing:
- it reduces the high lock contention
- it allows to perform operations on "free" and "busy"
trees in parallel on different CPUs. Please note it
does not solve scalability issue.
Test results:
In order to evaluate this patch, we can run "vmalloc test driver" to see
how many CPU cycles it takes to complete all test cases running
sequentially. All online CPUs run it so it will cause a high lock
contention.
HiKey 960, ARM64, 8xCPUs, big.LITTLE:
<snip>
sudo ./test_vmalloc.sh sequential_test_order=1
<snip>
<default>
[ 390.950557] All test took CPU0=457126382 cycles
[ 391.046690] All test took CPU1=454763452 cycles
[ 391.128586] All test took CPU2=454539334 cycles
[ 391.222669] All test took CPU3=455649517 cycles
[ 391.313946] All test took CPU4=388272196 cycles
[ 391.410425] All test took CPU5=384036264 cycles
[ 391.492219] All test took CPU6=387432964 cycles
[ 391.578433] All test took CPU7=387201996 cycles
<default>
<patched>
[ 304.721224] All test took CPU0=391521310 cycles
[ 304.821219] All test took CPU1=393533002 cycles
[ 304.917120] All test took CPU2=392243032 cycles
[ 305.008986] All test took CPU3=392353853 cycles
[ 305.108944] All test took CPU4=297630721 cycles
[ 305.196406] All test took CPU5=297548736 cycles
[ 305.288602] All test took CPU6=297092392 cycles
[ 305.381088] All test took CPU7=297293597 cycles
<patched>
~14%-23% patched variant is better.
Link: http://lkml.kernel.org/r/20191022155800.20468-1-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Acked-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:47 +08:00
|
|
|
spin_unlock(&free_vmap_area_lock);
|
|
|
|
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
if (unlikely(addr == vend))
|
mm: vmap area cache
Provide a free area cache for the vmalloc virtual address allocator, based
on the algorithm used by the user virtual memory allocator.
This reduces the number of rbtree operations and linear traversals over
the vmap extents in order to find a free area, by starting off at the last
point that a free area was found.
The free area cache is reset if areas are freed behind it, or if we are
searching for a smaller area or alignment than last time. So allocation
patterns are not changed (verified by corner-case and random test cases in
userspace testing).
This solves a regression caused by lazy vunmap TLB purging introduced in
db64fe02 (mm: rewrite vmap layer). That patch will leave extents in the
vmap allocator after they are vunmapped, and until a significant number
accumulate that can be flushed in a single batch. So in a workload that
vmalloc/vfree frequently, a chain of extents will build up from
VMALLOC_START address, which have to be iterated over each time (giving an
O(n) type of behaviour).
After this patch, the search will start from where it left off, giving
closer to an amortized O(1).
This is verified to solve regressions reported Steven in GFS2, and Avi in
KVM.
Hugh's update:
: I tried out the recent mmotm, and on one machine was fortunate to hit
: the BUG_ON(first->va_start < addr) which seems to have been stalling
: your vmap area cache patch ever since May.
: I can get you addresses etc, I did dump a few out; but once I stared
: at them, it was easier just to look at the code: and I cannot see how
: you would be so sure that first->va_start < addr, once you've done
: that addr = ALIGN(max(...), align) above, if align is over 0x1000
: (align was 0x8000 or 0x4000 in the cases I hit: ioremaps like Steve).
: I originally got around it by just changing the
: if (first->va_start < addr) {
: to
: while (first->va_start < addr) {
: without thinking about it any further; but that seemed unsatisfactory,
: why would we want to loop here when we've got another very similar
: loop just below it?
: I am never going to admit how long I've spent trying to grasp your
: "while (n)" rbtree loop just above this, the one with the peculiar
: if (!first && tmp->va_start < addr + size)
: in. That's unfamiliar to me, I'm guessing it's designed to save a
: subsequent rb_next() in a few circumstances (at risk of then setting
: a wrong cached_hole_size?); but they did appear few to me, and I didn't
: feel I could sign off something with that in when I don't grasp it,
: and it seems responsible for extra code and mistaken BUG_ON below it.
: I've reverted to the familiar rbtree loop that find_vma() does (but
: with va_end >= addr as you had, to respect the additional guard page):
: and then (given that cached_hole_size starts out 0) I don't see the
: need for any complications below it. If you do want to keep that loop
: as you had it, please add a comment to explain what it's trying to do,
: and where addr is relative to first when you emerge from it.
: Aren't your tests "size <= cached_hole_size" and
: "addr + size > first->va_start" forgetting the guard page we want
: before the next area? I've changed those.
: I have not changed your many "addr + size - 1 < addr" overflow tests,
: but have since come to wonder, shouldn't they be "addr + size < addr"
: tests - won't the vend checks go wrong if addr + size is 0?
: I have added a few comments - Wolfgang Wander's 2.6.13 description of
: 1363c3cd8603a913a27e2995dccbd70d5312d8e6 Avoiding mmap fragmentation
: helped me a lot, perhaps a pointer to that would be good too. And I found
: it easier to understand when I renamed cached_start slightly and moved the
: overflow label down.
: This patch would go after your mm-vmap-area-cache.patch in mmotm.
: Trivially, nobody is going to get that BUG_ON with this patch, and it
: appears to work fine on my machines; but I have not given it anything like
: the testing you did on your original, and may have broken all the
: performance you were aiming for. Please take a look and test it out
: integrate with yours if you're satisfied - thanks.
[akpm@linux-foundation.org: add locking comment]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Reported-and-tested-by: Steven Whitehouse <swhiteho@redhat.com>
Reported-and-tested-by: Avi Kivity <avi@redhat.com>
Tested-by: "Barry J. Marson" <bmarson@redhat.com>
Cc: Prarit Bhargava <prarit@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-03-23 07:30:36 +08:00
|
|
|
goto overflow;
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
|
|
|
|
va->va_start = addr;
|
|
|
|
va->va_end = addr + size;
|
2019-09-24 06:36:39 +08:00
|
|
|
va->vm = NULL;
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
|
2019-12-18 12:51:38 +08:00
|
|
|
|
mm/vmalloc: rework vmap_area_lock
With the new allocation approach introduced in the 5.2 kernel, it
becomes possible to get rid of one global spinlock. By doing that we
can further improve the KVA from the performance point of view.
Basically we can have two independent locks, one for allocation part and
another one for deallocation, because of two different entities: "free
data structures" and "busy data structures".
As a result, allocation/deallocation operations can still interfere
between each other in case of running simultaneously on different CPUs,
it means there is still dependency, but with two locks it becomes lower.
Summarizing:
- it reduces the high lock contention
- it allows to perform operations on "free" and "busy"
trees in parallel on different CPUs. Please note it
does not solve scalability issue.
Test results:
In order to evaluate this patch, we can run "vmalloc test driver" to see
how many CPU cycles it takes to complete all test cases running
sequentially. All online CPUs run it so it will cause a high lock
contention.
HiKey 960, ARM64, 8xCPUs, big.LITTLE:
<snip>
sudo ./test_vmalloc.sh sequential_test_order=1
<snip>
<default>
[ 390.950557] All test took CPU0=457126382 cycles
[ 391.046690] All test took CPU1=454763452 cycles
[ 391.128586] All test took CPU2=454539334 cycles
[ 391.222669] All test took CPU3=455649517 cycles
[ 391.313946] All test took CPU4=388272196 cycles
[ 391.410425] All test took CPU5=384036264 cycles
[ 391.492219] All test took CPU6=387432964 cycles
[ 391.578433] All test took CPU7=387201996 cycles
<default>
<patched>
[ 304.721224] All test took CPU0=391521310 cycles
[ 304.821219] All test took CPU1=393533002 cycles
[ 304.917120] All test took CPU2=392243032 cycles
[ 305.008986] All test took CPU3=392353853 cycles
[ 305.108944] All test took CPU4=297630721 cycles
[ 305.196406] All test took CPU5=297548736 cycles
[ 305.288602] All test took CPU6=297092392 cycles
[ 305.381088] All test took CPU7=297293597 cycles
<patched>
~14%-23% patched variant is better.
Link: http://lkml.kernel.org/r/20191022155800.20468-1-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Acked-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:47 +08:00
|
|
|
spin_lock(&vmap_area_lock);
|
|
|
|
insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
spin_unlock(&vmap_area_lock);
|
|
|
|
|
2016-01-16 08:57:19 +08:00
|
|
|
BUG_ON(!IS_ALIGNED(va->va_start, align));
|
mm: vmap area cache
Provide a free area cache for the vmalloc virtual address allocator, based
on the algorithm used by the user virtual memory allocator.
This reduces the number of rbtree operations and linear traversals over
the vmap extents in order to find a free area, by starting off at the last
point that a free area was found.
The free area cache is reset if areas are freed behind it, or if we are
searching for a smaller area or alignment than last time. So allocation
patterns are not changed (verified by corner-case and random test cases in
userspace testing).
This solves a regression caused by lazy vunmap TLB purging introduced in
db64fe02 (mm: rewrite vmap layer). That patch will leave extents in the
vmap allocator after they are vunmapped, and until a significant number
accumulate that can be flushed in a single batch. So in a workload that
vmalloc/vfree frequently, a chain of extents will build up from
VMALLOC_START address, which have to be iterated over each time (giving an
O(n) type of behaviour).
After this patch, the search will start from where it left off, giving
closer to an amortized O(1).
This is verified to solve regressions reported Steven in GFS2, and Avi in
KVM.
Hugh's update:
: I tried out the recent mmotm, and on one machine was fortunate to hit
: the BUG_ON(first->va_start < addr) which seems to have been stalling
: your vmap area cache patch ever since May.
: I can get you addresses etc, I did dump a few out; but once I stared
: at them, it was easier just to look at the code: and I cannot see how
: you would be so sure that first->va_start < addr, once you've done
: that addr = ALIGN(max(...), align) above, if align is over 0x1000
: (align was 0x8000 or 0x4000 in the cases I hit: ioremaps like Steve).
: I originally got around it by just changing the
: if (first->va_start < addr) {
: to
: while (first->va_start < addr) {
: without thinking about it any further; but that seemed unsatisfactory,
: why would we want to loop here when we've got another very similar
: loop just below it?
: I am never going to admit how long I've spent trying to grasp your
: "while (n)" rbtree loop just above this, the one with the peculiar
: if (!first && tmp->va_start < addr + size)
: in. That's unfamiliar to me, I'm guessing it's designed to save a
: subsequent rb_next() in a few circumstances (at risk of then setting
: a wrong cached_hole_size?); but they did appear few to me, and I didn't
: feel I could sign off something with that in when I don't grasp it,
: and it seems responsible for extra code and mistaken BUG_ON below it.
: I've reverted to the familiar rbtree loop that find_vma() does (but
: with va_end >= addr as you had, to respect the additional guard page):
: and then (given that cached_hole_size starts out 0) I don't see the
: need for any complications below it. If you do want to keep that loop
: as you had it, please add a comment to explain what it's trying to do,
: and where addr is relative to first when you emerge from it.
: Aren't your tests "size <= cached_hole_size" and
: "addr + size > first->va_start" forgetting the guard page we want
: before the next area? I've changed those.
: I have not changed your many "addr + size - 1 < addr" overflow tests,
: but have since come to wonder, shouldn't they be "addr + size < addr"
: tests - won't the vend checks go wrong if addr + size is 0?
: I have added a few comments - Wolfgang Wander's 2.6.13 description of
: 1363c3cd8603a913a27e2995dccbd70d5312d8e6 Avoiding mmap fragmentation
: helped me a lot, perhaps a pointer to that would be good too. And I found
: it easier to understand when I renamed cached_start slightly and moved the
: overflow label down.
: This patch would go after your mm-vmap-area-cache.patch in mmotm.
: Trivially, nobody is going to get that BUG_ON with this patch, and it
: appears to work fine on my machines; but I have not given it anything like
: the testing you did on your original, and may have broken all the
: performance you were aiming for. Please take a look and test it out
: integrate with yours if you're satisfied - thanks.
[akpm@linux-foundation.org: add locking comment]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Reported-and-tested-by: Steven Whitehouse <swhiteho@redhat.com>
Reported-and-tested-by: Avi Kivity <avi@redhat.com>
Tested-by: "Barry J. Marson" <bmarson@redhat.com>
Cc: Prarit Bhargava <prarit@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-03-23 07:30:36 +08:00
|
|
|
BUG_ON(va->va_start < vstart);
|
|
|
|
BUG_ON(va->va_end > vend);
|
|
|
|
|
2019-12-18 12:51:38 +08:00
|
|
|
ret = kasan_populate_vmalloc(addr, size);
|
|
|
|
if (ret) {
|
|
|
|
free_vmap_area(va);
|
|
|
|
return ERR_PTR(ret);
|
|
|
|
}
|
|
|
|
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
return va;
|
mm: vmap area cache
Provide a free area cache for the vmalloc virtual address allocator, based
on the algorithm used by the user virtual memory allocator.
This reduces the number of rbtree operations and linear traversals over
the vmap extents in order to find a free area, by starting off at the last
point that a free area was found.
The free area cache is reset if areas are freed behind it, or if we are
searching for a smaller area or alignment than last time. So allocation
patterns are not changed (verified by corner-case and random test cases in
userspace testing).
This solves a regression caused by lazy vunmap TLB purging introduced in
db64fe02 (mm: rewrite vmap layer). That patch will leave extents in the
vmap allocator after they are vunmapped, and until a significant number
accumulate that can be flushed in a single batch. So in a workload that
vmalloc/vfree frequently, a chain of extents will build up from
VMALLOC_START address, which have to be iterated over each time (giving an
O(n) type of behaviour).
After this patch, the search will start from where it left off, giving
closer to an amortized O(1).
This is verified to solve regressions reported Steven in GFS2, and Avi in
KVM.
Hugh's update:
: I tried out the recent mmotm, and on one machine was fortunate to hit
: the BUG_ON(first->va_start < addr) which seems to have been stalling
: your vmap area cache patch ever since May.
: I can get you addresses etc, I did dump a few out; but once I stared
: at them, it was easier just to look at the code: and I cannot see how
: you would be so sure that first->va_start < addr, once you've done
: that addr = ALIGN(max(...), align) above, if align is over 0x1000
: (align was 0x8000 or 0x4000 in the cases I hit: ioremaps like Steve).
: I originally got around it by just changing the
: if (first->va_start < addr) {
: to
: while (first->va_start < addr) {
: without thinking about it any further; but that seemed unsatisfactory,
: why would we want to loop here when we've got another very similar
: loop just below it?
: I am never going to admit how long I've spent trying to grasp your
: "while (n)" rbtree loop just above this, the one with the peculiar
: if (!first && tmp->va_start < addr + size)
: in. That's unfamiliar to me, I'm guessing it's designed to save a
: subsequent rb_next() in a few circumstances (at risk of then setting
: a wrong cached_hole_size?); but they did appear few to me, and I didn't
: feel I could sign off something with that in when I don't grasp it,
: and it seems responsible for extra code and mistaken BUG_ON below it.
: I've reverted to the familiar rbtree loop that find_vma() does (but
: with va_end >= addr as you had, to respect the additional guard page):
: and then (given that cached_hole_size starts out 0) I don't see the
: need for any complications below it. If you do want to keep that loop
: as you had it, please add a comment to explain what it's trying to do,
: and where addr is relative to first when you emerge from it.
: Aren't your tests "size <= cached_hole_size" and
: "addr + size > first->va_start" forgetting the guard page we want
: before the next area? I've changed those.
: I have not changed your many "addr + size - 1 < addr" overflow tests,
: but have since come to wonder, shouldn't they be "addr + size < addr"
: tests - won't the vend checks go wrong if addr + size is 0?
: I have added a few comments - Wolfgang Wander's 2.6.13 description of
: 1363c3cd8603a913a27e2995dccbd70d5312d8e6 Avoiding mmap fragmentation
: helped me a lot, perhaps a pointer to that would be good too. And I found
: it easier to understand when I renamed cached_start slightly and moved the
: overflow label down.
: This patch would go after your mm-vmap-area-cache.patch in mmotm.
: Trivially, nobody is going to get that BUG_ON with this patch, and it
: appears to work fine on my machines; but I have not given it anything like
: the testing you did on your original, and may have broken all the
: performance you were aiming for. Please take a look and test it out
: integrate with yours if you're satisfied - thanks.
[akpm@linux-foundation.org: add locking comment]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Reported-and-tested-by: Steven Whitehouse <swhiteho@redhat.com>
Reported-and-tested-by: Avi Kivity <avi@redhat.com>
Tested-by: "Barry J. Marson" <bmarson@redhat.com>
Cc: Prarit Bhargava <prarit@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-03-23 07:30:36 +08:00
|
|
|
|
|
|
|
overflow:
|
|
|
|
if (!purged) {
|
|
|
|
purge_vmap_area_lazy();
|
|
|
|
purged = 1;
|
|
|
|
goto retry;
|
|
|
|
}
|
2016-04-04 21:46:42 +08:00
|
|
|
|
|
|
|
if (gfpflags_allow_blocking(gfp_mask)) {
|
|
|
|
unsigned long freed = 0;
|
|
|
|
blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
|
|
|
|
if (freed > 0) {
|
|
|
|
purged = 0;
|
|
|
|
goto retry;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2017-04-28 02:19:00 +08:00
|
|
|
if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
|
2016-03-18 05:19:47 +08:00
|
|
|
pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
|
|
|
|
size);
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
|
|
|
|
kmem_cache_free(vmap_area_cachep, va);
|
mm: vmap area cache
Provide a free area cache for the vmalloc virtual address allocator, based
on the algorithm used by the user virtual memory allocator.
This reduces the number of rbtree operations and linear traversals over
the vmap extents in order to find a free area, by starting off at the last
point that a free area was found.
The free area cache is reset if areas are freed behind it, or if we are
searching for a smaller area or alignment than last time. So allocation
patterns are not changed (verified by corner-case and random test cases in
userspace testing).
This solves a regression caused by lazy vunmap TLB purging introduced in
db64fe02 (mm: rewrite vmap layer). That patch will leave extents in the
vmap allocator after they are vunmapped, and until a significant number
accumulate that can be flushed in a single batch. So in a workload that
vmalloc/vfree frequently, a chain of extents will build up from
VMALLOC_START address, which have to be iterated over each time (giving an
O(n) type of behaviour).
After this patch, the search will start from where it left off, giving
closer to an amortized O(1).
This is verified to solve regressions reported Steven in GFS2, and Avi in
KVM.
Hugh's update:
: I tried out the recent mmotm, and on one machine was fortunate to hit
: the BUG_ON(first->va_start < addr) which seems to have been stalling
: your vmap area cache patch ever since May.
: I can get you addresses etc, I did dump a few out; but once I stared
: at them, it was easier just to look at the code: and I cannot see how
: you would be so sure that first->va_start < addr, once you've done
: that addr = ALIGN(max(...), align) above, if align is over 0x1000
: (align was 0x8000 or 0x4000 in the cases I hit: ioremaps like Steve).
: I originally got around it by just changing the
: if (first->va_start < addr) {
: to
: while (first->va_start < addr) {
: without thinking about it any further; but that seemed unsatisfactory,
: why would we want to loop here when we've got another very similar
: loop just below it?
: I am never going to admit how long I've spent trying to grasp your
: "while (n)" rbtree loop just above this, the one with the peculiar
: if (!first && tmp->va_start < addr + size)
: in. That's unfamiliar to me, I'm guessing it's designed to save a
: subsequent rb_next() in a few circumstances (at risk of then setting
: a wrong cached_hole_size?); but they did appear few to me, and I didn't
: feel I could sign off something with that in when I don't grasp it,
: and it seems responsible for extra code and mistaken BUG_ON below it.
: I've reverted to the familiar rbtree loop that find_vma() does (but
: with va_end >= addr as you had, to respect the additional guard page):
: and then (given that cached_hole_size starts out 0) I don't see the
: need for any complications below it. If you do want to keep that loop
: as you had it, please add a comment to explain what it's trying to do,
: and where addr is relative to first when you emerge from it.
: Aren't your tests "size <= cached_hole_size" and
: "addr + size > first->va_start" forgetting the guard page we want
: before the next area? I've changed those.
: I have not changed your many "addr + size - 1 < addr" overflow tests,
: but have since come to wonder, shouldn't they be "addr + size < addr"
: tests - won't the vend checks go wrong if addr + size is 0?
: I have added a few comments - Wolfgang Wander's 2.6.13 description of
: 1363c3cd8603a913a27e2995dccbd70d5312d8e6 Avoiding mmap fragmentation
: helped me a lot, perhaps a pointer to that would be good too. And I found
: it easier to understand when I renamed cached_start slightly and moved the
: overflow label down.
: This patch would go after your mm-vmap-area-cache.patch in mmotm.
: Trivially, nobody is going to get that BUG_ON with this patch, and it
: appears to work fine on my machines; but I have not given it anything like
: the testing you did on your original, and may have broken all the
: performance you were aiming for. Please take a look and test it out
: integrate with yours if you're satisfied - thanks.
[akpm@linux-foundation.org: add locking comment]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Hugh Dickins <hughd@google.com>
Reviewed-by: Minchan Kim <minchan.kim@gmail.com>
Reported-and-tested-by: Steven Whitehouse <swhiteho@redhat.com>
Reported-and-tested-by: Avi Kivity <avi@redhat.com>
Tested-by: "Barry J. Marson" <bmarson@redhat.com>
Cc: Prarit Bhargava <prarit@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-03-23 07:30:36 +08:00
|
|
|
return ERR_PTR(-EBUSY);
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
}
|
|
|
|
|
2016-04-04 21:46:42 +08:00
|
|
|
int register_vmap_purge_notifier(struct notifier_block *nb)
|
|
|
|
{
|
|
|
|
return blocking_notifier_chain_register(&vmap_notify_list, nb);
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
|
|
|
|
|
|
|
|
int unregister_vmap_purge_notifier(struct notifier_block *nb)
|
|
|
|
{
|
|
|
|
return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
|
|
|
|
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
/*
|
|
|
|
* Clear the pagetable entries of a given vmap_area
|
|
|
|
*/
|
|
|
|
static void unmap_vmap_area(struct vmap_area *va)
|
|
|
|
{
|
|
|
|
vunmap_page_range(va->va_start, va->va_end);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* lazy_max_pages is the maximum amount of virtual address space we gather up
|
|
|
|
* before attempting to purge with a TLB flush.
|
|
|
|
*
|
|
|
|
* There is a tradeoff here: a larger number will cover more kernel page tables
|
|
|
|
* and take slightly longer to purge, but it will linearly reduce the number of
|
|
|
|
* global TLB flushes that must be performed. It would seem natural to scale
|
|
|
|
* this number up linearly with the number of CPUs (because vmapping activity
|
|
|
|
* could also scale linearly with the number of CPUs), however it is likely
|
|
|
|
* that in practice, workloads might be constrained in other ways that mean
|
|
|
|
* vmap activity will not scale linearly with CPUs. Also, I want to be
|
|
|
|
* conservative and not introduce a big latency on huge systems, so go with
|
|
|
|
* a less aggressive log scale. It will still be an improvement over the old
|
|
|
|
* code, and it will be simple to change the scale factor if we find that it
|
|
|
|
* becomes a problem on bigger systems.
|
|
|
|
*/
|
|
|
|
static unsigned long lazy_max_pages(void)
|
|
|
|
{
|
|
|
|
unsigned int log;
|
|
|
|
|
|
|
|
log = fls(num_online_cpus());
|
|
|
|
|
|
|
|
return log * (32UL * 1024 * 1024 / PAGE_SIZE);
|
|
|
|
}
|
|
|
|
|
2019-05-15 06:41:25 +08:00
|
|
|
static atomic_long_t vmap_lazy_nr = ATOMIC_LONG_INIT(0);
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
|
2016-12-13 08:44:07 +08:00
|
|
|
/*
|
|
|
|
* Serialize vmap purging. There is no actual criticial section protected
|
|
|
|
* by this look, but we want to avoid concurrent calls for performance
|
|
|
|
* reasons and to make the pcpu_get_vm_areas more deterministic.
|
|
|
|
*/
|
2016-12-13 08:44:23 +08:00
|
|
|
static DEFINE_MUTEX(vmap_purge_lock);
|
2016-12-13 08:44:07 +08:00
|
|
|
|
2010-02-01 19:25:57 +08:00
|
|
|
/* for per-CPU blocks */
|
|
|
|
static void purge_fragmented_blocks_allcpus(void);
|
|
|
|
|
2010-09-17 00:44:02 +08:00
|
|
|
/*
|
|
|
|
* called before a call to iounmap() if the caller wants vm_area_struct's
|
|
|
|
* immediately freed.
|
|
|
|
*/
|
|
|
|
void set_iounmap_nonlazy(void)
|
|
|
|
{
|
2019-05-15 06:41:25 +08:00
|
|
|
atomic_long_set(&vmap_lazy_nr, lazy_max_pages()+1);
|
2010-09-17 00:44:02 +08:00
|
|
|
}
|
|
|
|
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
/*
|
|
|
|
* Purges all lazily-freed vmap areas.
|
|
|
|
*/
|
2016-12-13 08:44:07 +08:00
|
|
|
static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
{
|
2019-05-15 06:41:25 +08:00
|
|
|
unsigned long resched_threshold;
|
2016-05-21 07:57:38 +08:00
|
|
|
struct llist_node *valist;
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
struct vmap_area *va;
|
2009-02-28 06:03:04 +08:00
|
|
|
struct vmap_area *n_va;
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
|
2016-12-13 08:44:07 +08:00
|
|
|
lockdep_assert_held(&vmap_purge_lock);
|
2010-02-01 19:25:57 +08:00
|
|
|
|
2016-05-21 07:57:38 +08:00
|
|
|
valist = llist_del_all(&vmap_purge_list);
|
2019-05-15 06:41:22 +08:00
|
|
|
if (unlikely(valist == NULL))
|
|
|
|
return false;
|
|
|
|
|
2019-07-20 02:46:52 +08:00
|
|
|
/*
|
|
|
|
* First make sure the mappings are removed from all page-tables
|
|
|
|
* before they are freed.
|
|
|
|
*/
|
|
|
|
vmalloc_sync_all();
|
|
|
|
|
2019-05-15 06:41:22 +08:00
|
|
|
/*
|
|
|
|
* TODO: to calculate a flush range without looping.
|
|
|
|
* The list can be up to lazy_max_pages() elements.
|
|
|
|
*/
|
2016-05-21 07:57:38 +08:00
|
|
|
llist_for_each_entry(va, valist, purge_list) {
|
2016-12-13 08:44:07 +08:00
|
|
|
if (va->va_start < start)
|
|
|
|
start = va->va_start;
|
|
|
|
if (va->va_end > end)
|
|
|
|
end = va->va_end;
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
}
|
|
|
|
|
2016-12-13 08:44:07 +08:00
|
|
|
flush_tlb_kernel_range(start, end);
|
2019-05-15 06:41:25 +08:00
|
|
|
resched_threshold = lazy_max_pages() << 1;
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
|
mm/vmalloc: rework vmap_area_lock
With the new allocation approach introduced in the 5.2 kernel, it
becomes possible to get rid of one global spinlock. By doing that we
can further improve the KVA from the performance point of view.
Basically we can have two independent locks, one for allocation part and
another one for deallocation, because of two different entities: "free
data structures" and "busy data structures".
As a result, allocation/deallocation operations can still interfere
between each other in case of running simultaneously on different CPUs,
it means there is still dependency, but with two locks it becomes lower.
Summarizing:
- it reduces the high lock contention
- it allows to perform operations on "free" and "busy"
trees in parallel on different CPUs. Please note it
does not solve scalability issue.
Test results:
In order to evaluate this patch, we can run "vmalloc test driver" to see
how many CPU cycles it takes to complete all test cases running
sequentially. All online CPUs run it so it will cause a high lock
contention.
HiKey 960, ARM64, 8xCPUs, big.LITTLE:
<snip>
sudo ./test_vmalloc.sh sequential_test_order=1
<snip>
<default>
[ 390.950557] All test took CPU0=457126382 cycles
[ 391.046690] All test took CPU1=454763452 cycles
[ 391.128586] All test took CPU2=454539334 cycles
[ 391.222669] All test took CPU3=455649517 cycles
[ 391.313946] All test took CPU4=388272196 cycles
[ 391.410425] All test took CPU5=384036264 cycles
[ 391.492219] All test took CPU6=387432964 cycles
[ 391.578433] All test took CPU7=387201996 cycles
<default>
<patched>
[ 304.721224] All test took CPU0=391521310 cycles
[ 304.821219] All test took CPU1=393533002 cycles
[ 304.917120] All test took CPU2=392243032 cycles
[ 305.008986] All test took CPU3=392353853 cycles
[ 305.108944] All test took CPU4=297630721 cycles
[ 305.196406] All test took CPU5=297548736 cycles
[ 305.288602] All test took CPU6=297092392 cycles
[ 305.381088] All test took CPU7=297293597 cycles
<patched>
~14%-23% patched variant is better.
Link: http://lkml.kernel.org/r/20191022155800.20468-1-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Acked-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:47 +08:00
|
|
|
spin_lock(&free_vmap_area_lock);
|
2016-12-13 08:44:26 +08:00
|
|
|
llist_for_each_entry_safe(va, n_va, valist, purge_list) {
|
2019-05-15 06:41:25 +08:00
|
|
|
unsigned long nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
|
kasan: support backing vmalloc space with real shadow memory
Patch series "kasan: support backing vmalloc space with real shadow
memory", v11.
Currently, vmalloc space is backed by the early shadow page. This means
that kasan is incompatible with VMAP_STACK.
This series provides a mechanism to back vmalloc space with real,
dynamically allocated memory. I have only wired up x86, because that's
the only currently supported arch I can work with easily, but it's very
easy to wire up other architectures, and it appears that there is some
work-in-progress code to do this on arm64 and s390.
This has been discussed before in the context of VMAP_STACK:
- https://bugzilla.kernel.org/show_bug.cgi?id=202009
- https://lkml.org/lkml/2018/7/22/198
- https://lkml.org/lkml/2019/7/19/822
In terms of implementation details:
Most mappings in vmalloc space are small, requiring less than a full
page of shadow space. Allocating a full shadow page per mapping would
therefore be wasteful. Furthermore, to ensure that different mappings
use different shadow pages, mappings would have to be aligned to
KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE.
Instead, share backing space across multiple mappings. Allocate a
backing page when a mapping in vmalloc space uses a particular page of
the shadow region. This page can be shared by other vmalloc mappings
later on.
We hook in to the vmap infrastructure to lazily clean up unused shadow
memory.
Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that:
- Turning on KASAN, inline instrumentation, without vmalloc, introuduces
a 4.1x-4.2x slowdown in vmalloc operations.
- Turning this on introduces the following slowdowns over KASAN:
* ~1.76x slower single-threaded (test_vmalloc.sh performance)
* ~2.18x slower when both cpus are performing operations
simultaneously (test_vmalloc.sh sequential_test_order=1)
This is unfortunate but given that this is a debug feature only, not the
end of the world. The benchmarks are also a stress-test for the vmalloc
subsystem: they're not indicative of an overall 2x slowdown!
This patch (of 4):
Hook into vmalloc and vmap, and dynamically allocate real shadow memory
to back the mappings.
Most mappings in vmalloc space are small, requiring less than a full
page of shadow space. Allocating a full shadow page per mapping would
therefore be wasteful. Furthermore, to ensure that different mappings
use different shadow pages, mappings would have to be aligned to
KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE.
Instead, share backing space across multiple mappings. Allocate a
backing page when a mapping in vmalloc space uses a particular page of
the shadow region. This page can be shared by other vmalloc mappings
later on.
We hook in to the vmap infrastructure to lazily clean up unused shadow
memory.
To avoid the difficulties around swapping mappings around, this code
expects that the part of the shadow region that covers the vmalloc space
will not be covered by the early shadow page, but will be left unmapped.
This will require changes in arch-specific code.
This allows KASAN with VMAP_STACK, and may be helpful for architectures
that do not have a separate module space (e.g. powerpc64, which I am
currently working on). It also allows relaxing the module alignment
back to PAGE_SIZE.
Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that:
- Turning on KASAN, inline instrumentation, without vmalloc, introuduces
a 4.1x-4.2x slowdown in vmalloc operations.
- Turning this on introduces the following slowdowns over KASAN:
* ~1.76x slower single-threaded (test_vmalloc.sh performance)
* ~2.18x slower when both cpus are performing operations
simultaneously (test_vmalloc.sh sequential_test_order=3D1)
This is unfortunate but given that this is a debug feature only, not the
end of the world.
The full benchmark results are:
Performance
No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN
fix_size_alloc_test 662004 11404956 17.23 19144610 28.92 1.68
full_fit_alloc_test 710950 12029752 16.92 13184651 18.55 1.10
long_busy_list_alloc_test 9431875 43990172 4.66 82970178 8.80 1.89
random_size_alloc_test 5033626 23061762 4.58 47158834 9.37 2.04
fix_align_alloc_test 1252514 15276910 12.20 31266116 24.96 2.05
random_size_align_alloc_te 1648501 14578321 8.84 25560052 15.51 1.75
align_shift_alloc_test 147 830 5.65 5692 38.72 6.86
pcpu_alloc_test 80732 125520 1.55 140864 1.74 1.12
Total Cycles 119240774314 763211341128 6.40 1390338696894 11.66 1.82
Sequential, 2 cpus
No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN
fix_size_alloc_test 1423150 14276550 10.03 27733022 19.49 1.94
full_fit_alloc_test 1754219 14722640 8.39 15030786 8.57 1.02
long_busy_list_alloc_test 11451858 52154973 4.55 107016027 9.34 2.05
random_size_alloc_test 5989020 26735276 4.46 68885923 11.50 2.58
fix_align_alloc_test 2050976 20166900 9.83 50491675 24.62 2.50
random_size_align_alloc_te 2858229 17971700 6.29 38730225 13.55 2.16
align_shift_alloc_test 405 6428 15.87 26253 64.82 4.08
pcpu_alloc_test 127183 151464 1.19 216263 1.70 1.43
Total Cycles 54181269392 308723699764 5.70 650772566394 12.01 2.11
fix_size_alloc_test 1420404 14289308 10.06 27790035 19.56 1.94
full_fit_alloc_test 1736145 14806234 8.53 15274301 8.80 1.03
long_busy_list_alloc_test 11404638 52270785 4.58 107550254 9.43 2.06
random_size_alloc_test 6017006 26650625 4.43 68696127 11.42 2.58
fix_align_alloc_test 2045504 20280985 9.91 50414862 24.65 2.49
random_size_align_alloc_te 2845338 17931018 6.30 38510276 13.53 2.15
align_shift_alloc_test 472 3760 7.97 9656 20.46 2.57
pcpu_alloc_test 118643 132732 1.12 146504 1.23 1.10
Total Cycles 54040011688 309102805492 5.72 651325675652 12.05 2.11
[dja@axtens.net: fixups]
Link: http://lkml.kernel.org/r/20191120052719.7201-1-dja@axtens.net
Link: https://bugzilla.kernel.org/show_bug.cgi?id=3D202009
Link: http://lkml.kernel.org/r/20191031093909.9228-2-dja@axtens.net
Signed-off-by: Mark Rutland <mark.rutland@arm.com> [shadow rework]
Signed-off-by: Daniel Axtens <dja@axtens.net>
Co-developed-by: Mark Rutland <mark.rutland@arm.com>
Acked-by: Vasily Gorbik <gor@linux.ibm.com>
Reviewed-by: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Christophe Leroy <christophe.leroy@c-s.fr>
Cc: Qian Cai <cai@lca.pw>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:50 +08:00
|
|
|
unsigned long orig_start = va->va_start;
|
|
|
|
unsigned long orig_end = va->va_end;
|
2016-12-13 08:44:26 +08:00
|
|
|
|
2019-09-24 06:36:36 +08:00
|
|
|
/*
|
|
|
|
* Finally insert or merge lazily-freed area. It is
|
|
|
|
* detached and there is no need to "unlink" it from
|
|
|
|
* anything.
|
|
|
|
*/
|
kasan: support backing vmalloc space with real shadow memory
Patch series "kasan: support backing vmalloc space with real shadow
memory", v11.
Currently, vmalloc space is backed by the early shadow page. This means
that kasan is incompatible with VMAP_STACK.
This series provides a mechanism to back vmalloc space with real,
dynamically allocated memory. I have only wired up x86, because that's
the only currently supported arch I can work with easily, but it's very
easy to wire up other architectures, and it appears that there is some
work-in-progress code to do this on arm64 and s390.
This has been discussed before in the context of VMAP_STACK:
- https://bugzilla.kernel.org/show_bug.cgi?id=202009
- https://lkml.org/lkml/2018/7/22/198
- https://lkml.org/lkml/2019/7/19/822
In terms of implementation details:
Most mappings in vmalloc space are small, requiring less than a full
page of shadow space. Allocating a full shadow page per mapping would
therefore be wasteful. Furthermore, to ensure that different mappings
use different shadow pages, mappings would have to be aligned to
KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE.
Instead, share backing space across multiple mappings. Allocate a
backing page when a mapping in vmalloc space uses a particular page of
the shadow region. This page can be shared by other vmalloc mappings
later on.
We hook in to the vmap infrastructure to lazily clean up unused shadow
memory.
Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that:
- Turning on KASAN, inline instrumentation, without vmalloc, introuduces
a 4.1x-4.2x slowdown in vmalloc operations.
- Turning this on introduces the following slowdowns over KASAN:
* ~1.76x slower single-threaded (test_vmalloc.sh performance)
* ~2.18x slower when both cpus are performing operations
simultaneously (test_vmalloc.sh sequential_test_order=1)
This is unfortunate but given that this is a debug feature only, not the
end of the world. The benchmarks are also a stress-test for the vmalloc
subsystem: they're not indicative of an overall 2x slowdown!
This patch (of 4):
Hook into vmalloc and vmap, and dynamically allocate real shadow memory
to back the mappings.
Most mappings in vmalloc space are small, requiring less than a full
page of shadow space. Allocating a full shadow page per mapping would
therefore be wasteful. Furthermore, to ensure that different mappings
use different shadow pages, mappings would have to be aligned to
KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE.
Instead, share backing space across multiple mappings. Allocate a
backing page when a mapping in vmalloc space uses a particular page of
the shadow region. This page can be shared by other vmalloc mappings
later on.
We hook in to the vmap infrastructure to lazily clean up unused shadow
memory.
To avoid the difficulties around swapping mappings around, this code
expects that the part of the shadow region that covers the vmalloc space
will not be covered by the early shadow page, but will be left unmapped.
This will require changes in arch-specific code.
This allows KASAN with VMAP_STACK, and may be helpful for architectures
that do not have a separate module space (e.g. powerpc64, which I am
currently working on). It also allows relaxing the module alignment
back to PAGE_SIZE.
Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that:
- Turning on KASAN, inline instrumentation, without vmalloc, introuduces
a 4.1x-4.2x slowdown in vmalloc operations.
- Turning this on introduces the following slowdowns over KASAN:
* ~1.76x slower single-threaded (test_vmalloc.sh performance)
* ~2.18x slower when both cpus are performing operations
simultaneously (test_vmalloc.sh sequential_test_order=3D1)
This is unfortunate but given that this is a debug feature only, not the
end of the world.
The full benchmark results are:
Performance
No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN
fix_size_alloc_test 662004 11404956 17.23 19144610 28.92 1.68
full_fit_alloc_test 710950 12029752 16.92 13184651 18.55 1.10
long_busy_list_alloc_test 9431875 43990172 4.66 82970178 8.80 1.89
random_size_alloc_test 5033626 23061762 4.58 47158834 9.37 2.04
fix_align_alloc_test 1252514 15276910 12.20 31266116 24.96 2.05
random_size_align_alloc_te 1648501 14578321 8.84 25560052 15.51 1.75
align_shift_alloc_test 147 830 5.65 5692 38.72 6.86
pcpu_alloc_test 80732 125520 1.55 140864 1.74 1.12
Total Cycles 119240774314 763211341128 6.40 1390338696894 11.66 1.82
Sequential, 2 cpus
No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN
fix_size_alloc_test 1423150 14276550 10.03 27733022 19.49 1.94
full_fit_alloc_test 1754219 14722640 8.39 15030786 8.57 1.02
long_busy_list_alloc_test 11451858 52154973 4.55 107016027 9.34 2.05
random_size_alloc_test 5989020 26735276 4.46 68885923 11.50 2.58
fix_align_alloc_test 2050976 20166900 9.83 50491675 24.62 2.50
random_size_align_alloc_te 2858229 17971700 6.29 38730225 13.55 2.16
align_shift_alloc_test 405 6428 15.87 26253 64.82 4.08
pcpu_alloc_test 127183 151464 1.19 216263 1.70 1.43
Total Cycles 54181269392 308723699764 5.70 650772566394 12.01 2.11
fix_size_alloc_test 1420404 14289308 10.06 27790035 19.56 1.94
full_fit_alloc_test 1736145 14806234 8.53 15274301 8.80 1.03
long_busy_list_alloc_test 11404638 52270785 4.58 107550254 9.43 2.06
random_size_alloc_test 6017006 26650625 4.43 68696127 11.42 2.58
fix_align_alloc_test 2045504 20280985 9.91 50414862 24.65 2.49
random_size_align_alloc_te 2845338 17931018 6.30 38510276 13.53 2.15
align_shift_alloc_test 472 3760 7.97 9656 20.46 2.57
pcpu_alloc_test 118643 132732 1.12 146504 1.23 1.10
Total Cycles 54040011688 309102805492 5.72 651325675652 12.05 2.11
[dja@axtens.net: fixups]
Link: http://lkml.kernel.org/r/20191120052719.7201-1-dja@axtens.net
Link: https://bugzilla.kernel.org/show_bug.cgi?id=3D202009
Link: http://lkml.kernel.org/r/20191031093909.9228-2-dja@axtens.net
Signed-off-by: Mark Rutland <mark.rutland@arm.com> [shadow rework]
Signed-off-by: Daniel Axtens <dja@axtens.net>
Co-developed-by: Mark Rutland <mark.rutland@arm.com>
Acked-by: Vasily Gorbik <gor@linux.ibm.com>
Reviewed-by: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Christophe Leroy <christophe.leroy@c-s.fr>
Cc: Qian Cai <cai@lca.pw>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:50 +08:00
|
|
|
va = merge_or_add_vmap_area(va, &free_vmap_area_root,
|
|
|
|
&free_vmap_area_list);
|
|
|
|
|
|
|
|
if (is_vmalloc_or_module_addr((void *)orig_start))
|
|
|
|
kasan_release_vmalloc(orig_start, orig_end,
|
|
|
|
va->va_start, va->va_end);
|
2019-09-24 06:36:36 +08:00
|
|
|
|
2019-05-15 06:41:25 +08:00
|
|
|
atomic_long_sub(nr, &vmap_lazy_nr);
|
2019-05-15 06:41:22 +08:00
|
|
|
|
2019-05-15 06:41:25 +08:00
|
|
|
if (atomic_long_read(&vmap_lazy_nr) < resched_threshold)
|
mm/vmalloc: rework vmap_area_lock
With the new allocation approach introduced in the 5.2 kernel, it
becomes possible to get rid of one global spinlock. By doing that we
can further improve the KVA from the performance point of view.
Basically we can have two independent locks, one for allocation part and
another one for deallocation, because of two different entities: "free
data structures" and "busy data structures".
As a result, allocation/deallocation operations can still interfere
between each other in case of running simultaneously on different CPUs,
it means there is still dependency, but with two locks it becomes lower.
Summarizing:
- it reduces the high lock contention
- it allows to perform operations on "free" and "busy"
trees in parallel on different CPUs. Please note it
does not solve scalability issue.
Test results:
In order to evaluate this patch, we can run "vmalloc test driver" to see
how many CPU cycles it takes to complete all test cases running
sequentially. All online CPUs run it so it will cause a high lock
contention.
HiKey 960, ARM64, 8xCPUs, big.LITTLE:
<snip>
sudo ./test_vmalloc.sh sequential_test_order=1
<snip>
<default>
[ 390.950557] All test took CPU0=457126382 cycles
[ 391.046690] All test took CPU1=454763452 cycles
[ 391.128586] All test took CPU2=454539334 cycles
[ 391.222669] All test took CPU3=455649517 cycles
[ 391.313946] All test took CPU4=388272196 cycles
[ 391.410425] All test took CPU5=384036264 cycles
[ 391.492219] All test took CPU6=387432964 cycles
[ 391.578433] All test took CPU7=387201996 cycles
<default>
<patched>
[ 304.721224] All test took CPU0=391521310 cycles
[ 304.821219] All test took CPU1=393533002 cycles
[ 304.917120] All test took CPU2=392243032 cycles
[ 305.008986] All test took CPU3=392353853 cycles
[ 305.108944] All test took CPU4=297630721 cycles
[ 305.196406] All test took CPU5=297548736 cycles
[ 305.288602] All test took CPU6=297092392 cycles
[ 305.381088] All test took CPU7=297293597 cycles
<patched>
~14%-23% patched variant is better.
Link: http://lkml.kernel.org/r/20191022155800.20468-1-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Acked-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:47 +08:00
|
|
|
cond_resched_lock(&free_vmap_area_lock);
|
2016-12-13 08:44:26 +08:00
|
|
|
}
|
mm/vmalloc: rework vmap_area_lock
With the new allocation approach introduced in the 5.2 kernel, it
becomes possible to get rid of one global spinlock. By doing that we
can further improve the KVA from the performance point of view.
Basically we can have two independent locks, one for allocation part and
another one for deallocation, because of two different entities: "free
data structures" and "busy data structures".
As a result, allocation/deallocation operations can still interfere
between each other in case of running simultaneously on different CPUs,
it means there is still dependency, but with two locks it becomes lower.
Summarizing:
- it reduces the high lock contention
- it allows to perform operations on "free" and "busy"
trees in parallel on different CPUs. Please note it
does not solve scalability issue.
Test results:
In order to evaluate this patch, we can run "vmalloc test driver" to see
how many CPU cycles it takes to complete all test cases running
sequentially. All online CPUs run it so it will cause a high lock
contention.
HiKey 960, ARM64, 8xCPUs, big.LITTLE:
<snip>
sudo ./test_vmalloc.sh sequential_test_order=1
<snip>
<default>
[ 390.950557] All test took CPU0=457126382 cycles
[ 391.046690] All test took CPU1=454763452 cycles
[ 391.128586] All test took CPU2=454539334 cycles
[ 391.222669] All test took CPU3=455649517 cycles
[ 391.313946] All test took CPU4=388272196 cycles
[ 391.410425] All test took CPU5=384036264 cycles
[ 391.492219] All test took CPU6=387432964 cycles
[ 391.578433] All test took CPU7=387201996 cycles
<default>
<patched>
[ 304.721224] All test took CPU0=391521310 cycles
[ 304.821219] All test took CPU1=393533002 cycles
[ 304.917120] All test took CPU2=392243032 cycles
[ 305.008986] All test took CPU3=392353853 cycles
[ 305.108944] All test took CPU4=297630721 cycles
[ 305.196406] All test took CPU5=297548736 cycles
[ 305.288602] All test took CPU6=297092392 cycles
[ 305.381088] All test took CPU7=297293597 cycles
<patched>
~14%-23% patched variant is better.
Link: http://lkml.kernel.org/r/20191022155800.20468-1-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Acked-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:47 +08:00
|
|
|
spin_unlock(&free_vmap_area_lock);
|
2016-12-13 08:44:07 +08:00
|
|
|
return true;
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
}
|
|
|
|
|
2008-11-20 07:36:33 +08:00
|
|
|
/*
|
|
|
|
* Kick off a purge of the outstanding lazy areas. Don't bother if somebody
|
|
|
|
* is already purging.
|
|
|
|
*/
|
|
|
|
static void try_purge_vmap_area_lazy(void)
|
|
|
|
{
|
2016-12-13 08:44:23 +08:00
|
|
|
if (mutex_trylock(&vmap_purge_lock)) {
|
2016-12-13 08:44:07 +08:00
|
|
|
__purge_vmap_area_lazy(ULONG_MAX, 0);
|
2016-12-13 08:44:23 +08:00
|
|
|
mutex_unlock(&vmap_purge_lock);
|
2016-12-13 08:44:07 +08:00
|
|
|
}
|
2008-11-20 07:36:33 +08:00
|
|
|
}
|
|
|
|
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
/*
|
|
|
|
* Kick off a purge of the outstanding lazy areas.
|
|
|
|
*/
|
|
|
|
static void purge_vmap_area_lazy(void)
|
|
|
|
{
|
2016-12-13 08:44:23 +08:00
|
|
|
mutex_lock(&vmap_purge_lock);
|
2016-12-13 08:44:07 +08:00
|
|
|
purge_fragmented_blocks_allcpus();
|
|
|
|
__purge_vmap_area_lazy(ULONG_MAX, 0);
|
2016-12-13 08:44:23 +08:00
|
|
|
mutex_unlock(&vmap_purge_lock);
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
vmalloc: eagerly clear ptes on vunmap
On stock 2.6.37-rc4, running:
# mount lilith:/export /mnt/lilith
# find /mnt/lilith/ -type f -print0 | xargs -0 file
crashes the machine fairly quickly under Xen. Often it results in oops
messages, but the couple of times I tried just now, it just hung quietly
and made Xen print some rude messages:
(XEN) mm.c:2389:d80 Bad type (saw 7400000000000001 != exp
3000000000000000) for mfn 1d7058 (pfn 18fa7)
(XEN) mm.c:964:d80 Attempt to create linear p.t. with write perms
(XEN) mm.c:2389:d80 Bad type (saw 7400000000000010 != exp
1000000000000000) for mfn 1d2e04 (pfn 1d1fb)
(XEN) mm.c:2965:d80 Error while pinning mfn 1d2e04
Which means the domain tried to map a pagetable page RW, which would
allow it to map arbitrary memory, so Xen stopped it. This is because
vm_unmap_ram() left some pages mapped in the vmalloc area after NFS had
finished with them, and those pages got recycled as pagetable pages
while still having these RW aliases.
Removing those mappings immediately removes the Xen-visible aliases, and
so it has no problem with those pages being reused as pagetable pages.
Deferring the TLB flush doesn't upset Xen because it can flush the TLB
itself as needed to maintain its invariants.
When unmapping a region in the vmalloc space, clear the ptes
immediately. There's no point in deferring this because there's no
amortization benefit.
The TLBs are left dirty, and they are flushed lazily to amortize the
cost of the IPIs.
This specific motivation for this patch is an oops-causing regression
since 2.6.36 when using NFS under Xen, triggered by the NFS client's use
of vm_map_ram() introduced in 56e4ebf877b60 ("NFS: readdir with vmapped
pages") . XFS also uses vm_map_ram() and could cause similar problems.
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Cc: Nick Piggin <npiggin@kernel.dk>
Cc: Bryan Schumaker <bjschuma@netapp.com>
Cc: Trond Myklebust <Trond.Myklebust@netapp.com>
Cc: Alex Elder <aelder@sgi.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Christoph Hellwig <hch@lst.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-12-03 06:31:18 +08:00
|
|
|
* Free a vmap area, caller ensuring that the area has been unmapped
|
|
|
|
* and flush_cache_vunmap had been called for the correct range
|
|
|
|
* previously.
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
*/
|
vmalloc: eagerly clear ptes on vunmap
On stock 2.6.37-rc4, running:
# mount lilith:/export /mnt/lilith
# find /mnt/lilith/ -type f -print0 | xargs -0 file
crashes the machine fairly quickly under Xen. Often it results in oops
messages, but the couple of times I tried just now, it just hung quietly
and made Xen print some rude messages:
(XEN) mm.c:2389:d80 Bad type (saw 7400000000000001 != exp
3000000000000000) for mfn 1d7058 (pfn 18fa7)
(XEN) mm.c:964:d80 Attempt to create linear p.t. with write perms
(XEN) mm.c:2389:d80 Bad type (saw 7400000000000010 != exp
1000000000000000) for mfn 1d2e04 (pfn 1d1fb)
(XEN) mm.c:2965:d80 Error while pinning mfn 1d2e04
Which means the domain tried to map a pagetable page RW, which would
allow it to map arbitrary memory, so Xen stopped it. This is because
vm_unmap_ram() left some pages mapped in the vmalloc area after NFS had
finished with them, and those pages got recycled as pagetable pages
while still having these RW aliases.
Removing those mappings immediately removes the Xen-visible aliases, and
so it has no problem with those pages being reused as pagetable pages.
Deferring the TLB flush doesn't upset Xen because it can flush the TLB
itself as needed to maintain its invariants.
When unmapping a region in the vmalloc space, clear the ptes
immediately. There's no point in deferring this because there's no
amortization benefit.
The TLBs are left dirty, and they are flushed lazily to amortize the
cost of the IPIs.
This specific motivation for this patch is an oops-causing regression
since 2.6.36 when using NFS under Xen, triggered by the NFS client's use
of vm_map_ram() introduced in 56e4ebf877b60 ("NFS: readdir with vmapped
pages") . XFS also uses vm_map_ram() and could cause similar problems.
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Cc: Nick Piggin <npiggin@kernel.dk>
Cc: Bryan Schumaker <bjschuma@netapp.com>
Cc: Trond Myklebust <Trond.Myklebust@netapp.com>
Cc: Alex Elder <aelder@sgi.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Christoph Hellwig <hch@lst.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-12-03 06:31:18 +08:00
|
|
|
static void free_vmap_area_noflush(struct vmap_area *va)
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
{
|
2019-05-15 06:41:25 +08:00
|
|
|
unsigned long nr_lazy;
|
2016-05-21 07:57:38 +08:00
|
|
|
|
2019-09-24 06:36:36 +08:00
|
|
|
spin_lock(&vmap_area_lock);
|
|
|
|
unlink_va(va, &vmap_area_root);
|
|
|
|
spin_unlock(&vmap_area_lock);
|
|
|
|
|
2019-05-15 06:41:25 +08:00
|
|
|
nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >>
|
|
|
|
PAGE_SHIFT, &vmap_lazy_nr);
|
2016-05-21 07:57:38 +08:00
|
|
|
|
|
|
|
/* After this point, we may free va at any time */
|
|
|
|
llist_add(&va->purge_list, &vmap_purge_list);
|
|
|
|
|
|
|
|
if (unlikely(nr_lazy > lazy_max_pages()))
|
2008-11-20 07:36:33 +08:00
|
|
|
try_purge_vmap_area_lazy();
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
}
|
|
|
|
|
2008-12-02 05:13:47 +08:00
|
|
|
/*
|
|
|
|
* Free and unmap a vmap area
|
|
|
|
*/
|
|
|
|
static void free_unmap_vmap_area(struct vmap_area *va)
|
|
|
|
{
|
|
|
|
flush_cache_vunmap(va->va_start, va->va_end);
|
2016-12-13 08:44:01 +08:00
|
|
|
unmap_vmap_area(va);
|
mm, debug_pagealloc: don't rely on static keys too early
Commit 96a2b03f281d ("mm, debug_pagelloc: use static keys to enable
debugging") has introduced a static key to reduce overhead when
debug_pagealloc is compiled in but not enabled. It relied on the
assumption that jump_label_init() is called before parse_early_param()
as in start_kernel(), so when the "debug_pagealloc=on" option is parsed,
it is safe to enable the static key.
However, it turns out multiple architectures call parse_early_param()
earlier from their setup_arch(). x86 also calls jump_label_init() even
earlier, so no issue was found while testing the commit, but same is not
true for e.g. ppc64 and s390 where the kernel would not boot with
debug_pagealloc=on as found by our QA.
To fix this without tricky changes to init code of multiple
architectures, this patch partially reverts the static key conversion
from 96a2b03f281d. Init-time and non-fastpath calls (such as in arch
code) of debug_pagealloc_enabled() will again test a simple bool
variable. Fastpath mm code is converted to a new
debug_pagealloc_enabled_static() variant that relies on the static key,
which is enabled in a well-defined point in mm_init() where it's
guaranteed that jump_label_init() has been called, regardless of
architecture.
[sfr@canb.auug.org.au: export _debug_pagealloc_enabled_early]
Link: http://lkml.kernel.org/r/20200106164944.063ac07b@canb.auug.org.au
Link: http://lkml.kernel.org/r/20191219130612.23171-1-vbabka@suse.cz
Fixes: 96a2b03f281d ("mm, debug_pagelloc: use static keys to enable debugging")
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Qian Cai <cai@lca.pw>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-01-14 08:29:20 +08:00
|
|
|
if (debug_pagealloc_enabled_static())
|
2018-06-08 08:06:46 +08:00
|
|
|
flush_tlb_kernel_range(va->va_start, va->va_end);
|
|
|
|
|
2016-12-13 08:44:01 +08:00
|
|
|
free_vmap_area_noflush(va);
|
2008-12-02 05:13:47 +08:00
|
|
|
}
|
|
|
|
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
static struct vmap_area *find_vmap_area(unsigned long addr)
|
|
|
|
{
|
|
|
|
struct vmap_area *va;
|
|
|
|
|
|
|
|
spin_lock(&vmap_area_lock);
|
|
|
|
va = __find_vmap_area(addr);
|
|
|
|
spin_unlock(&vmap_area_lock);
|
|
|
|
|
|
|
|
return va;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*** Per cpu kva allocator ***/
|
|
|
|
|
|
|
|
/*
|
|
|
|
* vmap space is limited especially on 32 bit architectures. Ensure there is
|
|
|
|
* room for at least 16 percpu vmap blocks per CPU.
|
|
|
|
*/
|
|
|
|
/*
|
|
|
|
* If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
|
|
|
|
* to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
|
|
|
|
* instead (we just need a rough idea)
|
|
|
|
*/
|
|
|
|
#if BITS_PER_LONG == 32
|
|
|
|
#define VMALLOC_SPACE (128UL*1024*1024)
|
|
|
|
#else
|
|
|
|
#define VMALLOC_SPACE (128UL*1024*1024*1024)
|
|
|
|
#endif
|
|
|
|
|
|
|
|
#define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
|
|
|
|
#define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
|
|
|
|
#define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
|
|
|
|
#define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
|
|
|
|
#define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
|
|
|
|
#define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
|
2011-06-22 04:09:50 +08:00
|
|
|
#define VMAP_BBMAP_BITS \
|
|
|
|
VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
|
|
|
|
VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
|
|
|
|
VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
|
|
|
|
#define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
|
|
|
|
|
|
|
|
struct vmap_block_queue {
|
|
|
|
spinlock_t lock;
|
|
|
|
struct list_head free;
|
|
|
|
};
|
|
|
|
|
|
|
|
struct vmap_block {
|
|
|
|
spinlock_t lock;
|
|
|
|
struct vmap_area *va;
|
|
|
|
unsigned long free, dirty;
|
mm/vmalloc: get rid of dirty bitmap inside vmap_block structure
In original implementation of vm_map_ram made by Nick Piggin there were
two bitmaps: alloc_map and dirty_map. None of them were used as supposed
to be: finding a suitable free hole for next allocation in block.
vm_map_ram allocates space sequentially in block and on free call marks
pages as dirty, so freed space can't be reused anymore.
Actually it would be very interesting to know the real meaning of those
bitmaps, maybe implementation was incomplete, etc.
But long time ago Zhang Yanfei removed alloc_map by these two commits:
mm/vmalloc.c: remove dead code in vb_alloc
3fcd76e8028e0be37b02a2002b4f56755daeda06
mm/vmalloc.c: remove alloc_map from vmap_block
b8e748b6c32999f221ea4786557b8e7e6c4e4e7a
In this patch I replaced dirty_map with two range variables: dirty min and
max. These variables store minimum and maximum position of dirty space in
a block, since we need only to know the dirty range, not exact position of
dirty pages.
Why it was made? Several reasons: at first glance it seems that
vm_map_ram allocator concerns about fragmentation thus it uses bitmaps for
finding free hole, but it is not true. To avoid complexity seems it is
better to use something simple, like min or max range values. Secondly,
code also becomes simpler, without iteration over bitmap, just comparing
values in min and max macros. Thirdly, bitmap occupies up to 1024 bits
(4MB is a max size of a block). Here I replaced the whole bitmap with two
longs.
Finally vm_unmap_aliases should be slightly faster and the whole
vmap_block structure occupies less memory.
Signed-off-by: Roman Pen <r.peniaev@gmail.com>
Cc: Zhang Yanfei <zhangyanfei@cn.fujitsu.com>
Cc: Eric Dumazet <edumazet@google.com>
Acked-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: David Rientjes <rientjes@google.com>
Cc: WANG Chao <chaowang@redhat.com>
Cc: Fabian Frederick <fabf@skynet.be>
Cc: Christoph Lameter <cl@linux.com>
Cc: Gioh Kim <gioh.kim@lge.com>
Cc: Rob Jones <rob.jones@codethink.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-04-16 07:13:55 +08:00
|
|
|
unsigned long dirty_min, dirty_max; /*< dirty range */
|
2010-02-01 19:24:18 +08:00
|
|
|
struct list_head free_list;
|
|
|
|
struct rcu_head rcu_head;
|
2010-02-01 19:25:57 +08:00
|
|
|
struct list_head purge;
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
};
|
|
|
|
|
|
|
|
/* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
|
|
|
|
static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
|
|
|
|
* in the free path. Could get rid of this if we change the API to return a
|
|
|
|
* "cookie" from alloc, to be passed to free. But no big deal yet.
|
|
|
|
*/
|
|
|
|
static DEFINE_SPINLOCK(vmap_block_tree_lock);
|
|
|
|
static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* We should probably have a fallback mechanism to allocate virtual memory
|
|
|
|
* out of partially filled vmap blocks. However vmap block sizing should be
|
|
|
|
* fairly reasonable according to the vmalloc size, so it shouldn't be a
|
|
|
|
* big problem.
|
|
|
|
*/
|
|
|
|
|
|
|
|
static unsigned long addr_to_vb_idx(unsigned long addr)
|
|
|
|
{
|
|
|
|
addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
|
|
|
|
addr /= VMAP_BLOCK_SIZE;
|
|
|
|
return addr;
|
|
|
|
}
|
|
|
|
|
2015-04-16 07:13:52 +08:00
|
|
|
static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
|
|
|
|
{
|
|
|
|
unsigned long addr;
|
|
|
|
|
|
|
|
addr = va_start + (pages_off << PAGE_SHIFT);
|
|
|
|
BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
|
|
|
|
return (void *)addr;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
|
|
|
|
* block. Of course pages number can't exceed VMAP_BBMAP_BITS
|
|
|
|
* @order: how many 2^order pages should be occupied in newly allocated block
|
|
|
|
* @gfp_mask: flags for the page level allocator
|
|
|
|
*
|
2019-03-06 07:48:42 +08:00
|
|
|
* Return: virtual address in a newly allocated block or ERR_PTR(-errno)
|
2015-04-16 07:13:52 +08:00
|
|
|
*/
|
|
|
|
static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
{
|
|
|
|
struct vmap_block_queue *vbq;
|
|
|
|
struct vmap_block *vb;
|
|
|
|
struct vmap_area *va;
|
|
|
|
unsigned long vb_idx;
|
|
|
|
int node, err;
|
2015-04-16 07:13:52 +08:00
|
|
|
void *vaddr;
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
|
|
|
|
node = numa_node_id();
|
|
|
|
|
|
|
|
vb = kmalloc_node(sizeof(struct vmap_block),
|
|
|
|
gfp_mask & GFP_RECLAIM_MASK, node);
|
|
|
|
if (unlikely(!vb))
|
|
|
|
return ERR_PTR(-ENOMEM);
|
|
|
|
|
|
|
|
va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
|
|
|
|
VMALLOC_START, VMALLOC_END,
|
|
|
|
node, gfp_mask);
|
2011-01-14 07:46:15 +08:00
|
|
|
if (IS_ERR(va)) {
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
kfree(vb);
|
2010-08-10 08:18:28 +08:00
|
|
|
return ERR_CAST(va);
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
err = radix_tree_preload(gfp_mask);
|
|
|
|
if (unlikely(err)) {
|
|
|
|
kfree(vb);
|
|
|
|
free_vmap_area(va);
|
|
|
|
return ERR_PTR(err);
|
|
|
|
}
|
|
|
|
|
2015-04-16 07:13:52 +08:00
|
|
|
vaddr = vmap_block_vaddr(va->va_start, 0);
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
spin_lock_init(&vb->lock);
|
|
|
|
vb->va = va;
|
2015-04-16 07:13:52 +08:00
|
|
|
/* At least something should be left free */
|
|
|
|
BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
|
|
|
|
vb->free = VMAP_BBMAP_BITS - (1UL << order);
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
vb->dirty = 0;
|
mm/vmalloc: get rid of dirty bitmap inside vmap_block structure
In original implementation of vm_map_ram made by Nick Piggin there were
two bitmaps: alloc_map and dirty_map. None of them were used as supposed
to be: finding a suitable free hole for next allocation in block.
vm_map_ram allocates space sequentially in block and on free call marks
pages as dirty, so freed space can't be reused anymore.
Actually it would be very interesting to know the real meaning of those
bitmaps, maybe implementation was incomplete, etc.
But long time ago Zhang Yanfei removed alloc_map by these two commits:
mm/vmalloc.c: remove dead code in vb_alloc
3fcd76e8028e0be37b02a2002b4f56755daeda06
mm/vmalloc.c: remove alloc_map from vmap_block
b8e748b6c32999f221ea4786557b8e7e6c4e4e7a
In this patch I replaced dirty_map with two range variables: dirty min and
max. These variables store minimum and maximum position of dirty space in
a block, since we need only to know the dirty range, not exact position of
dirty pages.
Why it was made? Several reasons: at first glance it seems that
vm_map_ram allocator concerns about fragmentation thus it uses bitmaps for
finding free hole, but it is not true. To avoid complexity seems it is
better to use something simple, like min or max range values. Secondly,
code also becomes simpler, without iteration over bitmap, just comparing
values in min and max macros. Thirdly, bitmap occupies up to 1024 bits
(4MB is a max size of a block). Here I replaced the whole bitmap with two
longs.
Finally vm_unmap_aliases should be slightly faster and the whole
vmap_block structure occupies less memory.
Signed-off-by: Roman Pen <r.peniaev@gmail.com>
Cc: Zhang Yanfei <zhangyanfei@cn.fujitsu.com>
Cc: Eric Dumazet <edumazet@google.com>
Acked-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: David Rientjes <rientjes@google.com>
Cc: WANG Chao <chaowang@redhat.com>
Cc: Fabian Frederick <fabf@skynet.be>
Cc: Christoph Lameter <cl@linux.com>
Cc: Gioh Kim <gioh.kim@lge.com>
Cc: Rob Jones <rob.jones@codethink.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-04-16 07:13:55 +08:00
|
|
|
vb->dirty_min = VMAP_BBMAP_BITS;
|
|
|
|
vb->dirty_max = 0;
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
INIT_LIST_HEAD(&vb->free_list);
|
|
|
|
|
|
|
|
vb_idx = addr_to_vb_idx(va->va_start);
|
|
|
|
spin_lock(&vmap_block_tree_lock);
|
|
|
|
err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
|
|
|
|
spin_unlock(&vmap_block_tree_lock);
|
|
|
|
BUG_ON(err);
|
|
|
|
radix_tree_preload_end();
|
|
|
|
|
|
|
|
vbq = &get_cpu_var(vmap_block_queue);
|
|
|
|
spin_lock(&vbq->lock);
|
mm/vmalloc: fix possible exhaustion of vmalloc space caused by vm_map_ram allocator
Recently I came across high fragmentation of vm_map_ram allocator:
vmap_block has free space, but still new blocks continue to appear.
Further investigation showed that certain mapping/unmapping sequences
can exhaust vmalloc space. On small 32bit systems that's not a big
problem, cause purging will be called soon on a first allocation failure
(alloc_vmap_area), but on 64bit machines, e.g. x86_64 has 45 bits of
vmalloc space, that can be a disaster.
1) I came up with a simple allocation sequence, which exhausts virtual
space very quickly:
while (iters) {
/* Map/unmap big chunk */
vaddr = vm_map_ram(pages, 16, -1, PAGE_KERNEL);
vm_unmap_ram(vaddr, 16);
/* Map/unmap small chunks.
*
* -1 for hole, which should be left at the end of each block
* to keep it partially used, with some free space available */
for (i = 0; i < (VMAP_BBMAP_BITS - 16) / 8 - 1; i++) {
vaddr = vm_map_ram(pages, 8, -1, PAGE_KERNEL);
vm_unmap_ram(vaddr, 8);
}
}
The idea behind is simple:
1. We have to map a big chunk, e.g. 16 pages.
2. Then we have to occupy the remaining space with smaller chunks, i.e.
8 pages. At the end small hole should remain to keep block in free list,
but do not let big chunk to occupy remaining space.
3. Goto 1 - allocation request of 16 pages can't be completed (only 8 slots
are left free in the block in the #2 step), new block will be allocated,
all further requests will lay into newly allocated block.
To have some measurement numbers for all further tests I setup ftrace and
enabled 4 basic calls in a function profile:
echo vm_map_ram > /sys/kernel/debug/tracing/set_ftrace_filter;
echo alloc_vmap_area >> /sys/kernel/debug/tracing/set_ftrace_filter;
echo vm_unmap_ram >> /sys/kernel/debug/tracing/set_ftrace_filter;
echo free_vmap_block >> /sys/kernel/debug/tracing/set_ftrace_filter;
So for this scenario I got these results:
BEFORE (all new blocks are put to the head of a free list)
# cat /sys/kernel/debug/tracing/trace_stat/function0
Function Hit Time Avg s^2
-------- --- ---- --- ---
vm_map_ram 126000 30683.30 us 0.243 us 30819.36 us
vm_unmap_ram 126000 22003.24 us 0.174 us 340.886 us
alloc_vmap_area 1000 4132.065 us 4.132 us 0.903 us
AFTER (all new blocks are put to the tail of a free list)
# cat /sys/kernel/debug/tracing/trace_stat/function0
Function Hit Time Avg s^2
-------- --- ---- --- ---
vm_map_ram 126000 28713.13 us 0.227 us 24944.70 us
vm_unmap_ram 126000 20403.96 us 0.161 us 1429.872 us
alloc_vmap_area 993 3916.795 us 3.944 us 29.370 us
free_vmap_block 992 654.157 us 0.659 us 1.273 us
SUMMARY:
The most interesting numbers in those tables are numbers of block
allocations and deallocations: alloc_vmap_area and free_vmap_block
calls, which show that before the change blocks were not freed, and
virtual space and physical memory (vmap_block structure allocations,
etc) were consumed.
Average time which were spent in vm_map_ram/vm_unmap_ram became slightly
better. That can be explained with a reasonable amount of blocks in a
free list, which we need to iterate to find a suitable free block.
2) Another scenario is a random allocation:
while (iters) {
/* Randomly take number from a range [1..32/64] */
nr = rand(1, VMAP_MAX_ALLOC);
vaddr = vm_map_ram(pages, nr, -1, PAGE_KERNEL);
vm_unmap_ram(vaddr, nr);
}
I chose mersenne twister PRNG to generate persistent random state to
guarantee that both runs have the same random sequence. For each
vm_map_ram call random number from [1..32/64] was taken to represent
amount of pages which I do map.
I did 10'000 vm_map_ram calls and got these two tables:
BEFORE (all new blocks are put to the head of a free list)
# cat /sys/kernel/debug/tracing/trace_stat/function0
Function Hit Time Avg s^2
-------- --- ---- --- ---
vm_map_ram 10000 10170.01 us 1.017 us 993.609 us
vm_unmap_ram 10000 5321.823 us 0.532 us 59.789 us
alloc_vmap_area 420 2150.239 us 5.119 us 3.307 us
free_vmap_block 37 159.587 us 4.313 us 134.344 us
AFTER (all new blocks are put to the tail of a free list)
# cat /sys/kernel/debug/tracing/trace_stat/function0
Function Hit Time Avg s^2
-------- --- ---- --- ---
vm_map_ram 10000 7745.637 us 0.774 us 395.229 us
vm_unmap_ram 10000 5460.573 us 0.546 us 67.187 us
alloc_vmap_area 414 2201.650 us 5.317 us 5.591 us
free_vmap_block 412 574.421 us 1.394 us 15.138 us
SUMMARY:
'BEFORE' table shows, that 420 blocks were allocated and only 37 were
freed. Remained 383 blocks are still in a free list, consuming virtual
space and physical memory.
'AFTER' table shows, that 414 blocks were allocated and 412 were really
freed. 2 blocks remained in a free list.
So fragmentation was dramatically reduced. Why? Because when we put
newly allocated block to the head, all further requests will occupy new
block, regardless remained space in other blocks. In this scenario all
requests come randomly. Eventually remained free space will be less
than requested size, free list will be iterated and it is possible that
nothing will be found there - finally new block will be created. So
exhaustion in random scenario happens for the maximum possible
allocation size: 32 pages for 32-bit system and 64 pages for 64-bit
system.
Also average cost of vm_map_ram was reduced from 1.017 us to 0.774 us.
Again this can be explained by iteration through smaller list of free
blocks.
3) Next simple scenario is a sequential allocation, when the allocation
order is increased for each block. This scenario forces allocator to
reach maximum amount of partially free blocks in a free list:
while (iters) {
/* Populate free list with blocks with remaining space */
for (order = 0; order <= ilog2(VMAP_MAX_ALLOC); order++) {
nr = VMAP_BBMAP_BITS / (1 << order);
/* Leave a hole */
nr -= 1;
for (i = 0; i < nr; i++) {
vaddr = vm_map_ram(pages, (1 << order), -1, PAGE_KERNEL);
vm_unmap_ram(vaddr, (1 << order));
}
/* Completely occupy blocks from a free list */
for (order = 0; order <= ilog2(VMAP_MAX_ALLOC); order++) {
vaddr = vm_map_ram(pages, (1 << order), -1, PAGE_KERNEL);
vm_unmap_ram(vaddr, (1 << order));
}
}
Results which I got:
BEFORE (all new blocks are put to the head of a free list)
# cat /sys/kernel/debug/tracing/trace_stat/function0
Function Hit Time Avg s^2
-------- --- ---- --- ---
vm_map_ram 2032000 399545.2 us 0.196 us 467123.7 us
vm_unmap_ram 2032000 363225.7 us 0.178 us 111405.9 us
alloc_vmap_area 7001 30627.76 us 4.374 us 495.755 us
free_vmap_block 6993 7011.685 us 1.002 us 159.090 us
AFTER (all new blocks are put to the tail of a free list)
# cat /sys/kernel/debug/tracing/trace_stat/function0
Function Hit Time Avg s^2
-------- --- ---- --- ---
vm_map_ram 2032000 394259.7 us 0.194 us 589395.9 us
vm_unmap_ram 2032000 292500.7 us 0.143 us 94181.08 us
alloc_vmap_area 7000 31103.11 us 4.443 us 703.225 us
free_vmap_block 7000 6750.844 us 0.964 us 119.112 us
SUMMARY:
No surprises here, almost all numbers are the same.
Fixing this fragmentation problem I also did some improvements in a
allocation logic of a new vmap block: occupy block immediately and get
rid of extra search in a free list.
Also I replaced dirty bitmap with min/max dirty range values to make the
logic simpler and slightly faster, since two longs comparison costs
less, than loop thru bitmap.
This patchset raises several questions:
Q: Think the problem you comments is already known so that I wrote comments
about it as "it could consume lots of address space through fragmentation".
Could you tell me about your situation and reason why it should be avoided?
Gioh Kim
A: Indeed, there was a commit 364376383 which adds explicit comment about
fragmentation. But fragmentation which is described in this comment caused
by mixing of long-lived and short-lived objects, when a whole block is pinned
in memory because some page slots are still in use. But here I am talking
about blocks which are free, nobody uses them, and allocator keeps them alive
forever, continuously allocating new blocks.
Q: I think that if you put newly allocated block to the tail of a free
list, below example would results in enormous performance degradation.
new block: 1MB (256 pages)
while (iters--) {
vm_map_ram(3 or something else not dividable for 256) * 85
vm_unmap_ram(3) * 85
}
On every iteration, it needs newly allocated block and it is put to the
tail of a free list so finding it consumes large amount of time.
Joonsoo Kim
A: Second patch in current patchset gets rid of extra search in a free list,
so new block will be immediately occupied..
Also, the scenario above is impossible, cause vm_map_ram allocates virtual
range in orders, i.e. 2^n. I.e. passing 3 to vm_map_ram you will allocate
4 slots in a block and 256 slots (capacity of a block) of course dividable
on 4, so block will be completely occupied.
But there is a worst case which we can achieve: each free block has a hole
equal to order size.
The maximum size of allocation is 64 pages for 64-bit system
(if you try to map more, original alloc_vmap_area will be called).
So the maximum order is 6. That means that worst case, before allocator
makes a decision to allocate a new block, is to iterate 7 blocks:
HEAD
1st block - has 1 page slot free (order 0)
2nd block - has 2 page slots free (order 1)
3rd block - has 4 page slots free (order 2)
4th block - has 8 page slots free (order 3)
5th block - has 16 page slots free (order 4)
6th block - has 32 page slots free (order 5)
7th block - has 64 page slots free (order 6)
TAIL
So the worst scenario on 64-bit system is that each CPU queue can have 7
blocks in a free list.
This can happen only and only if you allocate blocks increasing the order.
(as I did in the function written in the comment of the first patch)
This is weird and rare case, but still it is possible. Afterwards you will
get 7 blocks in a list.
All further requests should be placed in a newly allocated block or some
free slots should be found in a free list.
Seems it does not look dramatically awful.
This patch (of 3):
If suitable block can't be found, new block is allocated and put into a
head of a free list, so on next iteration this new block will be found
first.
That's bad, because old blocks in a free list will not get a chance to be
fully used, thus fragmentation will grow.
Let's consider this simple example:
#1 We have one block in a free list which is partially used, and where only
one page is free:
HEAD |xxxxxxxxx-| TAIL
^
free space for 1 page, order 0
#2 New allocation request of order 1 (2 pages) comes, new block is allocated
since we do not have free space to complete this request. New block is put
into a head of a free list:
HEAD |----------|xxxxxxxxx-| TAIL
#3 Two pages were occupied in a new found block:
HEAD |xx--------|xxxxxxxxx-| TAIL
^
two pages mapped here
#4 New allocation request of order 0 (1 page) comes. Block, which was created
on #2 step, is located at the beginning of a free list, so it will be found
first:
HEAD |xxX-------|xxxxxxxxx-| TAIL
^ ^
page mapped here, but better to use this hole
It is obvious, that it is better to complete request of #4 step using the
old block, where free space is left, because in other case fragmentation
will be highly increased.
But fragmentation is not only the case. The worst thing is that I can
easily create scenario, when the whole vmalloc space is exhausted by
blocks, which are not used, but already dirty and have several free pages.
Let's consider this function which execution should be pinned to one CPU:
static void exhaust_virtual_space(struct page *pages[16], int iters)
{
/* Firstly we have to map a big chunk, e.g. 16 pages.
* Then we have to occupy the remaining space with smaller
* chunks, i.e. 8 pages. At the end small hole should remain.
* So at the end of our allocation sequence block looks like
* this:
* XX big chunk
* |XXxxxxxxx-| x small chunk
* - hole, which is enough for a small chunk,
* but is not enough for a big chunk
*/
while (iters--) {
int i;
void *vaddr;
/* Map/unmap big chunk */
vaddr = vm_map_ram(pages, 16, -1, PAGE_KERNEL);
vm_unmap_ram(vaddr, 16);
/* Map/unmap small chunks.
*
* -1 for hole, which should be left at the end of each block
* to keep it partially used, with some free space available */
for (i = 0; i < (VMAP_BBMAP_BITS - 16) / 8 - 1; i++) {
vaddr = vm_map_ram(pages, 8, -1, PAGE_KERNEL);
vm_unmap_ram(vaddr, 8);
}
}
}
On every iteration new block (1MB of vm area in my case) will be
allocated and then will be occupied, without attempt to resolve small
allocation request using previously allocated blocks in a free list.
In case of random allocation (size should be randomly taken from the
range [1..64] in 64-bit case or [1..32] in 32-bit case) situation is the
same: new blocks continue to appear if maximum possible allocation size
(32 or 64) passed to the allocator, because all remaining blocks in a
free list do not have enough free space to complete this allocation
request.
In summary if new blocks are put into the head of a free list eventually
virtual space will be exhausted.
In current patch I simply put newly allocated block to the tail of a
free list, thus reduce fragmentation, giving a chance to resolve
allocation request using older blocks with possible holes left.
Signed-off-by: Roman Pen <r.peniaev@gmail.com>
Cc: Eric Dumazet <edumazet@google.com>
Acked-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: David Rientjes <rientjes@google.com>
Cc: WANG Chao <chaowang@redhat.com>
Cc: Fabian Frederick <fabf@skynet.be>
Cc: Christoph Lameter <cl@linux.com>
Cc: Gioh Kim <gioh.kim@lge.com>
Cc: Rob Jones <rob.jones@codethink.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-04-16 07:13:48 +08:00
|
|
|
list_add_tail_rcu(&vb->free_list, &vbq->free);
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
spin_unlock(&vbq->lock);
|
2009-10-29 21:34:12 +08:00
|
|
|
put_cpu_var(vmap_block_queue);
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
|
2015-04-16 07:13:52 +08:00
|
|
|
return vaddr;
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
static void free_vmap_block(struct vmap_block *vb)
|
|
|
|
{
|
|
|
|
struct vmap_block *tmp;
|
|
|
|
unsigned long vb_idx;
|
|
|
|
|
|
|
|
vb_idx = addr_to_vb_idx(vb->va->va_start);
|
|
|
|
spin_lock(&vmap_block_tree_lock);
|
|
|
|
tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
|
|
|
|
spin_unlock(&vmap_block_tree_lock);
|
|
|
|
BUG_ON(tmp != vb);
|
|
|
|
|
vmalloc: eagerly clear ptes on vunmap
On stock 2.6.37-rc4, running:
# mount lilith:/export /mnt/lilith
# find /mnt/lilith/ -type f -print0 | xargs -0 file
crashes the machine fairly quickly under Xen. Often it results in oops
messages, but the couple of times I tried just now, it just hung quietly
and made Xen print some rude messages:
(XEN) mm.c:2389:d80 Bad type (saw 7400000000000001 != exp
3000000000000000) for mfn 1d7058 (pfn 18fa7)
(XEN) mm.c:964:d80 Attempt to create linear p.t. with write perms
(XEN) mm.c:2389:d80 Bad type (saw 7400000000000010 != exp
1000000000000000) for mfn 1d2e04 (pfn 1d1fb)
(XEN) mm.c:2965:d80 Error while pinning mfn 1d2e04
Which means the domain tried to map a pagetable page RW, which would
allow it to map arbitrary memory, so Xen stopped it. This is because
vm_unmap_ram() left some pages mapped in the vmalloc area after NFS had
finished with them, and those pages got recycled as pagetable pages
while still having these RW aliases.
Removing those mappings immediately removes the Xen-visible aliases, and
so it has no problem with those pages being reused as pagetable pages.
Deferring the TLB flush doesn't upset Xen because it can flush the TLB
itself as needed to maintain its invariants.
When unmapping a region in the vmalloc space, clear the ptes
immediately. There's no point in deferring this because there's no
amortization benefit.
The TLBs are left dirty, and they are flushed lazily to amortize the
cost of the IPIs.
This specific motivation for this patch is an oops-causing regression
since 2.6.36 when using NFS under Xen, triggered by the NFS client's use
of vm_map_ram() introduced in 56e4ebf877b60 ("NFS: readdir with vmapped
pages") . XFS also uses vm_map_ram() and could cause similar problems.
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Cc: Nick Piggin <npiggin@kernel.dk>
Cc: Bryan Schumaker <bjschuma@netapp.com>
Cc: Trond Myklebust <Trond.Myklebust@netapp.com>
Cc: Alex Elder <aelder@sgi.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Christoph Hellwig <hch@lst.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-12-03 06:31:18 +08:00
|
|
|
free_vmap_area_noflush(vb->va);
|
2011-03-18 12:13:08 +08:00
|
|
|
kfree_rcu(vb, rcu_head);
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
}
|
|
|
|
|
2010-02-01 19:25:57 +08:00
|
|
|
static void purge_fragmented_blocks(int cpu)
|
|
|
|
{
|
|
|
|
LIST_HEAD(purge);
|
|
|
|
struct vmap_block *vb;
|
|
|
|
struct vmap_block *n_vb;
|
|
|
|
struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
|
|
|
|
|
|
|
|
rcu_read_lock();
|
|
|
|
list_for_each_entry_rcu(vb, &vbq->free, free_list) {
|
|
|
|
|
|
|
|
if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
spin_lock(&vb->lock);
|
|
|
|
if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
|
|
|
|
vb->free = 0; /* prevent further allocs after releasing lock */
|
|
|
|
vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
|
mm/vmalloc: get rid of dirty bitmap inside vmap_block structure
In original implementation of vm_map_ram made by Nick Piggin there were
two bitmaps: alloc_map and dirty_map. None of them were used as supposed
to be: finding a suitable free hole for next allocation in block.
vm_map_ram allocates space sequentially in block and on free call marks
pages as dirty, so freed space can't be reused anymore.
Actually it would be very interesting to know the real meaning of those
bitmaps, maybe implementation was incomplete, etc.
But long time ago Zhang Yanfei removed alloc_map by these two commits:
mm/vmalloc.c: remove dead code in vb_alloc
3fcd76e8028e0be37b02a2002b4f56755daeda06
mm/vmalloc.c: remove alloc_map from vmap_block
b8e748b6c32999f221ea4786557b8e7e6c4e4e7a
In this patch I replaced dirty_map with two range variables: dirty min and
max. These variables store minimum and maximum position of dirty space in
a block, since we need only to know the dirty range, not exact position of
dirty pages.
Why it was made? Several reasons: at first glance it seems that
vm_map_ram allocator concerns about fragmentation thus it uses bitmaps for
finding free hole, but it is not true. To avoid complexity seems it is
better to use something simple, like min or max range values. Secondly,
code also becomes simpler, without iteration over bitmap, just comparing
values in min and max macros. Thirdly, bitmap occupies up to 1024 bits
(4MB is a max size of a block). Here I replaced the whole bitmap with two
longs.
Finally vm_unmap_aliases should be slightly faster and the whole
vmap_block structure occupies less memory.
Signed-off-by: Roman Pen <r.peniaev@gmail.com>
Cc: Zhang Yanfei <zhangyanfei@cn.fujitsu.com>
Cc: Eric Dumazet <edumazet@google.com>
Acked-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: David Rientjes <rientjes@google.com>
Cc: WANG Chao <chaowang@redhat.com>
Cc: Fabian Frederick <fabf@skynet.be>
Cc: Christoph Lameter <cl@linux.com>
Cc: Gioh Kim <gioh.kim@lge.com>
Cc: Rob Jones <rob.jones@codethink.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-04-16 07:13:55 +08:00
|
|
|
vb->dirty_min = 0;
|
|
|
|
vb->dirty_max = VMAP_BBMAP_BITS;
|
2010-02-01 19:25:57 +08:00
|
|
|
spin_lock(&vbq->lock);
|
|
|
|
list_del_rcu(&vb->free_list);
|
|
|
|
spin_unlock(&vbq->lock);
|
|
|
|
spin_unlock(&vb->lock);
|
|
|
|
list_add_tail(&vb->purge, &purge);
|
|
|
|
} else
|
|
|
|
spin_unlock(&vb->lock);
|
|
|
|
}
|
|
|
|
rcu_read_unlock();
|
|
|
|
|
|
|
|
list_for_each_entry_safe(vb, n_vb, &purge, purge) {
|
|
|
|
list_del(&vb->purge);
|
|
|
|
free_vmap_block(vb);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static void purge_fragmented_blocks_allcpus(void)
|
|
|
|
{
|
|
|
|
int cpu;
|
|
|
|
|
|
|
|
for_each_possible_cpu(cpu)
|
|
|
|
purge_fragmented_blocks(cpu);
|
|
|
|
}
|
|
|
|
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
|
|
|
|
{
|
|
|
|
struct vmap_block_queue *vbq;
|
|
|
|
struct vmap_block *vb;
|
2015-04-16 07:13:52 +08:00
|
|
|
void *vaddr = NULL;
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
unsigned int order;
|
|
|
|
|
2015-11-06 10:46:51 +08:00
|
|
|
BUG_ON(offset_in_page(size));
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
|
2012-08-01 07:41:37 +08:00
|
|
|
if (WARN_ON(size == 0)) {
|
|
|
|
/*
|
|
|
|
* Allocating 0 bytes isn't what caller wants since
|
|
|
|
* get_order(0) returns funny result. Just warn and terminate
|
|
|
|
* early.
|
|
|
|
*/
|
|
|
|
return NULL;
|
|
|
|
}
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
order = get_order(size);
|
|
|
|
|
|
|
|
rcu_read_lock();
|
|
|
|
vbq = &get_cpu_var(vmap_block_queue);
|
|
|
|
list_for_each_entry_rcu(vb, &vbq->free, free_list) {
|
2015-04-16 07:13:52 +08:00
|
|
|
unsigned long pages_off;
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
|
|
|
|
spin_lock(&vb->lock);
|
2015-04-16 07:13:52 +08:00
|
|
|
if (vb->free < (1UL << order)) {
|
|
|
|
spin_unlock(&vb->lock);
|
|
|
|
continue;
|
|
|
|
}
|
2010-02-01 19:25:57 +08:00
|
|
|
|
2015-04-16 07:13:52 +08:00
|
|
|
pages_off = VMAP_BBMAP_BITS - vb->free;
|
|
|
|
vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
|
2010-02-01 19:25:57 +08:00
|
|
|
vb->free -= 1UL << order;
|
|
|
|
if (vb->free == 0) {
|
|
|
|
spin_lock(&vbq->lock);
|
|
|
|
list_del_rcu(&vb->free_list);
|
|
|
|
spin_unlock(&vbq->lock);
|
|
|
|
}
|
2015-04-16 07:13:52 +08:00
|
|
|
|
2010-02-01 19:25:57 +08:00
|
|
|
spin_unlock(&vb->lock);
|
|
|
|
break;
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
}
|
2010-02-01 19:25:57 +08:00
|
|
|
|
2009-10-29 21:34:12 +08:00
|
|
|
put_cpu_var(vmap_block_queue);
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
rcu_read_unlock();
|
|
|
|
|
2015-04-16 07:13:52 +08:00
|
|
|
/* Allocate new block if nothing was found */
|
|
|
|
if (!vaddr)
|
|
|
|
vaddr = new_vmap_block(order, gfp_mask);
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
|
2015-04-16 07:13:52 +08:00
|
|
|
return vaddr;
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
static void vb_free(const void *addr, unsigned long size)
|
|
|
|
{
|
|
|
|
unsigned long offset;
|
|
|
|
unsigned long vb_idx;
|
|
|
|
unsigned int order;
|
|
|
|
struct vmap_block *vb;
|
|
|
|
|
2015-11-06 10:46:51 +08:00
|
|
|
BUG_ON(offset_in_page(size));
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
|
2008-12-02 05:13:47 +08:00
|
|
|
|
|
|
|
flush_cache_vunmap((unsigned long)addr, (unsigned long)addr + size);
|
|
|
|
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
order = get_order(size);
|
|
|
|
|
|
|
|
offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
|
mm/vmalloc: get rid of dirty bitmap inside vmap_block structure
In original implementation of vm_map_ram made by Nick Piggin there were
two bitmaps: alloc_map and dirty_map. None of them were used as supposed
to be: finding a suitable free hole for next allocation in block.
vm_map_ram allocates space sequentially in block and on free call marks
pages as dirty, so freed space can't be reused anymore.
Actually it would be very interesting to know the real meaning of those
bitmaps, maybe implementation was incomplete, etc.
But long time ago Zhang Yanfei removed alloc_map by these two commits:
mm/vmalloc.c: remove dead code in vb_alloc
3fcd76e8028e0be37b02a2002b4f56755daeda06
mm/vmalloc.c: remove alloc_map from vmap_block
b8e748b6c32999f221ea4786557b8e7e6c4e4e7a
In this patch I replaced dirty_map with two range variables: dirty min and
max. These variables store minimum and maximum position of dirty space in
a block, since we need only to know the dirty range, not exact position of
dirty pages.
Why it was made? Several reasons: at first glance it seems that
vm_map_ram allocator concerns about fragmentation thus it uses bitmaps for
finding free hole, but it is not true. To avoid complexity seems it is
better to use something simple, like min or max range values. Secondly,
code also becomes simpler, without iteration over bitmap, just comparing
values in min and max macros. Thirdly, bitmap occupies up to 1024 bits
(4MB is a max size of a block). Here I replaced the whole bitmap with two
longs.
Finally vm_unmap_aliases should be slightly faster and the whole
vmap_block structure occupies less memory.
Signed-off-by: Roman Pen <r.peniaev@gmail.com>
Cc: Zhang Yanfei <zhangyanfei@cn.fujitsu.com>
Cc: Eric Dumazet <edumazet@google.com>
Acked-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: David Rientjes <rientjes@google.com>
Cc: WANG Chao <chaowang@redhat.com>
Cc: Fabian Frederick <fabf@skynet.be>
Cc: Christoph Lameter <cl@linux.com>
Cc: Gioh Kim <gioh.kim@lge.com>
Cc: Rob Jones <rob.jones@codethink.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-04-16 07:13:55 +08:00
|
|
|
offset >>= PAGE_SHIFT;
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
|
|
|
|
vb_idx = addr_to_vb_idx((unsigned long)addr);
|
|
|
|
rcu_read_lock();
|
|
|
|
vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
|
|
|
|
rcu_read_unlock();
|
|
|
|
BUG_ON(!vb);
|
|
|
|
|
vmalloc: eagerly clear ptes on vunmap
On stock 2.6.37-rc4, running:
# mount lilith:/export /mnt/lilith
# find /mnt/lilith/ -type f -print0 | xargs -0 file
crashes the machine fairly quickly under Xen. Often it results in oops
messages, but the couple of times I tried just now, it just hung quietly
and made Xen print some rude messages:
(XEN) mm.c:2389:d80 Bad type (saw 7400000000000001 != exp
3000000000000000) for mfn 1d7058 (pfn 18fa7)
(XEN) mm.c:964:d80 Attempt to create linear p.t. with write perms
(XEN) mm.c:2389:d80 Bad type (saw 7400000000000010 != exp
1000000000000000) for mfn 1d2e04 (pfn 1d1fb)
(XEN) mm.c:2965:d80 Error while pinning mfn 1d2e04
Which means the domain tried to map a pagetable page RW, which would
allow it to map arbitrary memory, so Xen stopped it. This is because
vm_unmap_ram() left some pages mapped in the vmalloc area after NFS had
finished with them, and those pages got recycled as pagetable pages
while still having these RW aliases.
Removing those mappings immediately removes the Xen-visible aliases, and
so it has no problem with those pages being reused as pagetable pages.
Deferring the TLB flush doesn't upset Xen because it can flush the TLB
itself as needed to maintain its invariants.
When unmapping a region in the vmalloc space, clear the ptes
immediately. There's no point in deferring this because there's no
amortization benefit.
The TLBs are left dirty, and they are flushed lazily to amortize the
cost of the IPIs.
This specific motivation for this patch is an oops-causing regression
since 2.6.36 when using NFS under Xen, triggered by the NFS client's use
of vm_map_ram() introduced in 56e4ebf877b60 ("NFS: readdir with vmapped
pages") . XFS also uses vm_map_ram() and could cause similar problems.
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Cc: Nick Piggin <npiggin@kernel.dk>
Cc: Bryan Schumaker <bjschuma@netapp.com>
Cc: Trond Myklebust <Trond.Myklebust@netapp.com>
Cc: Alex Elder <aelder@sgi.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Christoph Hellwig <hch@lst.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-12-03 06:31:18 +08:00
|
|
|
vunmap_page_range((unsigned long)addr, (unsigned long)addr + size);
|
|
|
|
|
mm, debug_pagealloc: don't rely on static keys too early
Commit 96a2b03f281d ("mm, debug_pagelloc: use static keys to enable
debugging") has introduced a static key to reduce overhead when
debug_pagealloc is compiled in but not enabled. It relied on the
assumption that jump_label_init() is called before parse_early_param()
as in start_kernel(), so when the "debug_pagealloc=on" option is parsed,
it is safe to enable the static key.
However, it turns out multiple architectures call parse_early_param()
earlier from their setup_arch(). x86 also calls jump_label_init() even
earlier, so no issue was found while testing the commit, but same is not
true for e.g. ppc64 and s390 where the kernel would not boot with
debug_pagealloc=on as found by our QA.
To fix this without tricky changes to init code of multiple
architectures, this patch partially reverts the static key conversion
from 96a2b03f281d. Init-time and non-fastpath calls (such as in arch
code) of debug_pagealloc_enabled() will again test a simple bool
variable. Fastpath mm code is converted to a new
debug_pagealloc_enabled_static() variant that relies on the static key,
which is enabled in a well-defined point in mm_init() where it's
guaranteed that jump_label_init() has been called, regardless of
architecture.
[sfr@canb.auug.org.au: export _debug_pagealloc_enabled_early]
Link: http://lkml.kernel.org/r/20200106164944.063ac07b@canb.auug.org.au
Link: http://lkml.kernel.org/r/20191219130612.23171-1-vbabka@suse.cz
Fixes: 96a2b03f281d ("mm, debug_pagelloc: use static keys to enable debugging")
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Qian Cai <cai@lca.pw>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-01-14 08:29:20 +08:00
|
|
|
if (debug_pagealloc_enabled_static())
|
2018-06-08 08:06:46 +08:00
|
|
|
flush_tlb_kernel_range((unsigned long)addr,
|
|
|
|
(unsigned long)addr + size);
|
|
|
|
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
spin_lock(&vb->lock);
|
mm/vmalloc: get rid of dirty bitmap inside vmap_block structure
In original implementation of vm_map_ram made by Nick Piggin there were
two bitmaps: alloc_map and dirty_map. None of them were used as supposed
to be: finding a suitable free hole for next allocation in block.
vm_map_ram allocates space sequentially in block and on free call marks
pages as dirty, so freed space can't be reused anymore.
Actually it would be very interesting to know the real meaning of those
bitmaps, maybe implementation was incomplete, etc.
But long time ago Zhang Yanfei removed alloc_map by these two commits:
mm/vmalloc.c: remove dead code in vb_alloc
3fcd76e8028e0be37b02a2002b4f56755daeda06
mm/vmalloc.c: remove alloc_map from vmap_block
b8e748b6c32999f221ea4786557b8e7e6c4e4e7a
In this patch I replaced dirty_map with two range variables: dirty min and
max. These variables store minimum and maximum position of dirty space in
a block, since we need only to know the dirty range, not exact position of
dirty pages.
Why it was made? Several reasons: at first glance it seems that
vm_map_ram allocator concerns about fragmentation thus it uses bitmaps for
finding free hole, but it is not true. To avoid complexity seems it is
better to use something simple, like min or max range values. Secondly,
code also becomes simpler, without iteration over bitmap, just comparing
values in min and max macros. Thirdly, bitmap occupies up to 1024 bits
(4MB is a max size of a block). Here I replaced the whole bitmap with two
longs.
Finally vm_unmap_aliases should be slightly faster and the whole
vmap_block structure occupies less memory.
Signed-off-by: Roman Pen <r.peniaev@gmail.com>
Cc: Zhang Yanfei <zhangyanfei@cn.fujitsu.com>
Cc: Eric Dumazet <edumazet@google.com>
Acked-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: David Rientjes <rientjes@google.com>
Cc: WANG Chao <chaowang@redhat.com>
Cc: Fabian Frederick <fabf@skynet.be>
Cc: Christoph Lameter <cl@linux.com>
Cc: Gioh Kim <gioh.kim@lge.com>
Cc: Rob Jones <rob.jones@codethink.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-04-16 07:13:55 +08:00
|
|
|
|
|
|
|
/* Expand dirty range */
|
|
|
|
vb->dirty_min = min(vb->dirty_min, offset);
|
|
|
|
vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
|
2009-04-01 06:19:26 +08:00
|
|
|
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
vb->dirty += 1UL << order;
|
|
|
|
if (vb->dirty == VMAP_BBMAP_BITS) {
|
2010-02-01 19:24:18 +08:00
|
|
|
BUG_ON(vb->free);
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
spin_unlock(&vb->lock);
|
|
|
|
free_vmap_block(vb);
|
|
|
|
} else
|
|
|
|
spin_unlock(&vb->lock);
|
|
|
|
}
|
|
|
|
|
2019-04-26 08:11:36 +08:00
|
|
|
static void _vm_unmap_aliases(unsigned long start, unsigned long end, int flush)
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
{
|
|
|
|
int cpu;
|
|
|
|
|
2008-10-28 16:22:34 +08:00
|
|
|
if (unlikely(!vmap_initialized))
|
|
|
|
return;
|
|
|
|
|
2016-12-13 08:44:20 +08:00
|
|
|
might_sleep();
|
|
|
|
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
for_each_possible_cpu(cpu) {
|
|
|
|
struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
|
|
|
|
struct vmap_block *vb;
|
|
|
|
|
|
|
|
rcu_read_lock();
|
|
|
|
list_for_each_entry_rcu(vb, &vbq->free, free_list) {
|
|
|
|
spin_lock(&vb->lock);
|
mm/vmalloc: get rid of dirty bitmap inside vmap_block structure
In original implementation of vm_map_ram made by Nick Piggin there were
two bitmaps: alloc_map and dirty_map. None of them were used as supposed
to be: finding a suitable free hole for next allocation in block.
vm_map_ram allocates space sequentially in block and on free call marks
pages as dirty, so freed space can't be reused anymore.
Actually it would be very interesting to know the real meaning of those
bitmaps, maybe implementation was incomplete, etc.
But long time ago Zhang Yanfei removed alloc_map by these two commits:
mm/vmalloc.c: remove dead code in vb_alloc
3fcd76e8028e0be37b02a2002b4f56755daeda06
mm/vmalloc.c: remove alloc_map from vmap_block
b8e748b6c32999f221ea4786557b8e7e6c4e4e7a
In this patch I replaced dirty_map with two range variables: dirty min and
max. These variables store minimum and maximum position of dirty space in
a block, since we need only to know the dirty range, not exact position of
dirty pages.
Why it was made? Several reasons: at first glance it seems that
vm_map_ram allocator concerns about fragmentation thus it uses bitmaps for
finding free hole, but it is not true. To avoid complexity seems it is
better to use something simple, like min or max range values. Secondly,
code also becomes simpler, without iteration over bitmap, just comparing
values in min and max macros. Thirdly, bitmap occupies up to 1024 bits
(4MB is a max size of a block). Here I replaced the whole bitmap with two
longs.
Finally vm_unmap_aliases should be slightly faster and the whole
vmap_block structure occupies less memory.
Signed-off-by: Roman Pen <r.peniaev@gmail.com>
Cc: Zhang Yanfei <zhangyanfei@cn.fujitsu.com>
Cc: Eric Dumazet <edumazet@google.com>
Acked-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: David Rientjes <rientjes@google.com>
Cc: WANG Chao <chaowang@redhat.com>
Cc: Fabian Frederick <fabf@skynet.be>
Cc: Christoph Lameter <cl@linux.com>
Cc: Gioh Kim <gioh.kim@lge.com>
Cc: Rob Jones <rob.jones@codethink.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-04-16 07:13:55 +08:00
|
|
|
if (vb->dirty) {
|
|
|
|
unsigned long va_start = vb->va->va_start;
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
unsigned long s, e;
|
2013-09-12 05:21:40 +08:00
|
|
|
|
mm/vmalloc: get rid of dirty bitmap inside vmap_block structure
In original implementation of vm_map_ram made by Nick Piggin there were
two bitmaps: alloc_map and dirty_map. None of them were used as supposed
to be: finding a suitable free hole for next allocation in block.
vm_map_ram allocates space sequentially in block and on free call marks
pages as dirty, so freed space can't be reused anymore.
Actually it would be very interesting to know the real meaning of those
bitmaps, maybe implementation was incomplete, etc.
But long time ago Zhang Yanfei removed alloc_map by these two commits:
mm/vmalloc.c: remove dead code in vb_alloc
3fcd76e8028e0be37b02a2002b4f56755daeda06
mm/vmalloc.c: remove alloc_map from vmap_block
b8e748b6c32999f221ea4786557b8e7e6c4e4e7a
In this patch I replaced dirty_map with two range variables: dirty min and
max. These variables store minimum and maximum position of dirty space in
a block, since we need only to know the dirty range, not exact position of
dirty pages.
Why it was made? Several reasons: at first glance it seems that
vm_map_ram allocator concerns about fragmentation thus it uses bitmaps for
finding free hole, but it is not true. To avoid complexity seems it is
better to use something simple, like min or max range values. Secondly,
code also becomes simpler, without iteration over bitmap, just comparing
values in min and max macros. Thirdly, bitmap occupies up to 1024 bits
(4MB is a max size of a block). Here I replaced the whole bitmap with two
longs.
Finally vm_unmap_aliases should be slightly faster and the whole
vmap_block structure occupies less memory.
Signed-off-by: Roman Pen <r.peniaev@gmail.com>
Cc: Zhang Yanfei <zhangyanfei@cn.fujitsu.com>
Cc: Eric Dumazet <edumazet@google.com>
Acked-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: David Rientjes <rientjes@google.com>
Cc: WANG Chao <chaowang@redhat.com>
Cc: Fabian Frederick <fabf@skynet.be>
Cc: Christoph Lameter <cl@linux.com>
Cc: Gioh Kim <gioh.kim@lge.com>
Cc: Rob Jones <rob.jones@codethink.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-04-16 07:13:55 +08:00
|
|
|
s = va_start + (vb->dirty_min << PAGE_SHIFT);
|
|
|
|
e = va_start + (vb->dirty_max << PAGE_SHIFT);
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
|
mm/vmalloc: get rid of dirty bitmap inside vmap_block structure
In original implementation of vm_map_ram made by Nick Piggin there were
two bitmaps: alloc_map and dirty_map. None of them were used as supposed
to be: finding a suitable free hole for next allocation in block.
vm_map_ram allocates space sequentially in block and on free call marks
pages as dirty, so freed space can't be reused anymore.
Actually it would be very interesting to know the real meaning of those
bitmaps, maybe implementation was incomplete, etc.
But long time ago Zhang Yanfei removed alloc_map by these two commits:
mm/vmalloc.c: remove dead code in vb_alloc
3fcd76e8028e0be37b02a2002b4f56755daeda06
mm/vmalloc.c: remove alloc_map from vmap_block
b8e748b6c32999f221ea4786557b8e7e6c4e4e7a
In this patch I replaced dirty_map with two range variables: dirty min and
max. These variables store minimum and maximum position of dirty space in
a block, since we need only to know the dirty range, not exact position of
dirty pages.
Why it was made? Several reasons: at first glance it seems that
vm_map_ram allocator concerns about fragmentation thus it uses bitmaps for
finding free hole, but it is not true. To avoid complexity seems it is
better to use something simple, like min or max range values. Secondly,
code also becomes simpler, without iteration over bitmap, just comparing
values in min and max macros. Thirdly, bitmap occupies up to 1024 bits
(4MB is a max size of a block). Here I replaced the whole bitmap with two
longs.
Finally vm_unmap_aliases should be slightly faster and the whole
vmap_block structure occupies less memory.
Signed-off-by: Roman Pen <r.peniaev@gmail.com>
Cc: Zhang Yanfei <zhangyanfei@cn.fujitsu.com>
Cc: Eric Dumazet <edumazet@google.com>
Acked-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: David Rientjes <rientjes@google.com>
Cc: WANG Chao <chaowang@redhat.com>
Cc: Fabian Frederick <fabf@skynet.be>
Cc: Christoph Lameter <cl@linux.com>
Cc: Gioh Kim <gioh.kim@lge.com>
Cc: Rob Jones <rob.jones@codethink.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-04-16 07:13:55 +08:00
|
|
|
start = min(s, start);
|
|
|
|
end = max(e, end);
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
|
mm/vmalloc: get rid of dirty bitmap inside vmap_block structure
In original implementation of vm_map_ram made by Nick Piggin there were
two bitmaps: alloc_map and dirty_map. None of them were used as supposed
to be: finding a suitable free hole for next allocation in block.
vm_map_ram allocates space sequentially in block and on free call marks
pages as dirty, so freed space can't be reused anymore.
Actually it would be very interesting to know the real meaning of those
bitmaps, maybe implementation was incomplete, etc.
But long time ago Zhang Yanfei removed alloc_map by these two commits:
mm/vmalloc.c: remove dead code in vb_alloc
3fcd76e8028e0be37b02a2002b4f56755daeda06
mm/vmalloc.c: remove alloc_map from vmap_block
b8e748b6c32999f221ea4786557b8e7e6c4e4e7a
In this patch I replaced dirty_map with two range variables: dirty min and
max. These variables store minimum and maximum position of dirty space in
a block, since we need only to know the dirty range, not exact position of
dirty pages.
Why it was made? Several reasons: at first glance it seems that
vm_map_ram allocator concerns about fragmentation thus it uses bitmaps for
finding free hole, but it is not true. To avoid complexity seems it is
better to use something simple, like min or max range values. Secondly,
code also becomes simpler, without iteration over bitmap, just comparing
values in min and max macros. Thirdly, bitmap occupies up to 1024 bits
(4MB is a max size of a block). Here I replaced the whole bitmap with two
longs.
Finally vm_unmap_aliases should be slightly faster and the whole
vmap_block structure occupies less memory.
Signed-off-by: Roman Pen <r.peniaev@gmail.com>
Cc: Zhang Yanfei <zhangyanfei@cn.fujitsu.com>
Cc: Eric Dumazet <edumazet@google.com>
Acked-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: David Rientjes <rientjes@google.com>
Cc: WANG Chao <chaowang@redhat.com>
Cc: Fabian Frederick <fabf@skynet.be>
Cc: Christoph Lameter <cl@linux.com>
Cc: Gioh Kim <gioh.kim@lge.com>
Cc: Rob Jones <rob.jones@codethink.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-04-16 07:13:55 +08:00
|
|
|
flush = 1;
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
}
|
|
|
|
spin_unlock(&vb->lock);
|
|
|
|
}
|
|
|
|
rcu_read_unlock();
|
|
|
|
}
|
|
|
|
|
2016-12-13 08:44:23 +08:00
|
|
|
mutex_lock(&vmap_purge_lock);
|
2016-12-13 08:44:07 +08:00
|
|
|
purge_fragmented_blocks_allcpus();
|
|
|
|
if (!__purge_vmap_area_lazy(start, end) && flush)
|
|
|
|
flush_tlb_kernel_range(start, end);
|
2016-12-13 08:44:23 +08:00
|
|
|
mutex_unlock(&vmap_purge_lock);
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
}
|
2019-04-26 08:11:36 +08:00
|
|
|
|
|
|
|
/**
|
|
|
|
* vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
|
|
|
|
*
|
|
|
|
* The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
|
|
|
|
* to amortize TLB flushing overheads. What this means is that any page you
|
|
|
|
* have now, may, in a former life, have been mapped into kernel virtual
|
|
|
|
* address by the vmap layer and so there might be some CPUs with TLB entries
|
|
|
|
* still referencing that page (additional to the regular 1:1 kernel mapping).
|
|
|
|
*
|
|
|
|
* vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
|
|
|
|
* be sure that none of the pages we have control over will have any aliases
|
|
|
|
* from the vmap layer.
|
|
|
|
*/
|
|
|
|
void vm_unmap_aliases(void)
|
|
|
|
{
|
|
|
|
unsigned long start = ULONG_MAX, end = 0;
|
|
|
|
int flush = 0;
|
|
|
|
|
|
|
|
_vm_unmap_aliases(start, end, flush);
|
|
|
|
}
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
EXPORT_SYMBOL_GPL(vm_unmap_aliases);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
|
|
|
|
* @mem: the pointer returned by vm_map_ram
|
|
|
|
* @count: the count passed to that vm_map_ram call (cannot unmap partial)
|
|
|
|
*/
|
|
|
|
void vm_unmap_ram(const void *mem, unsigned int count)
|
|
|
|
{
|
2016-06-04 05:55:33 +08:00
|
|
|
unsigned long size = (unsigned long)count << PAGE_SHIFT;
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
unsigned long addr = (unsigned long)mem;
|
2016-12-13 08:44:04 +08:00
|
|
|
struct vmap_area *va;
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
|
2016-12-13 08:44:20 +08:00
|
|
|
might_sleep();
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
BUG_ON(!addr);
|
|
|
|
BUG_ON(addr < VMALLOC_START);
|
|
|
|
BUG_ON(addr > VMALLOC_END);
|
2016-03-18 05:20:37 +08:00
|
|
|
BUG_ON(!PAGE_ALIGNED(addr));
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
|
2019-12-18 12:51:38 +08:00
|
|
|
kasan_poison_vmalloc(mem, size);
|
|
|
|
|
2016-12-13 08:44:04 +08:00
|
|
|
if (likely(count <= VMAP_MAX_ALLOC)) {
|
2018-06-08 08:06:53 +08:00
|
|
|
debug_check_no_locks_freed(mem, size);
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
vb_free(mem, size);
|
2016-12-13 08:44:04 +08:00
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
va = find_vmap_area(addr);
|
|
|
|
BUG_ON(!va);
|
2018-06-08 08:06:53 +08:00
|
|
|
debug_check_no_locks_freed((void *)va->va_start,
|
|
|
|
(va->va_end - va->va_start));
|
2016-12-13 08:44:04 +08:00
|
|
|
free_unmap_vmap_area(va);
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(vm_unmap_ram);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
|
|
|
|
* @pages: an array of pointers to the pages to be mapped
|
|
|
|
* @count: number of pages
|
|
|
|
* @node: prefer to allocate data structures on this node
|
|
|
|
* @prot: memory protection to use. PAGE_KERNEL for regular RAM
|
2008-10-30 05:01:09 +08:00
|
|
|
*
|
2014-04-08 06:37:37 +08:00
|
|
|
* If you use this function for less than VMAP_MAX_ALLOC pages, it could be
|
|
|
|
* faster than vmap so it's good. But if you mix long-life and short-life
|
|
|
|
* objects with vm_map_ram(), it could consume lots of address space through
|
|
|
|
* fragmentation (especially on a 32bit machine). You could see failures in
|
|
|
|
* the end. Please use this function for short-lived objects.
|
|
|
|
*
|
2008-10-30 05:01:09 +08:00
|
|
|
* Returns: a pointer to the address that has been mapped, or %NULL on failure
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
*/
|
|
|
|
void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
|
|
|
|
{
|
2016-06-04 05:55:33 +08:00
|
|
|
unsigned long size = (unsigned long)count << PAGE_SHIFT;
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
unsigned long addr;
|
|
|
|
void *mem;
|
|
|
|
|
|
|
|
if (likely(count <= VMAP_MAX_ALLOC)) {
|
|
|
|
mem = vb_alloc(size, GFP_KERNEL);
|
|
|
|
if (IS_ERR(mem))
|
|
|
|
return NULL;
|
|
|
|
addr = (unsigned long)mem;
|
|
|
|
} else {
|
|
|
|
struct vmap_area *va;
|
|
|
|
va = alloc_vmap_area(size, PAGE_SIZE,
|
|
|
|
VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
|
|
|
|
if (IS_ERR(va))
|
|
|
|
return NULL;
|
|
|
|
|
|
|
|
addr = va->va_start;
|
|
|
|
mem = (void *)addr;
|
|
|
|
}
|
2019-12-18 12:51:38 +08:00
|
|
|
|
|
|
|
kasan_unpoison_vmalloc(mem, size);
|
|
|
|
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
|
|
|
|
vm_unmap_ram(mem, count);
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
return mem;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(vm_map_ram);
|
|
|
|
|
2013-04-30 06:07:39 +08:00
|
|
|
static struct vm_struct *vmlist __initdata;
|
2019-03-06 07:48:36 +08:00
|
|
|
|
2011-08-25 12:24:21 +08:00
|
|
|
/**
|
|
|
|
* vm_area_add_early - add vmap area early during boot
|
|
|
|
* @vm: vm_struct to add
|
|
|
|
*
|
|
|
|
* This function is used to add fixed kernel vm area to vmlist before
|
|
|
|
* vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
|
|
|
|
* should contain proper values and the other fields should be zero.
|
|
|
|
*
|
|
|
|
* DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
|
|
|
|
*/
|
|
|
|
void __init vm_area_add_early(struct vm_struct *vm)
|
|
|
|
{
|
|
|
|
struct vm_struct *tmp, **p;
|
|
|
|
|
|
|
|
BUG_ON(vmap_initialized);
|
|
|
|
for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
|
|
|
|
if (tmp->addr >= vm->addr) {
|
|
|
|
BUG_ON(tmp->addr < vm->addr + vm->size);
|
|
|
|
break;
|
|
|
|
} else
|
|
|
|
BUG_ON(tmp->addr + tmp->size > vm->addr);
|
|
|
|
}
|
|
|
|
vm->next = *p;
|
|
|
|
*p = vm;
|
|
|
|
}
|
|
|
|
|
2009-02-20 15:29:08 +08:00
|
|
|
/**
|
|
|
|
* vm_area_register_early - register vmap area early during boot
|
|
|
|
* @vm: vm_struct to register
|
2009-02-24 10:57:21 +08:00
|
|
|
* @align: requested alignment
|
2009-02-20 15:29:08 +08:00
|
|
|
*
|
|
|
|
* This function is used to register kernel vm area before
|
|
|
|
* vmalloc_init() is called. @vm->size and @vm->flags should contain
|
|
|
|
* proper values on entry and other fields should be zero. On return,
|
|
|
|
* vm->addr contains the allocated address.
|
|
|
|
*
|
|
|
|
* DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
|
|
|
|
*/
|
2009-02-24 10:57:21 +08:00
|
|
|
void __init vm_area_register_early(struct vm_struct *vm, size_t align)
|
2009-02-20 15:29:08 +08:00
|
|
|
{
|
|
|
|
static size_t vm_init_off __initdata;
|
2009-02-24 10:57:21 +08:00
|
|
|
unsigned long addr;
|
|
|
|
|
|
|
|
addr = ALIGN(VMALLOC_START + vm_init_off, align);
|
|
|
|
vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
|
2009-02-20 15:29:08 +08:00
|
|
|
|
2009-02-24 10:57:21 +08:00
|
|
|
vm->addr = (void *)addr;
|
2009-02-20 15:29:08 +08:00
|
|
|
|
2011-08-25 12:24:21 +08:00
|
|
|
vm_area_add_early(vm);
|
2009-02-20 15:29:08 +08:00
|
|
|
}
|
|
|
|
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
static void vmap_init_free_space(void)
|
|
|
|
{
|
|
|
|
unsigned long vmap_start = 1;
|
|
|
|
const unsigned long vmap_end = ULONG_MAX;
|
|
|
|
struct vmap_area *busy, *free;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* B F B B B F
|
|
|
|
* -|-----|.....|-----|-----|-----|.....|-
|
|
|
|
* | The KVA space |
|
|
|
|
* |<--------------------------------->|
|
|
|
|
*/
|
|
|
|
list_for_each_entry(busy, &vmap_area_list, list) {
|
|
|
|
if (busy->va_start - vmap_start > 0) {
|
|
|
|
free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
|
|
|
|
if (!WARN_ON_ONCE(!free)) {
|
|
|
|
free->va_start = vmap_start;
|
|
|
|
free->va_end = busy->va_start;
|
|
|
|
|
|
|
|
insert_vmap_area_augment(free, NULL,
|
|
|
|
&free_vmap_area_root,
|
|
|
|
&free_vmap_area_list);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
vmap_start = busy->va_end;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (vmap_end - vmap_start > 0) {
|
|
|
|
free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
|
|
|
|
if (!WARN_ON_ONCE(!free)) {
|
|
|
|
free->va_start = vmap_start;
|
|
|
|
free->va_end = vmap_end;
|
|
|
|
|
|
|
|
insert_vmap_area_augment(free, NULL,
|
|
|
|
&free_vmap_area_root,
|
|
|
|
&free_vmap_area_list);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
void __init vmalloc_init(void)
|
|
|
|
{
|
2009-01-16 05:50:48 +08:00
|
|
|
struct vmap_area *va;
|
|
|
|
struct vm_struct *tmp;
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
int i;
|
|
|
|
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
/*
|
|
|
|
* Create the cache for vmap_area objects.
|
|
|
|
*/
|
|
|
|
vmap_area_cachep = KMEM_CACHE(vmap_area, SLAB_PANIC);
|
|
|
|
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
for_each_possible_cpu(i) {
|
|
|
|
struct vmap_block_queue *vbq;
|
2013-03-11 08:14:08 +08:00
|
|
|
struct vfree_deferred *p;
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
|
|
|
|
vbq = &per_cpu(vmap_block_queue, i);
|
|
|
|
spin_lock_init(&vbq->lock);
|
|
|
|
INIT_LIST_HEAD(&vbq->free);
|
2013-03-11 08:14:08 +08:00
|
|
|
p = &per_cpu(vfree_deferred, i);
|
|
|
|
init_llist_head(&p->list);
|
|
|
|
INIT_WORK(&p->wq, free_work);
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
}
|
2008-10-28 16:22:34 +08:00
|
|
|
|
2009-01-16 05:50:48 +08:00
|
|
|
/* Import existing vmlist entries. */
|
|
|
|
for (tmp = vmlist; tmp; tmp = tmp->next) {
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
va = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
|
|
|
|
if (WARN_ON_ONCE(!va))
|
|
|
|
continue;
|
|
|
|
|
2009-01-16 05:50:48 +08:00
|
|
|
va->va_start = (unsigned long)tmp->addr;
|
|
|
|
va->va_end = va->va_start + tmp->size;
|
2012-05-30 06:06:49 +08:00
|
|
|
va->vm = tmp;
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
|
2009-01-16 05:50:48 +08:00
|
|
|
}
|
2009-08-14 14:00:52 +08:00
|
|
|
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
/*
|
|
|
|
* Now we can initialize a free vmap space.
|
|
|
|
*/
|
|
|
|
vmap_init_free_space();
|
2008-10-28 16:22:34 +08:00
|
|
|
vmap_initialized = true;
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
}
|
|
|
|
|
2009-02-20 15:29:08 +08:00
|
|
|
/**
|
|
|
|
* map_kernel_range_noflush - map kernel VM area with the specified pages
|
|
|
|
* @addr: start of the VM area to map
|
|
|
|
* @size: size of the VM area to map
|
|
|
|
* @prot: page protection flags to use
|
|
|
|
* @pages: pages to map
|
|
|
|
*
|
|
|
|
* Map PFN_UP(@size) pages at @addr. The VM area @addr and @size
|
|
|
|
* specify should have been allocated using get_vm_area() and its
|
|
|
|
* friends.
|
|
|
|
*
|
|
|
|
* NOTE:
|
|
|
|
* This function does NOT do any cache flushing. The caller is
|
|
|
|
* responsible for calling flush_cache_vmap() on to-be-mapped areas
|
|
|
|
* before calling this function.
|
|
|
|
*
|
|
|
|
* RETURNS:
|
|
|
|
* The number of pages mapped on success, -errno on failure.
|
|
|
|
*/
|
|
|
|
int map_kernel_range_noflush(unsigned long addr, unsigned long size,
|
|
|
|
pgprot_t prot, struct page **pages)
|
|
|
|
{
|
|
|
|
return vmap_page_range_noflush(addr, addr + size, prot, pages);
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* unmap_kernel_range_noflush - unmap kernel VM area
|
|
|
|
* @addr: start of the VM area to unmap
|
|
|
|
* @size: size of the VM area to unmap
|
|
|
|
*
|
|
|
|
* Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size
|
|
|
|
* specify should have been allocated using get_vm_area() and its
|
|
|
|
* friends.
|
|
|
|
*
|
|
|
|
* NOTE:
|
|
|
|
* This function does NOT do any cache flushing. The caller is
|
|
|
|
* responsible for calling flush_cache_vunmap() on to-be-mapped areas
|
|
|
|
* before calling this function and flush_tlb_kernel_range() after.
|
|
|
|
*/
|
|
|
|
void unmap_kernel_range_noflush(unsigned long addr, unsigned long size)
|
|
|
|
{
|
|
|
|
vunmap_page_range(addr, addr + size);
|
|
|
|
}
|
ACPI, APEI, Generic Hardware Error Source POLL/IRQ/NMI notification type support
Generic Hardware Error Source provides a way to report platform
hardware errors (such as that from chipset). It works in so called
"Firmware First" mode, that is, hardware errors are reported to
firmware firstly, then reported to Linux by firmware. This way, some
non-standard hardware error registers or non-standard hardware link
can be checked by firmware to produce more valuable hardware error
information for Linux.
This patch adds POLL/IRQ/NMI notification types support.
Because the memory area used to transfer hardware error information
from BIOS to Linux can be determined only in NMI, IRQ or timer
handler, but general ioremap can not be used in atomic context, so a
special version of atomic ioremap is implemented for that.
Known issue:
- Error information can not be printed for recoverable errors notified
via NMI, because printk is not NMI-safe. Will fix this via delay
printing to IRQ context via irq_work or make printk NMI-safe.
v2:
- adjust printk format per comments.
Signed-off-by: Huang Ying <ying.huang@intel.com>
Reviewed-by: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Len Brown <len.brown@intel.com>
2011-01-12 14:44:55 +08:00
|
|
|
EXPORT_SYMBOL_GPL(unmap_kernel_range_noflush);
|
2009-02-20 15:29:08 +08:00
|
|
|
|
|
|
|
/**
|
|
|
|
* unmap_kernel_range - unmap kernel VM area and flush cache and TLB
|
|
|
|
* @addr: start of the VM area to unmap
|
|
|
|
* @size: size of the VM area to unmap
|
|
|
|
*
|
|
|
|
* Similar to unmap_kernel_range_noflush() but flushes vcache before
|
|
|
|
* the unmapping and tlb after.
|
|
|
|
*/
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
void unmap_kernel_range(unsigned long addr, unsigned long size)
|
|
|
|
{
|
|
|
|
unsigned long end = addr + size;
|
2009-02-21 07:38:48 +08:00
|
|
|
|
|
|
|
flush_cache_vunmap(addr, end);
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
vunmap_page_range(addr, end);
|
|
|
|
flush_tlb_kernel_range(addr, end);
|
|
|
|
}
|
2014-06-05 07:11:09 +08:00
|
|
|
EXPORT_SYMBOL_GPL(unmap_kernel_range);
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
|
2014-08-07 07:06:58 +08:00
|
|
|
int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page **pages)
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
{
|
|
|
|
unsigned long addr = (unsigned long)area->addr;
|
2013-09-12 05:22:42 +08:00
|
|
|
unsigned long end = addr + get_vm_area_size(area);
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
int err;
|
|
|
|
|
2014-08-07 07:06:58 +08:00
|
|
|
err = vmap_page_range(addr, end, prot, pages);
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
|
2014-08-07 07:06:58 +08:00
|
|
|
return err > 0 ? 0 : err;
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(map_vm_area);
|
|
|
|
|
mm/vmalloc: rework vmap_area_lock
With the new allocation approach introduced in the 5.2 kernel, it
becomes possible to get rid of one global spinlock. By doing that we
can further improve the KVA from the performance point of view.
Basically we can have two independent locks, one for allocation part and
another one for deallocation, because of two different entities: "free
data structures" and "busy data structures".
As a result, allocation/deallocation operations can still interfere
between each other in case of running simultaneously on different CPUs,
it means there is still dependency, but with two locks it becomes lower.
Summarizing:
- it reduces the high lock contention
- it allows to perform operations on "free" and "busy"
trees in parallel on different CPUs. Please note it
does not solve scalability issue.
Test results:
In order to evaluate this patch, we can run "vmalloc test driver" to see
how many CPU cycles it takes to complete all test cases running
sequentially. All online CPUs run it so it will cause a high lock
contention.
HiKey 960, ARM64, 8xCPUs, big.LITTLE:
<snip>
sudo ./test_vmalloc.sh sequential_test_order=1
<snip>
<default>
[ 390.950557] All test took CPU0=457126382 cycles
[ 391.046690] All test took CPU1=454763452 cycles
[ 391.128586] All test took CPU2=454539334 cycles
[ 391.222669] All test took CPU3=455649517 cycles
[ 391.313946] All test took CPU4=388272196 cycles
[ 391.410425] All test took CPU5=384036264 cycles
[ 391.492219] All test took CPU6=387432964 cycles
[ 391.578433] All test took CPU7=387201996 cycles
<default>
<patched>
[ 304.721224] All test took CPU0=391521310 cycles
[ 304.821219] All test took CPU1=393533002 cycles
[ 304.917120] All test took CPU2=392243032 cycles
[ 305.008986] All test took CPU3=392353853 cycles
[ 305.108944] All test took CPU4=297630721 cycles
[ 305.196406] All test took CPU5=297548736 cycles
[ 305.288602] All test took CPU6=297092392 cycles
[ 305.381088] All test took CPU7=297293597 cycles
<patched>
~14%-23% patched variant is better.
Link: http://lkml.kernel.org/r/20191022155800.20468-1-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Acked-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:47 +08:00
|
|
|
static inline void setup_vmalloc_vm_locked(struct vm_struct *vm,
|
|
|
|
struct vmap_area *va, unsigned long flags, const void *caller)
|
2009-08-14 14:00:52 +08:00
|
|
|
{
|
|
|
|
vm->flags = flags;
|
|
|
|
vm->addr = (void *)va->va_start;
|
|
|
|
vm->size = va->va_end - va->va_start;
|
|
|
|
vm->caller = caller;
|
2012-01-11 07:08:39 +08:00
|
|
|
va->vm = vm;
|
mm/vmalloc: rework vmap_area_lock
With the new allocation approach introduced in the 5.2 kernel, it
becomes possible to get rid of one global spinlock. By doing that we
can further improve the KVA from the performance point of view.
Basically we can have two independent locks, one for allocation part and
another one for deallocation, because of two different entities: "free
data structures" and "busy data structures".
As a result, allocation/deallocation operations can still interfere
between each other in case of running simultaneously on different CPUs,
it means there is still dependency, but with two locks it becomes lower.
Summarizing:
- it reduces the high lock contention
- it allows to perform operations on "free" and "busy"
trees in parallel on different CPUs. Please note it
does not solve scalability issue.
Test results:
In order to evaluate this patch, we can run "vmalloc test driver" to see
how many CPU cycles it takes to complete all test cases running
sequentially. All online CPUs run it so it will cause a high lock
contention.
HiKey 960, ARM64, 8xCPUs, big.LITTLE:
<snip>
sudo ./test_vmalloc.sh sequential_test_order=1
<snip>
<default>
[ 390.950557] All test took CPU0=457126382 cycles
[ 391.046690] All test took CPU1=454763452 cycles
[ 391.128586] All test took CPU2=454539334 cycles
[ 391.222669] All test took CPU3=455649517 cycles
[ 391.313946] All test took CPU4=388272196 cycles
[ 391.410425] All test took CPU5=384036264 cycles
[ 391.492219] All test took CPU6=387432964 cycles
[ 391.578433] All test took CPU7=387201996 cycles
<default>
<patched>
[ 304.721224] All test took CPU0=391521310 cycles
[ 304.821219] All test took CPU1=393533002 cycles
[ 304.917120] All test took CPU2=392243032 cycles
[ 305.008986] All test took CPU3=392353853 cycles
[ 305.108944] All test took CPU4=297630721 cycles
[ 305.196406] All test took CPU5=297548736 cycles
[ 305.288602] All test took CPU6=297092392 cycles
[ 305.381088] All test took CPU7=297293597 cycles
<patched>
~14%-23% patched variant is better.
Link: http://lkml.kernel.org/r/20191022155800.20468-1-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Acked-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:47 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
|
|
|
|
unsigned long flags, const void *caller)
|
|
|
|
{
|
|
|
|
spin_lock(&vmap_area_lock);
|
|
|
|
setup_vmalloc_vm_locked(vm, va, flags, caller);
|
2013-04-30 06:07:30 +08:00
|
|
|
spin_unlock(&vmap_area_lock);
|
2011-11-01 08:08:13 +08:00
|
|
|
}
|
2009-08-14 14:00:52 +08:00
|
|
|
|
2013-07-09 06:59:58 +08:00
|
|
|
static void clear_vm_uninitialized_flag(struct vm_struct *vm)
|
2011-11-01 08:08:13 +08:00
|
|
|
{
|
2013-04-30 06:07:35 +08:00
|
|
|
/*
|
2013-07-09 06:59:58 +08:00
|
|
|
* Before removing VM_UNINITIALIZED,
|
2013-04-30 06:07:35 +08:00
|
|
|
* we should make sure that vm has proper values.
|
|
|
|
* Pair with smp_rmb() in show_numa_info().
|
|
|
|
*/
|
|
|
|
smp_wmb();
|
2013-07-09 06:59:58 +08:00
|
|
|
vm->flags &= ~VM_UNINITIALIZED;
|
2009-08-14 14:00:52 +08:00
|
|
|
}
|
|
|
|
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
static struct vm_struct *__get_vm_area_node(unsigned long size,
|
2009-09-22 03:22:34 +08:00
|
|
|
unsigned long align, unsigned long flags, unsigned long start,
|
2012-04-13 18:32:09 +08:00
|
|
|
unsigned long end, int node, gfp_t gfp_mask, const void *caller)
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
{
|
2011-12-20 09:12:04 +08:00
|
|
|
struct vmap_area *va;
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
struct vm_struct *area;
|
2019-12-18 12:51:38 +08:00
|
|
|
unsigned long requested_size = size;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2006-10-29 01:38:34 +08:00
|
|
|
BUG_ON(in_interrupt());
|
2005-04-17 06:20:36 +08:00
|
|
|
size = PAGE_ALIGN(size);
|
2006-11-16 17:19:29 +08:00
|
|
|
if (unlikely(!size))
|
|
|
|
return NULL;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2016-10-08 07:57:26 +08:00
|
|
|
if (flags & VM_IOREMAP)
|
|
|
|
align = 1ul << clamp_t(int, get_count_order_long(size),
|
|
|
|
PAGE_SHIFT, IOREMAP_MAX_ORDER);
|
|
|
|
|
2009-08-14 14:00:52 +08:00
|
|
|
area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
|
2005-04-17 06:20:36 +08:00
|
|
|
if (unlikely(!area))
|
|
|
|
return NULL;
|
|
|
|
|
2015-02-14 06:40:03 +08:00
|
|
|
if (!(flags & VM_NO_GUARD))
|
|
|
|
size += PAGE_SIZE;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
|
|
|
|
if (IS_ERR(va)) {
|
|
|
|
kfree(area);
|
|
|
|
return NULL;
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
2019-12-18 12:51:38 +08:00
|
|
|
kasan_unpoison_vmalloc((void *)va->va_start, requested_size);
|
2011-11-01 08:08:13 +08:00
|
|
|
|
2019-12-18 12:51:38 +08:00
|
|
|
setup_vmalloc_vm(area, va, flags, caller);
|
kasan: support backing vmalloc space with real shadow memory
Patch series "kasan: support backing vmalloc space with real shadow
memory", v11.
Currently, vmalloc space is backed by the early shadow page. This means
that kasan is incompatible with VMAP_STACK.
This series provides a mechanism to back vmalloc space with real,
dynamically allocated memory. I have only wired up x86, because that's
the only currently supported arch I can work with easily, but it's very
easy to wire up other architectures, and it appears that there is some
work-in-progress code to do this on arm64 and s390.
This has been discussed before in the context of VMAP_STACK:
- https://bugzilla.kernel.org/show_bug.cgi?id=202009
- https://lkml.org/lkml/2018/7/22/198
- https://lkml.org/lkml/2019/7/19/822
In terms of implementation details:
Most mappings in vmalloc space are small, requiring less than a full
page of shadow space. Allocating a full shadow page per mapping would
therefore be wasteful. Furthermore, to ensure that different mappings
use different shadow pages, mappings would have to be aligned to
KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE.
Instead, share backing space across multiple mappings. Allocate a
backing page when a mapping in vmalloc space uses a particular page of
the shadow region. This page can be shared by other vmalloc mappings
later on.
We hook in to the vmap infrastructure to lazily clean up unused shadow
memory.
Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that:
- Turning on KASAN, inline instrumentation, without vmalloc, introuduces
a 4.1x-4.2x slowdown in vmalloc operations.
- Turning this on introduces the following slowdowns over KASAN:
* ~1.76x slower single-threaded (test_vmalloc.sh performance)
* ~2.18x slower when both cpus are performing operations
simultaneously (test_vmalloc.sh sequential_test_order=1)
This is unfortunate but given that this is a debug feature only, not the
end of the world. The benchmarks are also a stress-test for the vmalloc
subsystem: they're not indicative of an overall 2x slowdown!
This patch (of 4):
Hook into vmalloc and vmap, and dynamically allocate real shadow memory
to back the mappings.
Most mappings in vmalloc space are small, requiring less than a full
page of shadow space. Allocating a full shadow page per mapping would
therefore be wasteful. Furthermore, to ensure that different mappings
use different shadow pages, mappings would have to be aligned to
KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE.
Instead, share backing space across multiple mappings. Allocate a
backing page when a mapping in vmalloc space uses a particular page of
the shadow region. This page can be shared by other vmalloc mappings
later on.
We hook in to the vmap infrastructure to lazily clean up unused shadow
memory.
To avoid the difficulties around swapping mappings around, this code
expects that the part of the shadow region that covers the vmalloc space
will not be covered by the early shadow page, but will be left unmapped.
This will require changes in arch-specific code.
This allows KASAN with VMAP_STACK, and may be helpful for architectures
that do not have a separate module space (e.g. powerpc64, which I am
currently working on). It also allows relaxing the module alignment
back to PAGE_SIZE.
Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that:
- Turning on KASAN, inline instrumentation, without vmalloc, introuduces
a 4.1x-4.2x slowdown in vmalloc operations.
- Turning this on introduces the following slowdowns over KASAN:
* ~1.76x slower single-threaded (test_vmalloc.sh performance)
* ~2.18x slower when both cpus are performing operations
simultaneously (test_vmalloc.sh sequential_test_order=3D1)
This is unfortunate but given that this is a debug feature only, not the
end of the world.
The full benchmark results are:
Performance
No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN
fix_size_alloc_test 662004 11404956 17.23 19144610 28.92 1.68
full_fit_alloc_test 710950 12029752 16.92 13184651 18.55 1.10
long_busy_list_alloc_test 9431875 43990172 4.66 82970178 8.80 1.89
random_size_alloc_test 5033626 23061762 4.58 47158834 9.37 2.04
fix_align_alloc_test 1252514 15276910 12.20 31266116 24.96 2.05
random_size_align_alloc_te 1648501 14578321 8.84 25560052 15.51 1.75
align_shift_alloc_test 147 830 5.65 5692 38.72 6.86
pcpu_alloc_test 80732 125520 1.55 140864 1.74 1.12
Total Cycles 119240774314 763211341128 6.40 1390338696894 11.66 1.82
Sequential, 2 cpus
No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN
fix_size_alloc_test 1423150 14276550 10.03 27733022 19.49 1.94
full_fit_alloc_test 1754219 14722640 8.39 15030786 8.57 1.02
long_busy_list_alloc_test 11451858 52154973 4.55 107016027 9.34 2.05
random_size_alloc_test 5989020 26735276 4.46 68885923 11.50 2.58
fix_align_alloc_test 2050976 20166900 9.83 50491675 24.62 2.50
random_size_align_alloc_te 2858229 17971700 6.29 38730225 13.55 2.16
align_shift_alloc_test 405 6428 15.87 26253 64.82 4.08
pcpu_alloc_test 127183 151464 1.19 216263 1.70 1.43
Total Cycles 54181269392 308723699764 5.70 650772566394 12.01 2.11
fix_size_alloc_test 1420404 14289308 10.06 27790035 19.56 1.94
full_fit_alloc_test 1736145 14806234 8.53 15274301 8.80 1.03
long_busy_list_alloc_test 11404638 52270785 4.58 107550254 9.43 2.06
random_size_alloc_test 6017006 26650625 4.43 68696127 11.42 2.58
fix_align_alloc_test 2045504 20280985 9.91 50414862 24.65 2.49
random_size_align_alloc_te 2845338 17931018 6.30 38510276 13.53 2.15
align_shift_alloc_test 472 3760 7.97 9656 20.46 2.57
pcpu_alloc_test 118643 132732 1.12 146504 1.23 1.10
Total Cycles 54040011688 309102805492 5.72 651325675652 12.05 2.11
[dja@axtens.net: fixups]
Link: http://lkml.kernel.org/r/20191120052719.7201-1-dja@axtens.net
Link: https://bugzilla.kernel.org/show_bug.cgi?id=3D202009
Link: http://lkml.kernel.org/r/20191031093909.9228-2-dja@axtens.net
Signed-off-by: Mark Rutland <mark.rutland@arm.com> [shadow rework]
Signed-off-by: Daniel Axtens <dja@axtens.net>
Co-developed-by: Mark Rutland <mark.rutland@arm.com>
Acked-by: Vasily Gorbik <gor@linux.ibm.com>
Reviewed-by: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Christophe Leroy <christophe.leroy@c-s.fr>
Cc: Qian Cai <cai@lca.pw>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:50 +08:00
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
return area;
|
|
|
|
}
|
|
|
|
|
2005-10-30 09:15:41 +08:00
|
|
|
struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
|
|
|
|
unsigned long start, unsigned long end)
|
|
|
|
{
|
2013-02-23 08:35:36 +08:00
|
|
|
return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
|
|
|
|
GFP_KERNEL, __builtin_return_address(0));
|
2005-10-30 09:15:41 +08:00
|
|
|
}
|
2007-07-19 16:49:21 +08:00
|
|
|
EXPORT_SYMBOL_GPL(__get_vm_area);
|
2005-10-30 09:15:41 +08:00
|
|
|
|
2009-02-19 06:48:12 +08:00
|
|
|
struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
|
|
|
|
unsigned long start, unsigned long end,
|
2012-04-13 18:32:09 +08:00
|
|
|
const void *caller)
|
2009-02-19 06:48:12 +08:00
|
|
|
{
|
2013-02-23 08:35:36 +08:00
|
|
|
return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
|
|
|
|
GFP_KERNEL, caller);
|
2009-02-19 06:48:12 +08:00
|
|
|
}
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
/**
|
2019-03-06 07:48:36 +08:00
|
|
|
* get_vm_area - reserve a contiguous kernel virtual area
|
|
|
|
* @size: size of the area
|
|
|
|
* @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
|
2005-04-17 06:20:36 +08:00
|
|
|
*
|
2019-03-06 07:48:36 +08:00
|
|
|
* Search an area of @size in the kernel virtual mapping area,
|
|
|
|
* and reserved it for out purposes. Returns the area descriptor
|
|
|
|
* on success or %NULL on failure.
|
2019-03-06 07:48:42 +08:00
|
|
|
*
|
|
|
|
* Return: the area descriptor on success or %NULL on failure.
|
2005-04-17 06:20:36 +08:00
|
|
|
*/
|
|
|
|
struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
|
|
|
|
{
|
2009-09-22 03:22:34 +08:00
|
|
|
return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
|
2013-02-23 08:35:36 +08:00
|
|
|
NUMA_NO_NODE, GFP_KERNEL,
|
|
|
|
__builtin_return_address(0));
|
2008-04-28 17:12:42 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
|
2012-04-13 18:32:09 +08:00
|
|
|
const void *caller)
|
2008-04-28 17:12:42 +08:00
|
|
|
{
|
2009-09-22 03:22:34 +08:00
|
|
|
return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
|
2013-02-23 08:35:36 +08:00
|
|
|
NUMA_NO_NODE, GFP_KERNEL, caller);
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
2012-07-30 15:11:33 +08:00
|
|
|
/**
|
2019-03-06 07:48:36 +08:00
|
|
|
* find_vm_area - find a continuous kernel virtual area
|
|
|
|
* @addr: base address
|
2012-07-30 15:11:33 +08:00
|
|
|
*
|
2019-03-06 07:48:36 +08:00
|
|
|
* Search for the kernel VM area starting at @addr, and return it.
|
|
|
|
* It is up to the caller to do all required locking to keep the returned
|
|
|
|
* pointer valid.
|
2019-03-06 07:48:42 +08:00
|
|
|
*
|
|
|
|
* Return: pointer to the found area or %NULL on faulure
|
2012-07-30 15:11:33 +08:00
|
|
|
*/
|
|
|
|
struct vm_struct *find_vm_area(const void *addr)
|
2006-06-23 17:03:20 +08:00
|
|
|
{
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
struct vmap_area *va;
|
2006-06-23 17:03:20 +08:00
|
|
|
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
va = find_vmap_area((unsigned long)addr);
|
2019-09-24 06:36:39 +08:00
|
|
|
if (!va)
|
|
|
|
return NULL;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2019-09-24 06:36:39 +08:00
|
|
|
return va->vm;
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
2005-05-21 05:27:57 +08:00
|
|
|
/**
|
2019-03-06 07:48:36 +08:00
|
|
|
* remove_vm_area - find and remove a continuous kernel virtual area
|
|
|
|
* @addr: base address
|
2005-05-21 05:27:57 +08:00
|
|
|
*
|
2019-03-06 07:48:36 +08:00
|
|
|
* Search for the kernel VM area starting at @addr, and remove it.
|
|
|
|
* This function returns the found VM area, but using it is NOT safe
|
|
|
|
* on SMP machines, except for its size or flags.
|
2019-03-06 07:48:42 +08:00
|
|
|
*
|
|
|
|
* Return: pointer to the found area or %NULL on faulure
|
2005-05-21 05:27:57 +08:00
|
|
|
*/
|
2008-02-05 14:28:32 +08:00
|
|
|
struct vm_struct *remove_vm_area(const void *addr)
|
2005-05-21 05:27:57 +08:00
|
|
|
{
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
struct vmap_area *va;
|
|
|
|
|
2016-12-13 08:44:20 +08:00
|
|
|
might_sleep();
|
|
|
|
|
2019-09-24 06:36:36 +08:00
|
|
|
spin_lock(&vmap_area_lock);
|
|
|
|
va = __find_vmap_area((unsigned long)addr);
|
2019-09-24 06:36:39 +08:00
|
|
|
if (va && va->vm) {
|
2012-01-11 07:08:39 +08:00
|
|
|
struct vm_struct *vm = va->vm;
|
2011-11-01 08:08:13 +08:00
|
|
|
|
2013-04-30 06:07:30 +08:00
|
|
|
va->vm = NULL;
|
|
|
|
spin_unlock(&vmap_area_lock);
|
|
|
|
|
2015-03-13 07:26:11 +08:00
|
|
|
kasan_free_shadow(vm);
|
2009-09-22 08:02:32 +08:00
|
|
|
free_unmap_vmap_area(va);
|
|
|
|
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
return vm;
|
|
|
|
}
|
2019-09-24 06:36:36 +08:00
|
|
|
|
|
|
|
spin_unlock(&vmap_area_lock);
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
return NULL;
|
2005-05-21 05:27:57 +08:00
|
|
|
}
|
|
|
|
|
2019-04-26 08:11:36 +08:00
|
|
|
static inline void set_area_direct_map(const struct vm_struct *area,
|
|
|
|
int (*set_direct_map)(struct page *page))
|
|
|
|
{
|
|
|
|
int i;
|
|
|
|
|
|
|
|
for (i = 0; i < area->nr_pages; i++)
|
|
|
|
if (page_address(area->pages[i]))
|
|
|
|
set_direct_map(area->pages[i]);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Handle removing and resetting vm mappings related to the vm_struct. */
|
|
|
|
static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages)
|
|
|
|
{
|
|
|
|
unsigned long start = ULONG_MAX, end = 0;
|
|
|
|
int flush_reset = area->flags & VM_FLUSH_RESET_PERMS;
|
2019-05-28 05:10:58 +08:00
|
|
|
int flush_dmap = 0;
|
2019-04-26 08:11:36 +08:00
|
|
|
int i;
|
|
|
|
|
|
|
|
remove_vm_area(area->addr);
|
|
|
|
|
|
|
|
/* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */
|
|
|
|
if (!flush_reset)
|
|
|
|
return;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* If not deallocating pages, just do the flush of the VM area and
|
|
|
|
* return.
|
|
|
|
*/
|
|
|
|
if (!deallocate_pages) {
|
|
|
|
vm_unmap_aliases();
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* If execution gets here, flush the vm mapping and reset the direct
|
|
|
|
* map. Find the start and end range of the direct mappings to make sure
|
|
|
|
* the vm_unmap_aliases() flush includes the direct map.
|
|
|
|
*/
|
|
|
|
for (i = 0; i < area->nr_pages; i++) {
|
2019-05-28 05:10:57 +08:00
|
|
|
unsigned long addr = (unsigned long)page_address(area->pages[i]);
|
|
|
|
if (addr) {
|
2019-04-26 08:11:36 +08:00
|
|
|
start = min(addr, start);
|
2019-05-28 05:10:57 +08:00
|
|
|
end = max(addr + PAGE_SIZE, end);
|
2019-05-28 05:10:58 +08:00
|
|
|
flush_dmap = 1;
|
2019-04-26 08:11:36 +08:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Set direct map to something invalid so that it won't be cached if
|
|
|
|
* there are any accesses after the TLB flush, then flush the TLB and
|
|
|
|
* reset the direct map permissions to the default.
|
|
|
|
*/
|
|
|
|
set_area_direct_map(area, set_direct_map_invalid_noflush);
|
2019-05-28 05:10:58 +08:00
|
|
|
_vm_unmap_aliases(start, end, flush_dmap);
|
2019-04-26 08:11:36 +08:00
|
|
|
set_area_direct_map(area, set_direct_map_default_noflush);
|
|
|
|
}
|
|
|
|
|
2008-02-05 14:28:32 +08:00
|
|
|
static void __vunmap(const void *addr, int deallocate_pages)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
|
|
|
struct vm_struct *area;
|
|
|
|
|
|
|
|
if (!addr)
|
|
|
|
return;
|
|
|
|
|
2013-07-04 06:02:18 +08:00
|
|
|
if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
|
2013-07-09 06:59:53 +08:00
|
|
|
addr))
|
2005-04-17 06:20:36 +08:00
|
|
|
return;
|
|
|
|
|
2019-03-06 07:42:54 +08:00
|
|
|
area = find_vm_area(addr);
|
2005-04-17 06:20:36 +08:00
|
|
|
if (unlikely(!area)) {
|
2008-07-26 10:45:37 +08:00
|
|
|
WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
|
2005-04-17 06:20:36 +08:00
|
|
|
addr);
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2018-06-08 08:06:53 +08:00
|
|
|
debug_check_no_locks_freed(area->addr, get_vm_area_size(area));
|
|
|
|
debug_check_no_obj_freed(area->addr, get_vm_area_size(area));
|
2006-07-03 15:24:33 +08:00
|
|
|
|
2019-12-18 12:51:38 +08:00
|
|
|
kasan_poison_vmalloc(area->addr, area->size);
|
kasan: support backing vmalloc space with real shadow memory
Patch series "kasan: support backing vmalloc space with real shadow
memory", v11.
Currently, vmalloc space is backed by the early shadow page. This means
that kasan is incompatible with VMAP_STACK.
This series provides a mechanism to back vmalloc space with real,
dynamically allocated memory. I have only wired up x86, because that's
the only currently supported arch I can work with easily, but it's very
easy to wire up other architectures, and it appears that there is some
work-in-progress code to do this on arm64 and s390.
This has been discussed before in the context of VMAP_STACK:
- https://bugzilla.kernel.org/show_bug.cgi?id=202009
- https://lkml.org/lkml/2018/7/22/198
- https://lkml.org/lkml/2019/7/19/822
In terms of implementation details:
Most mappings in vmalloc space are small, requiring less than a full
page of shadow space. Allocating a full shadow page per mapping would
therefore be wasteful. Furthermore, to ensure that different mappings
use different shadow pages, mappings would have to be aligned to
KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE.
Instead, share backing space across multiple mappings. Allocate a
backing page when a mapping in vmalloc space uses a particular page of
the shadow region. This page can be shared by other vmalloc mappings
later on.
We hook in to the vmap infrastructure to lazily clean up unused shadow
memory.
Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that:
- Turning on KASAN, inline instrumentation, without vmalloc, introuduces
a 4.1x-4.2x slowdown in vmalloc operations.
- Turning this on introduces the following slowdowns over KASAN:
* ~1.76x slower single-threaded (test_vmalloc.sh performance)
* ~2.18x slower when both cpus are performing operations
simultaneously (test_vmalloc.sh sequential_test_order=1)
This is unfortunate but given that this is a debug feature only, not the
end of the world. The benchmarks are also a stress-test for the vmalloc
subsystem: they're not indicative of an overall 2x slowdown!
This patch (of 4):
Hook into vmalloc and vmap, and dynamically allocate real shadow memory
to back the mappings.
Most mappings in vmalloc space are small, requiring less than a full
page of shadow space. Allocating a full shadow page per mapping would
therefore be wasteful. Furthermore, to ensure that different mappings
use different shadow pages, mappings would have to be aligned to
KASAN_SHADOW_SCALE_SIZE * PAGE_SIZE.
Instead, share backing space across multiple mappings. Allocate a
backing page when a mapping in vmalloc space uses a particular page of
the shadow region. This page can be shared by other vmalloc mappings
later on.
We hook in to the vmap infrastructure to lazily clean up unused shadow
memory.
To avoid the difficulties around swapping mappings around, this code
expects that the part of the shadow region that covers the vmalloc space
will not be covered by the early shadow page, but will be left unmapped.
This will require changes in arch-specific code.
This allows KASAN with VMAP_STACK, and may be helpful for architectures
that do not have a separate module space (e.g. powerpc64, which I am
currently working on). It also allows relaxing the module alignment
back to PAGE_SIZE.
Testing with test_vmalloc.sh on an x86 VM with 2 vCPUs shows that:
- Turning on KASAN, inline instrumentation, without vmalloc, introuduces
a 4.1x-4.2x slowdown in vmalloc operations.
- Turning this on introduces the following slowdowns over KASAN:
* ~1.76x slower single-threaded (test_vmalloc.sh performance)
* ~2.18x slower when both cpus are performing operations
simultaneously (test_vmalloc.sh sequential_test_order=3D1)
This is unfortunate but given that this is a debug feature only, not the
end of the world.
The full benchmark results are:
Performance
No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN
fix_size_alloc_test 662004 11404956 17.23 19144610 28.92 1.68
full_fit_alloc_test 710950 12029752 16.92 13184651 18.55 1.10
long_busy_list_alloc_test 9431875 43990172 4.66 82970178 8.80 1.89
random_size_alloc_test 5033626 23061762 4.58 47158834 9.37 2.04
fix_align_alloc_test 1252514 15276910 12.20 31266116 24.96 2.05
random_size_align_alloc_te 1648501 14578321 8.84 25560052 15.51 1.75
align_shift_alloc_test 147 830 5.65 5692 38.72 6.86
pcpu_alloc_test 80732 125520 1.55 140864 1.74 1.12
Total Cycles 119240774314 763211341128 6.40 1390338696894 11.66 1.82
Sequential, 2 cpus
No KASAN KASAN original x baseline KASAN vmalloc x baseline x KASAN
fix_size_alloc_test 1423150 14276550 10.03 27733022 19.49 1.94
full_fit_alloc_test 1754219 14722640 8.39 15030786 8.57 1.02
long_busy_list_alloc_test 11451858 52154973 4.55 107016027 9.34 2.05
random_size_alloc_test 5989020 26735276 4.46 68885923 11.50 2.58
fix_align_alloc_test 2050976 20166900 9.83 50491675 24.62 2.50
random_size_align_alloc_te 2858229 17971700 6.29 38730225 13.55 2.16
align_shift_alloc_test 405 6428 15.87 26253 64.82 4.08
pcpu_alloc_test 127183 151464 1.19 216263 1.70 1.43
Total Cycles 54181269392 308723699764 5.70 650772566394 12.01 2.11
fix_size_alloc_test 1420404 14289308 10.06 27790035 19.56 1.94
full_fit_alloc_test 1736145 14806234 8.53 15274301 8.80 1.03
long_busy_list_alloc_test 11404638 52270785 4.58 107550254 9.43 2.06
random_size_alloc_test 6017006 26650625 4.43 68696127 11.42 2.58
fix_align_alloc_test 2045504 20280985 9.91 50414862 24.65 2.49
random_size_align_alloc_te 2845338 17931018 6.30 38510276 13.53 2.15
align_shift_alloc_test 472 3760 7.97 9656 20.46 2.57
pcpu_alloc_test 118643 132732 1.12 146504 1.23 1.10
Total Cycles 54040011688 309102805492 5.72 651325675652 12.05 2.11
[dja@axtens.net: fixups]
Link: http://lkml.kernel.org/r/20191120052719.7201-1-dja@axtens.net
Link: https://bugzilla.kernel.org/show_bug.cgi?id=3D202009
Link: http://lkml.kernel.org/r/20191031093909.9228-2-dja@axtens.net
Signed-off-by: Mark Rutland <mark.rutland@arm.com> [shadow rework]
Signed-off-by: Daniel Axtens <dja@axtens.net>
Co-developed-by: Mark Rutland <mark.rutland@arm.com>
Acked-by: Vasily Gorbik <gor@linux.ibm.com>
Reviewed-by: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Christophe Leroy <christophe.leroy@c-s.fr>
Cc: Qian Cai <cai@lca.pw>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:50 +08:00
|
|
|
|
2019-04-26 08:11:36 +08:00
|
|
|
vm_remove_mappings(area, deallocate_pages);
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
if (deallocate_pages) {
|
|
|
|
int i;
|
|
|
|
|
|
|
|
for (i = 0; i < area->nr_pages; i++) {
|
2008-02-05 14:28:34 +08:00
|
|
|
struct page *page = area->pages[i];
|
|
|
|
|
|
|
|
BUG_ON(!page);
|
mm: charge/uncharge kmemcg from generic page allocator paths
Currently, to charge a non-slab allocation to kmemcg one has to use
alloc_kmem_pages helper with __GFP_ACCOUNT flag. A page allocated with
this helper should finally be freed using free_kmem_pages, otherwise it
won't be uncharged.
This API suits its current users fine, but it turns out to be impossible
to use along with page reference counting, i.e. when an allocation is
supposed to be freed with put_page, as it is the case with pipe or unix
socket buffers.
To overcome this limitation, this patch moves charging/uncharging to
generic page allocator paths, i.e. to __alloc_pages_nodemask and
free_pages_prepare, and zaps alloc/free_kmem_pages helpers. This way,
one can use any of the available page allocation functions to get the
allocated page charged to kmemcg - it's enough to pass __GFP_ACCOUNT,
just like in case of kmalloc and friends. A charged page will be
automatically uncharged on free.
To make it possible, we need to mark pages charged to kmemcg somehow.
To avoid introducing a new page flag, we make use of page->_mapcount for
marking such pages. Since pages charged to kmemcg are not supposed to
be mapped to userspace, it should work just fine. There are other
(ab)users of page->_mapcount - buddy and balloon pages - but we don't
conflict with them.
In case kmemcg is compiled out or not used at runtime, this patch
introduces no overhead to generic page allocator paths. If kmemcg is
used, it will be plus one gfp flags check on alloc and plus one
page->_mapcount check on free, which shouldn't hurt performance, because
the data accessed are hot.
Link: http://lkml.kernel.org/r/a9736d856f895bcb465d9f257b54efe32eda6f99.1464079538.git.vdavydov@virtuozzo.com
Signed-off-by: Vladimir Davydov <vdavydov@virtuozzo.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-07-27 06:24:24 +08:00
|
|
|
__free_pages(page, 0);
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
2019-07-12 12:00:13 +08:00
|
|
|
atomic_long_sub(area->nr_pages, &nr_vmalloc_pages);
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2016-01-15 07:19:35 +08:00
|
|
|
kvfree(area->pages);
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
kfree(area);
|
|
|
|
return;
|
|
|
|
}
|
2016-12-13 08:44:10 +08:00
|
|
|
|
|
|
|
static inline void __vfree_deferred(const void *addr)
|
|
|
|
{
|
|
|
|
/*
|
|
|
|
* Use raw_cpu_ptr() because this can be called from preemptible
|
|
|
|
* context. Preemption is absolutely fine here, because the llist_add()
|
|
|
|
* implementation is lockless, so it works even if we are adding to
|
|
|
|
* nother cpu's list. schedule_work() should be fine with this too.
|
|
|
|
*/
|
|
|
|
struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
|
|
|
|
|
|
|
|
if (llist_add((struct llist_node *)addr, &p->list))
|
|
|
|
schedule_work(&p->wq);
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
2019-03-06 07:48:36 +08:00
|
|
|
* vfree_atomic - release memory allocated by vmalloc()
|
|
|
|
* @addr: memory base address
|
2016-12-13 08:44:10 +08:00
|
|
|
*
|
2019-03-06 07:48:36 +08:00
|
|
|
* This one is just like vfree() but can be called in any atomic context
|
|
|
|
* except NMIs.
|
2016-12-13 08:44:10 +08:00
|
|
|
*/
|
|
|
|
void vfree_atomic(const void *addr)
|
|
|
|
{
|
|
|
|
BUG_ON(in_nmi());
|
|
|
|
|
|
|
|
kmemleak_free(addr);
|
|
|
|
|
|
|
|
if (!addr)
|
|
|
|
return;
|
|
|
|
__vfree_deferred(addr);
|
|
|
|
}
|
|
|
|
|
2019-03-06 07:43:24 +08:00
|
|
|
static void __vfree(const void *addr)
|
|
|
|
{
|
|
|
|
if (unlikely(in_interrupt()))
|
|
|
|
__vfree_deferred(addr);
|
|
|
|
else
|
|
|
|
__vunmap(addr, 1);
|
|
|
|
}
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
/**
|
2019-03-06 07:48:36 +08:00
|
|
|
* vfree - release memory allocated by vmalloc()
|
|
|
|
* @addr: memory base address
|
2005-04-17 06:20:36 +08:00
|
|
|
*
|
2019-03-06 07:48:36 +08:00
|
|
|
* Free the virtually continuous memory area starting at @addr, as
|
|
|
|
* obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
|
|
|
|
* NULL, no operation is performed.
|
2005-04-17 06:20:36 +08:00
|
|
|
*
|
2019-03-06 07:48:36 +08:00
|
|
|
* Must not be called in NMI context (strictly speaking, only if we don't
|
|
|
|
* have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
|
|
|
|
* conventions for vfree() arch-depenedent would be a really bad idea)
|
2013-05-08 07:18:18 +08:00
|
|
|
*
|
2019-03-06 07:48:36 +08:00
|
|
|
* May sleep if called *not* from interrupt context.
|
2018-10-27 06:07:03 +08:00
|
|
|
*
|
2019-03-06 07:48:36 +08:00
|
|
|
* NOTE: assumes that the object at @addr has a size >= sizeof(llist_node)
|
2005-04-17 06:20:36 +08:00
|
|
|
*/
|
2008-02-05 14:28:32 +08:00
|
|
|
void vfree(const void *addr)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
2013-03-11 08:14:08 +08:00
|
|
|
BUG_ON(in_nmi());
|
2009-06-11 20:23:19 +08:00
|
|
|
|
|
|
|
kmemleak_free(addr);
|
|
|
|
|
2018-10-27 06:07:07 +08:00
|
|
|
might_sleep_if(!in_interrupt());
|
|
|
|
|
2013-03-11 08:14:08 +08:00
|
|
|
if (!addr)
|
|
|
|
return;
|
2019-03-06 07:43:24 +08:00
|
|
|
|
|
|
|
__vfree(addr);
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(vfree);
|
|
|
|
|
|
|
|
/**
|
2019-03-06 07:48:36 +08:00
|
|
|
* vunmap - release virtual mapping obtained by vmap()
|
|
|
|
* @addr: memory base address
|
2005-04-17 06:20:36 +08:00
|
|
|
*
|
2019-03-06 07:48:36 +08:00
|
|
|
* Free the virtually contiguous memory area starting at @addr,
|
|
|
|
* which was created from the page array passed to vmap().
|
2005-04-17 06:20:36 +08:00
|
|
|
*
|
2019-03-06 07:48:36 +08:00
|
|
|
* Must not be called in interrupt context.
|
2005-04-17 06:20:36 +08:00
|
|
|
*/
|
2008-02-05 14:28:32 +08:00
|
|
|
void vunmap(const void *addr)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
|
|
|
BUG_ON(in_interrupt());
|
2009-02-25 23:04:03 +08:00
|
|
|
might_sleep();
|
2013-03-11 08:14:08 +08:00
|
|
|
if (addr)
|
|
|
|
__vunmap(addr, 0);
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(vunmap);
|
|
|
|
|
|
|
|
/**
|
2019-03-06 07:48:36 +08:00
|
|
|
* vmap - map an array of pages into virtually contiguous space
|
|
|
|
* @pages: array of page pointers
|
|
|
|
* @count: number of pages to map
|
|
|
|
* @flags: vm_area->flags
|
|
|
|
* @prot: page protection for the mapping
|
|
|
|
*
|
|
|
|
* Maps @count pages from @pages into contiguous kernel virtual
|
|
|
|
* space.
|
2019-03-06 07:48:42 +08:00
|
|
|
*
|
|
|
|
* Return: the address of the area or %NULL on failure
|
2005-04-17 06:20:36 +08:00
|
|
|
*/
|
|
|
|
void *vmap(struct page **pages, unsigned int count,
|
2019-03-06 07:48:36 +08:00
|
|
|
unsigned long flags, pgprot_t prot)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
|
|
|
struct vm_struct *area;
|
2016-06-04 05:55:33 +08:00
|
|
|
unsigned long size; /* In bytes */
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2009-02-25 23:04:03 +08:00
|
|
|
might_sleep();
|
|
|
|
|
2018-12-28 16:34:29 +08:00
|
|
|
if (count > totalram_pages())
|
2005-04-17 06:20:36 +08:00
|
|
|
return NULL;
|
|
|
|
|
2016-06-04 05:55:33 +08:00
|
|
|
size = (unsigned long)count << PAGE_SHIFT;
|
|
|
|
area = get_vm_area_caller(size, flags, __builtin_return_address(0));
|
2005-04-17 06:20:36 +08:00
|
|
|
if (!area)
|
|
|
|
return NULL;
|
2008-04-28 17:12:42 +08:00
|
|
|
|
2014-08-07 07:06:58 +08:00
|
|
|
if (map_vm_area(area, prot, pages)) {
|
2005-04-17 06:20:36 +08:00
|
|
|
vunmap(area->addr);
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
return area->addr;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(vmap);
|
|
|
|
|
mm, vmalloc: fix vmalloc users tracking properly
Commit 1f5307b1e094 ("mm, vmalloc: properly track vmalloc users") has
pulled asm/pgtable.h include dependency to linux/vmalloc.h and that
turned out to be a bad idea for some architectures. E.g. m68k fails
with
In file included from arch/m68k/include/asm/pgtable_mm.h:145:0,
from arch/m68k/include/asm/pgtable.h:4,
from include/linux/vmalloc.h:9,
from arch/m68k/kernel/module.c:9:
arch/m68k/include/asm/mcf_pgtable.h: In function 'nocache_page':
>> arch/m68k/include/asm/mcf_pgtable.h:339:43: error: 'init_mm' undeclared (first use in this function)
#define pgd_offset_k(address) pgd_offset(&init_mm, address)
as spotted by kernel build bot. nios2 fails for other reason
In file included from include/asm-generic/io.h:767:0,
from arch/nios2/include/asm/io.h:61,
from include/linux/io.h:25,
from arch/nios2/include/asm/pgtable.h:18,
from include/linux/mm.h:70,
from include/linux/pid_namespace.h:6,
from include/linux/ptrace.h:9,
from arch/nios2/include/uapi/asm/elf.h:23,
from arch/nios2/include/asm/elf.h:22,
from include/linux/elf.h:4,
from include/linux/module.h:15,
from init/main.c:16:
include/linux/vmalloc.h: In function '__vmalloc_node_flags':
include/linux/vmalloc.h:99:40: error: 'PAGE_KERNEL' undeclared (first use in this function); did you mean 'GFP_KERNEL'?
which is due to the newly added #include <asm/pgtable.h>, which on nios2
includes <linux/io.h> and thus <asm/io.h> and <asm-generic/io.h> which
again includes <linux/vmalloc.h>.
Tweaking that around just turns out a bigger headache than necessary.
This patch reverts 1f5307b1e094 and reimplements the original fix in a
different way. __vmalloc_node_flags can stay static inline which will
cover vmalloc* functions. We only have one external user
(kvmalloc_node) and we can export __vmalloc_node_flags_caller and
provide the caller directly. This is much simpler and it doesn't really
need any games with header files.
[akpm@linux-foundation.org: coding-style fixes]
[mhocko@kernel.org: revert old comment]
Link: http://lkml.kernel.org/r/20170509211054.GB16325@dhcp22.suse.cz
Fixes: 1f5307b1e094 ("mm, vmalloc: properly track vmalloc users")
Link: http://lkml.kernel.org/r/20170509153702.GR6481@dhcp22.suse.cz
Signed-off-by: Michal Hocko <mhocko@suse.com>
Cc: Tobias Klauser <tklauser@distanz.ch>
Cc: Geert Uytterhoeven <geert@linux-m68k.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-13 06:46:41 +08:00
|
|
|
static void *__vmalloc_node(unsigned long size, unsigned long align,
|
|
|
|
gfp_t gfp_mask, pgprot_t prot,
|
|
|
|
int node, const void *caller);
|
2008-02-05 14:29:09 +08:00
|
|
|
static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
|
2013-11-13 07:07:29 +08:00
|
|
|
pgprot_t prot, int node)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
|
|
|
struct page **pages;
|
|
|
|
unsigned int nr_pages, array_size, i;
|
2014-08-07 07:06:28 +08:00
|
|
|
const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
|
2017-08-19 06:16:27 +08:00
|
|
|
const gfp_t alloc_mask = gfp_mask | __GFP_NOWARN;
|
|
|
|
const gfp_t highmem_mask = (gfp_mask & (GFP_DMA | GFP_DMA32)) ?
|
|
|
|
0 :
|
|
|
|
__GFP_HIGHMEM;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2013-09-12 05:22:42 +08:00
|
|
|
nr_pages = get_vm_area_size(area) >> PAGE_SHIFT;
|
2005-04-17 06:20:36 +08:00
|
|
|
array_size = (nr_pages * sizeof(struct page *));
|
|
|
|
|
|
|
|
/* Please note that the recursion is strictly bounded. */
|
2006-07-14 15:23:56 +08:00
|
|
|
if (array_size > PAGE_SIZE) {
|
2017-08-19 06:16:27 +08:00
|
|
|
pages = __vmalloc_node(array_size, 1, nested_gfp|highmem_mask,
|
2013-11-13 07:07:29 +08:00
|
|
|
PAGE_KERNEL, node, area->caller);
|
2006-10-17 15:09:57 +08:00
|
|
|
} else {
|
2009-12-15 09:58:39 +08:00
|
|
|
pages = kmalloc_node(array_size, nested_gfp, node);
|
2006-10-17 15:09:57 +08:00
|
|
|
}
|
2019-09-24 06:36:42 +08:00
|
|
|
|
|
|
|
if (!pages) {
|
2005-04-17 06:20:36 +08:00
|
|
|
remove_vm_area(area->addr);
|
|
|
|
kfree(area);
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
2019-09-24 06:36:42 +08:00
|
|
|
area->pages = pages;
|
|
|
|
area->nr_pages = nr_pages;
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
for (i = 0; i < area->nr_pages; i++) {
|
2008-02-05 14:28:34 +08:00
|
|
|
struct page *page;
|
|
|
|
|
2013-11-13 07:07:11 +08:00
|
|
|
if (node == NUMA_NO_NODE)
|
2017-08-19 06:16:27 +08:00
|
|
|
page = alloc_page(alloc_mask|highmem_mask);
|
2005-10-30 09:15:41 +08:00
|
|
|
else
|
2017-08-19 06:16:27 +08:00
|
|
|
page = alloc_pages_node(node, alloc_mask|highmem_mask, 0);
|
2008-02-05 14:28:34 +08:00
|
|
|
|
|
|
|
if (unlikely(!page)) {
|
2005-04-17 06:20:36 +08:00
|
|
|
/* Successfully allocated i pages, free them in __vunmap() */
|
|
|
|
area->nr_pages = i;
|
2019-07-12 12:00:13 +08:00
|
|
|
atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
|
2005-04-17 06:20:36 +08:00
|
|
|
goto fail;
|
|
|
|
}
|
2008-02-05 14:28:34 +08:00
|
|
|
area->pages[i] = page;
|
2019-12-01 09:54:30 +08:00
|
|
|
if (gfpflags_allow_blocking(gfp_mask))
|
2014-08-07 07:06:25 +08:00
|
|
|
cond_resched();
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
2019-07-12 12:00:13 +08:00
|
|
|
atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2014-08-07 07:06:58 +08:00
|
|
|
if (map_vm_area(area, prot, pages))
|
2005-04-17 06:20:36 +08:00
|
|
|
goto fail;
|
|
|
|
return area->addr;
|
|
|
|
|
|
|
|
fail:
|
2017-02-23 07:46:10 +08:00
|
|
|
warn_alloc(gfp_mask, NULL,
|
2016-10-08 08:01:55 +08:00
|
|
|
"vmalloc: allocation failure, allocated %ld of %ld bytes",
|
mm: print vmalloc() state after allocation failures
I was tracking down a page allocation failure that ended up in vmalloc().
Since vmalloc() uses 0-order pages, if somebody asks for an insane amount
of memory, we'll still get a warning with "order:0" in it. That's not
very useful.
During recovery, vmalloc() also nicely frees all of the memory that it got
up to the point of the failure. That is wonderful, but it also quickly
hides any issues. We have a much different sitation if vmalloc()
repeatedly fails 10GB in to:
vmalloc(100 * 1<<30);
versus repeatedly failing 4096 bytes in to a:
vmalloc(8192);
This patch will print out messages that look like this:
[ 68.123503] vmalloc: allocation failure, allocated 6680576 of 13426688 bytes
[ 68.124218] bash: page allocation failure: order:0, mode:0xd2
[ 68.124811] Pid: 3770, comm: bash Not tainted 2.6.39-rc3-00082-g85f2e68-dirty #333
[ 68.125579] Call Trace:
[ 68.125853] [<ffffffff810f6da6>] warn_alloc_failed+0x146/0x170
[ 68.126464] [<ffffffff8107e05c>] ? printk+0x6c/0x70
[ 68.126791] [<ffffffff8112b5d4>] ? alloc_pages_current+0x94/0xe0
[ 68.127661] [<ffffffff8111ed37>] __vmalloc_node_range+0x237/0x290
...
The 'order' variable is added for clarity when calling warn_alloc_failed()
to avoid having an unexplained '0' as an argument.
The 'tmp_mask' is because adding an open-coded '| __GFP_NOWARN' would take
us over 80 columns for the alloc_pages_node() call. If we are going to
add a line, it might as well be one that makes the sucker easier to read.
As a side issue, I also noticed that ctl_ioctl() does vmalloc() based
solely on an unverified value passed in from userspace. Granted, it's
under CAP_SYS_ADMIN, but it still frightens me a bit.
Signed-off-by: Dave Hansen <dave@linux.vnet.ibm.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Michal Nazarewicz <mina86@mina86.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-05-25 08:12:18 +08:00
|
|
|
(area->nr_pages*PAGE_SIZE), area->size);
|
2019-03-06 07:43:24 +08:00
|
|
|
__vfree(area->addr);
|
2005-04-17 06:20:36 +08:00
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
2019-03-06 07:48:36 +08:00
|
|
|
* __vmalloc_node_range - allocate virtually contiguous memory
|
|
|
|
* @size: allocation size
|
|
|
|
* @align: desired alignment
|
|
|
|
* @start: vm area range start
|
|
|
|
* @end: vm area range end
|
|
|
|
* @gfp_mask: flags for the page level allocator
|
|
|
|
* @prot: protection mask for the allocated pages
|
|
|
|
* @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
|
|
|
|
* @node: node to use for allocation or NUMA_NO_NODE
|
|
|
|
* @caller: caller's return address
|
|
|
|
*
|
|
|
|
* Allocate enough pages to cover @size from the page level
|
|
|
|
* allocator with @gfp_mask flags. Map them into contiguous
|
|
|
|
* kernel virtual space, using a pagetable protection of @prot.
|
2019-03-06 07:48:42 +08:00
|
|
|
*
|
|
|
|
* Return: the address of the area or %NULL on failure
|
2005-04-17 06:20:36 +08:00
|
|
|
*/
|
2011-01-14 07:46:02 +08:00
|
|
|
void *__vmalloc_node_range(unsigned long size, unsigned long align,
|
|
|
|
unsigned long start, unsigned long end, gfp_t gfp_mask,
|
2015-02-14 06:40:07 +08:00
|
|
|
pgprot_t prot, unsigned long vm_flags, int node,
|
|
|
|
const void *caller)
|
2005-04-17 06:20:36 +08:00
|
|
|
{
|
|
|
|
struct vm_struct *area;
|
2009-06-11 20:23:19 +08:00
|
|
|
void *addr;
|
|
|
|
unsigned long real_size = size;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
size = PAGE_ALIGN(size);
|
2018-12-28 16:34:29 +08:00
|
|
|
if (!size || (size >> PAGE_SHIFT) > totalram_pages())
|
2011-11-01 08:08:48 +08:00
|
|
|
goto fail;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2019-12-18 12:51:38 +08:00
|
|
|
area = __get_vm_area_node(real_size, align, VM_ALLOC | VM_UNINITIALIZED |
|
2015-02-14 06:40:07 +08:00
|
|
|
vm_flags, start, end, node, gfp_mask, caller);
|
2005-04-17 06:20:36 +08:00
|
|
|
if (!area)
|
2011-11-01 08:08:48 +08:00
|
|
|
goto fail;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
2013-11-13 07:07:29 +08:00
|
|
|
addr = __vmalloc_area_node(area, gfp_mask, prot, node);
|
2011-12-09 06:34:30 +08:00
|
|
|
if (!addr)
|
2013-11-13 07:07:33 +08:00
|
|
|
return NULL;
|
2009-06-11 20:23:19 +08:00
|
|
|
|
2011-11-01 08:08:13 +08:00
|
|
|
/*
|
2013-07-09 06:59:58 +08:00
|
|
|
* In this function, newly allocated vm_struct has VM_UNINITIALIZED
|
|
|
|
* flag. It means that vm_struct is not fully initialized.
|
2013-04-30 06:07:39 +08:00
|
|
|
* Now, it is fully initialized, so remove this flag here.
|
2011-11-01 08:08:13 +08:00
|
|
|
*/
|
2013-07-09 06:59:58 +08:00
|
|
|
clear_vm_uninitialized_flag(area);
|
2011-11-01 08:08:13 +08:00
|
|
|
|
2017-07-07 06:40:22 +08:00
|
|
|
kmemleak_vmalloc(area, size, gfp_mask);
|
2009-06-11 20:23:19 +08:00
|
|
|
|
|
|
|
return addr;
|
2011-11-01 08:08:48 +08:00
|
|
|
|
|
|
|
fail:
|
2017-02-23 07:46:10 +08:00
|
|
|
warn_alloc(gfp_mask, NULL,
|
2016-10-08 08:01:55 +08:00
|
|
|
"vmalloc: allocation failure: %lu bytes", real_size);
|
2011-11-01 08:08:48 +08:00
|
|
|
return NULL;
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
2019-03-06 07:43:30 +08:00
|
|
|
/*
|
|
|
|
* This is only for performance analysis of vmalloc and stress purpose.
|
|
|
|
* It is required by vmalloc test module, therefore do not use it other
|
|
|
|
* than that.
|
|
|
|
*/
|
|
|
|
#ifdef CONFIG_TEST_VMALLOC_MODULE
|
|
|
|
EXPORT_SYMBOL_GPL(__vmalloc_node_range);
|
|
|
|
#endif
|
|
|
|
|
2011-01-14 07:46:02 +08:00
|
|
|
/**
|
2019-03-06 07:48:36 +08:00
|
|
|
* __vmalloc_node - allocate virtually contiguous memory
|
|
|
|
* @size: allocation size
|
|
|
|
* @align: desired alignment
|
|
|
|
* @gfp_mask: flags for the page level allocator
|
|
|
|
* @prot: protection mask for the allocated pages
|
|
|
|
* @node: node to use for allocation or NUMA_NO_NODE
|
|
|
|
* @caller: caller's return address
|
2017-05-09 06:57:09 +08:00
|
|
|
*
|
2019-03-06 07:48:36 +08:00
|
|
|
* Allocate enough pages to cover @size from the page level
|
|
|
|
* allocator with @gfp_mask flags. Map them into contiguous
|
|
|
|
* kernel virtual space, using a pagetable protection of @prot.
|
2017-05-09 06:57:09 +08:00
|
|
|
*
|
2019-03-06 07:48:36 +08:00
|
|
|
* Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
|
|
|
|
* and __GFP_NOFAIL are not supported
|
2017-05-09 06:57:09 +08:00
|
|
|
*
|
2019-03-06 07:48:36 +08:00
|
|
|
* Any use of gfp flags outside of GFP_KERNEL should be consulted
|
|
|
|
* with mm people.
|
2019-03-06 07:48:42 +08:00
|
|
|
*
|
|
|
|
* Return: pointer to the allocated memory or %NULL on error
|
2011-01-14 07:46:02 +08:00
|
|
|
*/
|
mm, vmalloc: fix vmalloc users tracking properly
Commit 1f5307b1e094 ("mm, vmalloc: properly track vmalloc users") has
pulled asm/pgtable.h include dependency to linux/vmalloc.h and that
turned out to be a bad idea for some architectures. E.g. m68k fails
with
In file included from arch/m68k/include/asm/pgtable_mm.h:145:0,
from arch/m68k/include/asm/pgtable.h:4,
from include/linux/vmalloc.h:9,
from arch/m68k/kernel/module.c:9:
arch/m68k/include/asm/mcf_pgtable.h: In function 'nocache_page':
>> arch/m68k/include/asm/mcf_pgtable.h:339:43: error: 'init_mm' undeclared (first use in this function)
#define pgd_offset_k(address) pgd_offset(&init_mm, address)
as spotted by kernel build bot. nios2 fails for other reason
In file included from include/asm-generic/io.h:767:0,
from arch/nios2/include/asm/io.h:61,
from include/linux/io.h:25,
from arch/nios2/include/asm/pgtable.h:18,
from include/linux/mm.h:70,
from include/linux/pid_namespace.h:6,
from include/linux/ptrace.h:9,
from arch/nios2/include/uapi/asm/elf.h:23,
from arch/nios2/include/asm/elf.h:22,
from include/linux/elf.h:4,
from include/linux/module.h:15,
from init/main.c:16:
include/linux/vmalloc.h: In function '__vmalloc_node_flags':
include/linux/vmalloc.h:99:40: error: 'PAGE_KERNEL' undeclared (first use in this function); did you mean 'GFP_KERNEL'?
which is due to the newly added #include <asm/pgtable.h>, which on nios2
includes <linux/io.h> and thus <asm/io.h> and <asm-generic/io.h> which
again includes <linux/vmalloc.h>.
Tweaking that around just turns out a bigger headache than necessary.
This patch reverts 1f5307b1e094 and reimplements the original fix in a
different way. __vmalloc_node_flags can stay static inline which will
cover vmalloc* functions. We only have one external user
(kvmalloc_node) and we can export __vmalloc_node_flags_caller and
provide the caller directly. This is much simpler and it doesn't really
need any games with header files.
[akpm@linux-foundation.org: coding-style fixes]
[mhocko@kernel.org: revert old comment]
Link: http://lkml.kernel.org/r/20170509211054.GB16325@dhcp22.suse.cz
Fixes: 1f5307b1e094 ("mm, vmalloc: properly track vmalloc users")
Link: http://lkml.kernel.org/r/20170509153702.GR6481@dhcp22.suse.cz
Signed-off-by: Michal Hocko <mhocko@suse.com>
Cc: Tobias Klauser <tklauser@distanz.ch>
Cc: Geert Uytterhoeven <geert@linux-m68k.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-13 06:46:41 +08:00
|
|
|
static void *__vmalloc_node(unsigned long size, unsigned long align,
|
2011-01-14 07:46:02 +08:00
|
|
|
gfp_t gfp_mask, pgprot_t prot,
|
2012-04-13 18:32:09 +08:00
|
|
|
int node, const void *caller)
|
2011-01-14 07:46:02 +08:00
|
|
|
{
|
|
|
|
return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
|
2015-02-14 06:40:07 +08:00
|
|
|
gfp_mask, prot, 0, node, caller);
|
2011-01-14 07:46:02 +08:00
|
|
|
}
|
|
|
|
|
2005-10-30 09:15:41 +08:00
|
|
|
void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
|
|
|
|
{
|
2013-02-23 08:35:36 +08:00
|
|
|
return __vmalloc_node(size, 1, gfp_mask, prot, NUMA_NO_NODE,
|
2008-04-28 17:12:42 +08:00
|
|
|
__builtin_return_address(0));
|
2005-10-30 09:15:41 +08:00
|
|
|
}
|
2005-04-17 06:20:36 +08:00
|
|
|
EXPORT_SYMBOL(__vmalloc);
|
|
|
|
|
mm, vmalloc: fix vmalloc users tracking properly
Commit 1f5307b1e094 ("mm, vmalloc: properly track vmalloc users") has
pulled asm/pgtable.h include dependency to linux/vmalloc.h and that
turned out to be a bad idea for some architectures. E.g. m68k fails
with
In file included from arch/m68k/include/asm/pgtable_mm.h:145:0,
from arch/m68k/include/asm/pgtable.h:4,
from include/linux/vmalloc.h:9,
from arch/m68k/kernel/module.c:9:
arch/m68k/include/asm/mcf_pgtable.h: In function 'nocache_page':
>> arch/m68k/include/asm/mcf_pgtable.h:339:43: error: 'init_mm' undeclared (first use in this function)
#define pgd_offset_k(address) pgd_offset(&init_mm, address)
as spotted by kernel build bot. nios2 fails for other reason
In file included from include/asm-generic/io.h:767:0,
from arch/nios2/include/asm/io.h:61,
from include/linux/io.h:25,
from arch/nios2/include/asm/pgtable.h:18,
from include/linux/mm.h:70,
from include/linux/pid_namespace.h:6,
from include/linux/ptrace.h:9,
from arch/nios2/include/uapi/asm/elf.h:23,
from arch/nios2/include/asm/elf.h:22,
from include/linux/elf.h:4,
from include/linux/module.h:15,
from init/main.c:16:
include/linux/vmalloc.h: In function '__vmalloc_node_flags':
include/linux/vmalloc.h:99:40: error: 'PAGE_KERNEL' undeclared (first use in this function); did you mean 'GFP_KERNEL'?
which is due to the newly added #include <asm/pgtable.h>, which on nios2
includes <linux/io.h> and thus <asm/io.h> and <asm-generic/io.h> which
again includes <linux/vmalloc.h>.
Tweaking that around just turns out a bigger headache than necessary.
This patch reverts 1f5307b1e094 and reimplements the original fix in a
different way. __vmalloc_node_flags can stay static inline which will
cover vmalloc* functions. We only have one external user
(kvmalloc_node) and we can export __vmalloc_node_flags_caller and
provide the caller directly. This is much simpler and it doesn't really
need any games with header files.
[akpm@linux-foundation.org: coding-style fixes]
[mhocko@kernel.org: revert old comment]
Link: http://lkml.kernel.org/r/20170509211054.GB16325@dhcp22.suse.cz
Fixes: 1f5307b1e094 ("mm, vmalloc: properly track vmalloc users")
Link: http://lkml.kernel.org/r/20170509153702.GR6481@dhcp22.suse.cz
Signed-off-by: Michal Hocko <mhocko@suse.com>
Cc: Tobias Klauser <tklauser@distanz.ch>
Cc: Geert Uytterhoeven <geert@linux-m68k.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-13 06:46:41 +08:00
|
|
|
static inline void *__vmalloc_node_flags(unsigned long size,
|
|
|
|
int node, gfp_t flags)
|
|
|
|
{
|
|
|
|
return __vmalloc_node(size, 1, flags, PAGE_KERNEL,
|
|
|
|
node, __builtin_return_address(0));
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
void *__vmalloc_node_flags_caller(unsigned long size, int node, gfp_t flags,
|
|
|
|
void *caller)
|
|
|
|
{
|
|
|
|
return __vmalloc_node(size, 1, flags, PAGE_KERNEL, node, caller);
|
|
|
|
}
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
/**
|
2019-03-06 07:48:36 +08:00
|
|
|
* vmalloc - allocate virtually contiguous memory
|
|
|
|
* @size: allocation size
|
|
|
|
*
|
|
|
|
* Allocate enough pages to cover @size from the page level
|
|
|
|
* allocator and map them into contiguous kernel virtual space.
|
2005-04-17 06:20:36 +08:00
|
|
|
*
|
2019-03-06 07:48:36 +08:00
|
|
|
* For tight control over page level allocator and protection flags
|
|
|
|
* use __vmalloc() instead.
|
2019-03-06 07:48:42 +08:00
|
|
|
*
|
|
|
|
* Return: pointer to the allocated memory or %NULL on error
|
2005-04-17 06:20:36 +08:00
|
|
|
*/
|
|
|
|
void *vmalloc(unsigned long size)
|
|
|
|
{
|
2013-02-23 08:35:36 +08:00
|
|
|
return __vmalloc_node_flags(size, NUMA_NO_NODE,
|
2017-05-09 06:57:44 +08:00
|
|
|
GFP_KERNEL);
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(vmalloc);
|
|
|
|
|
2010-10-27 05:22:06 +08:00
|
|
|
/**
|
2019-03-06 07:48:36 +08:00
|
|
|
* vzalloc - allocate virtually contiguous memory with zero fill
|
|
|
|
* @size: allocation size
|
|
|
|
*
|
|
|
|
* Allocate enough pages to cover @size from the page level
|
|
|
|
* allocator and map them into contiguous kernel virtual space.
|
|
|
|
* The memory allocated is set to zero.
|
|
|
|
*
|
|
|
|
* For tight control over page level allocator and protection flags
|
|
|
|
* use __vmalloc() instead.
|
2019-03-06 07:48:42 +08:00
|
|
|
*
|
|
|
|
* Return: pointer to the allocated memory or %NULL on error
|
2010-10-27 05:22:06 +08:00
|
|
|
*/
|
|
|
|
void *vzalloc(unsigned long size)
|
|
|
|
{
|
2013-02-23 08:35:36 +08:00
|
|
|
return __vmalloc_node_flags(size, NUMA_NO_NODE,
|
2017-05-09 06:57:44 +08:00
|
|
|
GFP_KERNEL | __GFP_ZERO);
|
2010-10-27 05:22:06 +08:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(vzalloc);
|
|
|
|
|
2006-06-23 17:03:20 +08:00
|
|
|
/**
|
2006-09-27 16:50:13 +08:00
|
|
|
* vmalloc_user - allocate zeroed virtually contiguous memory for userspace
|
|
|
|
* @size: allocation size
|
2006-06-23 17:03:20 +08:00
|
|
|
*
|
2006-09-27 16:50:13 +08:00
|
|
|
* The resulting memory area is zeroed so it can be mapped to userspace
|
|
|
|
* without leaking data.
|
2019-03-06 07:48:42 +08:00
|
|
|
*
|
|
|
|
* Return: pointer to the allocated memory or %NULL on error
|
2006-06-23 17:03:20 +08:00
|
|
|
*/
|
|
|
|
void *vmalloc_user(unsigned long size)
|
|
|
|
{
|
2019-03-06 07:43:27 +08:00
|
|
|
return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
|
|
|
|
GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL,
|
|
|
|
VM_USERMAP, NUMA_NO_NODE,
|
|
|
|
__builtin_return_address(0));
|
2006-06-23 17:03:20 +08:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(vmalloc_user);
|
|
|
|
|
2005-10-30 09:15:41 +08:00
|
|
|
/**
|
2019-03-06 07:48:36 +08:00
|
|
|
* vmalloc_node - allocate memory on a specific node
|
|
|
|
* @size: allocation size
|
|
|
|
* @node: numa node
|
2005-10-30 09:15:41 +08:00
|
|
|
*
|
2019-03-06 07:48:36 +08:00
|
|
|
* Allocate enough pages to cover @size from the page level
|
|
|
|
* allocator and map them into contiguous kernel virtual space.
|
2005-10-30 09:15:41 +08:00
|
|
|
*
|
2019-03-06 07:48:36 +08:00
|
|
|
* For tight control over page level allocator and protection flags
|
|
|
|
* use __vmalloc() instead.
|
2019-03-06 07:48:42 +08:00
|
|
|
*
|
|
|
|
* Return: pointer to the allocated memory or %NULL on error
|
2005-10-30 09:15:41 +08:00
|
|
|
*/
|
|
|
|
void *vmalloc_node(unsigned long size, int node)
|
|
|
|
{
|
2017-05-09 06:57:44 +08:00
|
|
|
return __vmalloc_node(size, 1, GFP_KERNEL, PAGE_KERNEL,
|
2008-04-28 17:12:42 +08:00
|
|
|
node, __builtin_return_address(0));
|
2005-10-30 09:15:41 +08:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(vmalloc_node);
|
|
|
|
|
2010-10-27 05:22:06 +08:00
|
|
|
/**
|
|
|
|
* vzalloc_node - allocate memory on a specific node with zero fill
|
|
|
|
* @size: allocation size
|
|
|
|
* @node: numa node
|
|
|
|
*
|
|
|
|
* Allocate enough pages to cover @size from the page level
|
|
|
|
* allocator and map them into contiguous kernel virtual space.
|
|
|
|
* The memory allocated is set to zero.
|
|
|
|
*
|
|
|
|
* For tight control over page level allocator and protection flags
|
|
|
|
* use __vmalloc_node() instead.
|
2019-03-06 07:48:42 +08:00
|
|
|
*
|
|
|
|
* Return: pointer to the allocated memory or %NULL on error
|
2010-10-27 05:22:06 +08:00
|
|
|
*/
|
|
|
|
void *vzalloc_node(unsigned long size, int node)
|
|
|
|
{
|
|
|
|
return __vmalloc_node_flags(size, node,
|
2017-05-09 06:57:44 +08:00
|
|
|
GFP_KERNEL | __GFP_ZERO);
|
2010-10-27 05:22:06 +08:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(vzalloc_node);
|
|
|
|
|
bpf: Add mmap() support for BPF_MAP_TYPE_ARRAY
Add ability to memory-map contents of BPF array map. This is extremely useful
for working with BPF global data from userspace programs. It allows to avoid
typical bpf_map_{lookup,update}_elem operations, improving both performance
and usability.
There had to be special considerations for map freezing, to avoid having
writable memory view into a frozen map. To solve this issue, map freezing and
mmap-ing is happening under mutex now:
- if map is already frozen, no writable mapping is allowed;
- if map has writable memory mappings active (accounted in map->writecnt),
map freezing will keep failing with -EBUSY;
- once number of writable memory mappings drops to zero, map freezing can be
performed again.
Only non-per-CPU plain arrays are supported right now. Maps with spinlocks
can't be memory mapped either.
For BPF_F_MMAPABLE array, memory allocation has to be done through vmalloc()
to be mmap()'able. We also need to make sure that array data memory is
page-sized and page-aligned, so we over-allocate memory in such a way that
struct bpf_array is at the end of a single page of memory with array->value
being aligned with the start of the second page. On deallocation we need to
accomodate this memory arrangement to free vmalloc()'ed memory correctly.
One important consideration regarding how memory-mapping subsystem functions.
Memory-mapping subsystem provides few optional callbacks, among them open()
and close(). close() is called for each memory region that is unmapped, so
that users can decrease their reference counters and free up resources, if
necessary. open() is *almost* symmetrical: it's called for each memory region
that is being mapped, **except** the very first one. So bpf_map_mmap does
initial refcnt bump, while open() will do any extra ones after that. Thus
number of close() calls is equal to number of open() calls plus one more.
Signed-off-by: Andrii Nakryiko <andriin@fb.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Song Liu <songliubraving@fb.com>
Acked-by: John Fastabend <john.fastabend@gmail.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Link: https://lore.kernel.org/bpf/20191117172806.2195367-4-andriin@fb.com
2019-11-18 01:28:04 +08:00
|
|
|
/**
|
|
|
|
* vmalloc_user_node_flags - allocate memory for userspace on a specific node
|
|
|
|
* @size: allocation size
|
|
|
|
* @node: numa node
|
|
|
|
* @flags: flags for the page level allocator
|
|
|
|
*
|
|
|
|
* The resulting memory area is zeroed so it can be mapped to userspace
|
|
|
|
* without leaking data.
|
|
|
|
*
|
|
|
|
* Return: pointer to the allocated memory or %NULL on error
|
|
|
|
*/
|
|
|
|
void *vmalloc_user_node_flags(unsigned long size, int node, gfp_t flags)
|
|
|
|
{
|
|
|
|
return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
|
|
|
|
flags | __GFP_ZERO, PAGE_KERNEL,
|
|
|
|
VM_USERMAP, node,
|
|
|
|
__builtin_return_address(0));
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(vmalloc_user_node_flags);
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
/**
|
2019-03-06 07:48:36 +08:00
|
|
|
* vmalloc_exec - allocate virtually contiguous, executable memory
|
|
|
|
* @size: allocation size
|
2005-04-17 06:20:36 +08:00
|
|
|
*
|
2019-03-06 07:48:36 +08:00
|
|
|
* Kernel-internal function to allocate enough pages to cover @size
|
|
|
|
* the page level allocator and map them into contiguous and
|
|
|
|
* executable kernel virtual space.
|
2005-04-17 06:20:36 +08:00
|
|
|
*
|
2019-03-06 07:48:36 +08:00
|
|
|
* For tight control over page level allocator and protection flags
|
|
|
|
* use __vmalloc() instead.
|
2019-03-06 07:48:42 +08:00
|
|
|
*
|
|
|
|
* Return: pointer to the allocated memory or %NULL on error
|
2005-04-17 06:20:36 +08:00
|
|
|
*/
|
|
|
|
void *vmalloc_exec(unsigned long size)
|
|
|
|
{
|
2019-04-26 08:11:36 +08:00
|
|
|
return __vmalloc_node_range(size, 1, VMALLOC_START, VMALLOC_END,
|
|
|
|
GFP_KERNEL, PAGE_KERNEL_EXEC, VM_FLUSH_RESET_PERMS,
|
|
|
|
NUMA_NO_NODE, __builtin_return_address(0));
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
2007-05-03 01:27:12 +08:00
|
|
|
#if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
|
2018-02-22 06:46:01 +08:00
|
|
|
#define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
|
2007-05-03 01:27:12 +08:00
|
|
|
#elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
|
2018-02-22 06:46:01 +08:00
|
|
|
#define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
|
2007-05-03 01:27:12 +08:00
|
|
|
#else
|
2018-02-22 06:46:01 +08:00
|
|
|
/*
|
|
|
|
* 64b systems should always have either DMA or DMA32 zones. For others
|
|
|
|
* GFP_DMA32 should do the right thing and use the normal zone.
|
|
|
|
*/
|
|
|
|
#define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
|
2007-05-03 01:27:12 +08:00
|
|
|
#endif
|
|
|
|
|
2005-04-17 06:20:36 +08:00
|
|
|
/**
|
2019-03-06 07:48:36 +08:00
|
|
|
* vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
|
|
|
|
* @size: allocation size
|
2005-04-17 06:20:36 +08:00
|
|
|
*
|
2019-03-06 07:48:36 +08:00
|
|
|
* Allocate enough 32bit PA addressable pages to cover @size from the
|
|
|
|
* page level allocator and map them into contiguous kernel virtual space.
|
2019-03-06 07:48:42 +08:00
|
|
|
*
|
|
|
|
* Return: pointer to the allocated memory or %NULL on error
|
2005-04-17 06:20:36 +08:00
|
|
|
*/
|
|
|
|
void *vmalloc_32(unsigned long size)
|
|
|
|
{
|
2009-09-22 03:22:34 +08:00
|
|
|
return __vmalloc_node(size, 1, GFP_VMALLOC32, PAGE_KERNEL,
|
2013-02-23 08:35:36 +08:00
|
|
|
NUMA_NO_NODE, __builtin_return_address(0));
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(vmalloc_32);
|
|
|
|
|
2006-06-23 17:03:20 +08:00
|
|
|
/**
|
2006-09-27 16:50:13 +08:00
|
|
|
* vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
|
2019-03-06 07:48:36 +08:00
|
|
|
* @size: allocation size
|
2006-09-27 16:50:13 +08:00
|
|
|
*
|
|
|
|
* The resulting memory area is 32bit addressable and zeroed so it can be
|
|
|
|
* mapped to userspace without leaking data.
|
2019-03-06 07:48:42 +08:00
|
|
|
*
|
|
|
|
* Return: pointer to the allocated memory or %NULL on error
|
2006-06-23 17:03:20 +08:00
|
|
|
*/
|
|
|
|
void *vmalloc_32_user(unsigned long size)
|
|
|
|
{
|
2019-03-06 07:43:27 +08:00
|
|
|
return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
|
|
|
|
GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
|
|
|
|
VM_USERMAP, NUMA_NO_NODE,
|
|
|
|
__builtin_return_address(0));
|
2006-06-23 17:03:20 +08:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(vmalloc_32_user);
|
|
|
|
|
2009-09-22 08:02:34 +08:00
|
|
|
/*
|
|
|
|
* small helper routine , copy contents to buf from addr.
|
|
|
|
* If the page is not present, fill zero.
|
|
|
|
*/
|
|
|
|
|
|
|
|
static int aligned_vread(char *buf, char *addr, unsigned long count)
|
|
|
|
{
|
|
|
|
struct page *p;
|
|
|
|
int copied = 0;
|
|
|
|
|
|
|
|
while (count) {
|
|
|
|
unsigned long offset, length;
|
|
|
|
|
2015-11-06 10:46:51 +08:00
|
|
|
offset = offset_in_page(addr);
|
2009-09-22 08:02:34 +08:00
|
|
|
length = PAGE_SIZE - offset;
|
|
|
|
if (length > count)
|
|
|
|
length = count;
|
|
|
|
p = vmalloc_to_page(addr);
|
|
|
|
/*
|
|
|
|
* To do safe access to this _mapped_ area, we need
|
|
|
|
* lock. But adding lock here means that we need to add
|
|
|
|
* overhead of vmalloc()/vfree() calles for this _debug_
|
|
|
|
* interface, rarely used. Instead of that, we'll use
|
|
|
|
* kmap() and get small overhead in this access function.
|
|
|
|
*/
|
|
|
|
if (p) {
|
|
|
|
/*
|
|
|
|
* we can expect USER0 is not used (see vread/vwrite's
|
|
|
|
* function description)
|
|
|
|
*/
|
2011-11-25 23:14:39 +08:00
|
|
|
void *map = kmap_atomic(p);
|
2009-09-22 08:02:34 +08:00
|
|
|
memcpy(buf, map + offset, length);
|
2011-11-25 23:14:39 +08:00
|
|
|
kunmap_atomic(map);
|
2009-09-22 08:02:34 +08:00
|
|
|
} else
|
|
|
|
memset(buf, 0, length);
|
|
|
|
|
|
|
|
addr += length;
|
|
|
|
buf += length;
|
|
|
|
copied += length;
|
|
|
|
count -= length;
|
|
|
|
}
|
|
|
|
return copied;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int aligned_vwrite(char *buf, char *addr, unsigned long count)
|
|
|
|
{
|
|
|
|
struct page *p;
|
|
|
|
int copied = 0;
|
|
|
|
|
|
|
|
while (count) {
|
|
|
|
unsigned long offset, length;
|
|
|
|
|
2015-11-06 10:46:51 +08:00
|
|
|
offset = offset_in_page(addr);
|
2009-09-22 08:02:34 +08:00
|
|
|
length = PAGE_SIZE - offset;
|
|
|
|
if (length > count)
|
|
|
|
length = count;
|
|
|
|
p = vmalloc_to_page(addr);
|
|
|
|
/*
|
|
|
|
* To do safe access to this _mapped_ area, we need
|
|
|
|
* lock. But adding lock here means that we need to add
|
|
|
|
* overhead of vmalloc()/vfree() calles for this _debug_
|
|
|
|
* interface, rarely used. Instead of that, we'll use
|
|
|
|
* kmap() and get small overhead in this access function.
|
|
|
|
*/
|
|
|
|
if (p) {
|
|
|
|
/*
|
|
|
|
* we can expect USER0 is not used (see vread/vwrite's
|
|
|
|
* function description)
|
|
|
|
*/
|
2011-11-25 23:14:39 +08:00
|
|
|
void *map = kmap_atomic(p);
|
2009-09-22 08:02:34 +08:00
|
|
|
memcpy(map + offset, buf, length);
|
2011-11-25 23:14:39 +08:00
|
|
|
kunmap_atomic(map);
|
2009-09-22 08:02:34 +08:00
|
|
|
}
|
|
|
|
addr += length;
|
|
|
|
buf += length;
|
|
|
|
copied += length;
|
|
|
|
count -= length;
|
|
|
|
}
|
|
|
|
return copied;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
2019-03-06 07:48:36 +08:00
|
|
|
* vread() - read vmalloc area in a safe way.
|
|
|
|
* @buf: buffer for reading data
|
|
|
|
* @addr: vm address.
|
|
|
|
* @count: number of bytes to be read.
|
|
|
|
*
|
|
|
|
* This function checks that addr is a valid vmalloc'ed area, and
|
|
|
|
* copy data from that area to a given buffer. If the given memory range
|
|
|
|
* of [addr...addr+count) includes some valid address, data is copied to
|
|
|
|
* proper area of @buf. If there are memory holes, they'll be zero-filled.
|
|
|
|
* IOREMAP area is treated as memory hole and no copy is done.
|
|
|
|
*
|
|
|
|
* If [addr...addr+count) doesn't includes any intersects with alive
|
|
|
|
* vm_struct area, returns 0. @buf should be kernel's buffer.
|
|
|
|
*
|
|
|
|
* Note: In usual ops, vread() is never necessary because the caller
|
|
|
|
* should know vmalloc() area is valid and can use memcpy().
|
|
|
|
* This is for routines which have to access vmalloc area without
|
2019-07-12 11:59:06 +08:00
|
|
|
* any information, as /dev/kmem.
|
2019-03-06 07:48:42 +08:00
|
|
|
*
|
|
|
|
* Return: number of bytes for which addr and buf should be increased
|
|
|
|
* (same number as @count) or %0 if [addr...addr+count) doesn't
|
|
|
|
* include any intersection with valid vmalloc area
|
2009-09-22 08:02:34 +08:00
|
|
|
*/
|
2005-04-17 06:20:36 +08:00
|
|
|
long vread(char *buf, char *addr, unsigned long count)
|
|
|
|
{
|
mm, vmalloc: iterate vmap_area_list, instead of vmlist in vread/vwrite()
Now, when we hold a vmap_area_lock, va->vm can't be discarded. So we can
safely access to va->vm when iterating a vmap_area_list with holding a
vmap_area_lock. With this property, change iterating vmlist codes in
vread/vwrite() to iterating vmap_area_list.
There is a little difference relate to lock, because vmlist_lock is mutex,
but, vmap_area_lock is spin_lock. It may introduce a spinning overhead
during vread/vwrite() is executing. But, these are debug-oriented
functions, so this overhead is not real problem for common case.
Signed-off-by: Joonsoo Kim <js1304@gmail.com>
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Atsushi Kumagai <kumagai-atsushi@mxc.nes.nec.co.jp>
Cc: Chris Metcalf <cmetcalf@tilera.com>
Cc: Dave Anderson <anderson@redhat.com>
Cc: Eric Biederman <ebiederm@xmission.com>
Cc: Guan Xuetao <gxt@mprc.pku.edu.cn>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Vivek Goyal <vgoyal@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-04-30 06:07:32 +08:00
|
|
|
struct vmap_area *va;
|
|
|
|
struct vm_struct *vm;
|
2005-04-17 06:20:36 +08:00
|
|
|
char *vaddr, *buf_start = buf;
|
2009-09-22 08:02:34 +08:00
|
|
|
unsigned long buflen = count;
|
2005-04-17 06:20:36 +08:00
|
|
|
unsigned long n;
|
|
|
|
|
|
|
|
/* Don't allow overflow */
|
|
|
|
if ((unsigned long) addr + count < count)
|
|
|
|
count = -(unsigned long) addr;
|
|
|
|
|
mm, vmalloc: iterate vmap_area_list, instead of vmlist in vread/vwrite()
Now, when we hold a vmap_area_lock, va->vm can't be discarded. So we can
safely access to va->vm when iterating a vmap_area_list with holding a
vmap_area_lock. With this property, change iterating vmlist codes in
vread/vwrite() to iterating vmap_area_list.
There is a little difference relate to lock, because vmlist_lock is mutex,
but, vmap_area_lock is spin_lock. It may introduce a spinning overhead
during vread/vwrite() is executing. But, these are debug-oriented
functions, so this overhead is not real problem for common case.
Signed-off-by: Joonsoo Kim <js1304@gmail.com>
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Atsushi Kumagai <kumagai-atsushi@mxc.nes.nec.co.jp>
Cc: Chris Metcalf <cmetcalf@tilera.com>
Cc: Dave Anderson <anderson@redhat.com>
Cc: Eric Biederman <ebiederm@xmission.com>
Cc: Guan Xuetao <gxt@mprc.pku.edu.cn>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Vivek Goyal <vgoyal@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-04-30 06:07:32 +08:00
|
|
|
spin_lock(&vmap_area_lock);
|
|
|
|
list_for_each_entry(va, &vmap_area_list, list) {
|
|
|
|
if (!count)
|
|
|
|
break;
|
|
|
|
|
2019-09-24 06:36:39 +08:00
|
|
|
if (!va->vm)
|
mm, vmalloc: iterate vmap_area_list, instead of vmlist in vread/vwrite()
Now, when we hold a vmap_area_lock, va->vm can't be discarded. So we can
safely access to va->vm when iterating a vmap_area_list with holding a
vmap_area_lock. With this property, change iterating vmlist codes in
vread/vwrite() to iterating vmap_area_list.
There is a little difference relate to lock, because vmlist_lock is mutex,
but, vmap_area_lock is spin_lock. It may introduce a spinning overhead
during vread/vwrite() is executing. But, these are debug-oriented
functions, so this overhead is not real problem for common case.
Signed-off-by: Joonsoo Kim <js1304@gmail.com>
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Atsushi Kumagai <kumagai-atsushi@mxc.nes.nec.co.jp>
Cc: Chris Metcalf <cmetcalf@tilera.com>
Cc: Dave Anderson <anderson@redhat.com>
Cc: Eric Biederman <ebiederm@xmission.com>
Cc: Guan Xuetao <gxt@mprc.pku.edu.cn>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Vivek Goyal <vgoyal@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-04-30 06:07:32 +08:00
|
|
|
continue;
|
|
|
|
|
|
|
|
vm = va->vm;
|
|
|
|
vaddr = (char *) vm->addr;
|
2013-09-12 05:22:42 +08:00
|
|
|
if (addr >= vaddr + get_vm_area_size(vm))
|
2005-04-17 06:20:36 +08:00
|
|
|
continue;
|
|
|
|
while (addr < vaddr) {
|
|
|
|
if (count == 0)
|
|
|
|
goto finished;
|
|
|
|
*buf = '\0';
|
|
|
|
buf++;
|
|
|
|
addr++;
|
|
|
|
count--;
|
|
|
|
}
|
2013-09-12 05:22:42 +08:00
|
|
|
n = vaddr + get_vm_area_size(vm) - addr;
|
2009-09-22 08:02:34 +08:00
|
|
|
if (n > count)
|
|
|
|
n = count;
|
mm, vmalloc: iterate vmap_area_list, instead of vmlist in vread/vwrite()
Now, when we hold a vmap_area_lock, va->vm can't be discarded. So we can
safely access to va->vm when iterating a vmap_area_list with holding a
vmap_area_lock. With this property, change iterating vmlist codes in
vread/vwrite() to iterating vmap_area_list.
There is a little difference relate to lock, because vmlist_lock is mutex,
but, vmap_area_lock is spin_lock. It may introduce a spinning overhead
during vread/vwrite() is executing. But, these are debug-oriented
functions, so this overhead is not real problem for common case.
Signed-off-by: Joonsoo Kim <js1304@gmail.com>
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Atsushi Kumagai <kumagai-atsushi@mxc.nes.nec.co.jp>
Cc: Chris Metcalf <cmetcalf@tilera.com>
Cc: Dave Anderson <anderson@redhat.com>
Cc: Eric Biederman <ebiederm@xmission.com>
Cc: Guan Xuetao <gxt@mprc.pku.edu.cn>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Vivek Goyal <vgoyal@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-04-30 06:07:32 +08:00
|
|
|
if (!(vm->flags & VM_IOREMAP))
|
2009-09-22 08:02:34 +08:00
|
|
|
aligned_vread(buf, addr, n);
|
|
|
|
else /* IOREMAP area is treated as memory hole */
|
|
|
|
memset(buf, 0, n);
|
|
|
|
buf += n;
|
|
|
|
addr += n;
|
|
|
|
count -= n;
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
finished:
|
mm, vmalloc: iterate vmap_area_list, instead of vmlist in vread/vwrite()
Now, when we hold a vmap_area_lock, va->vm can't be discarded. So we can
safely access to va->vm when iterating a vmap_area_list with holding a
vmap_area_lock. With this property, change iterating vmlist codes in
vread/vwrite() to iterating vmap_area_list.
There is a little difference relate to lock, because vmlist_lock is mutex,
but, vmap_area_lock is spin_lock. It may introduce a spinning overhead
during vread/vwrite() is executing. But, these are debug-oriented
functions, so this overhead is not real problem for common case.
Signed-off-by: Joonsoo Kim <js1304@gmail.com>
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Atsushi Kumagai <kumagai-atsushi@mxc.nes.nec.co.jp>
Cc: Chris Metcalf <cmetcalf@tilera.com>
Cc: Dave Anderson <anderson@redhat.com>
Cc: Eric Biederman <ebiederm@xmission.com>
Cc: Guan Xuetao <gxt@mprc.pku.edu.cn>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Vivek Goyal <vgoyal@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-04-30 06:07:32 +08:00
|
|
|
spin_unlock(&vmap_area_lock);
|
2009-09-22 08:02:34 +08:00
|
|
|
|
|
|
|
if (buf == buf_start)
|
|
|
|
return 0;
|
|
|
|
/* zero-fill memory holes */
|
|
|
|
if (buf != buf_start + buflen)
|
|
|
|
memset(buf, 0, buflen - (buf - buf_start));
|
|
|
|
|
|
|
|
return buflen;
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
|
2009-09-22 08:02:34 +08:00
|
|
|
/**
|
2019-03-06 07:48:36 +08:00
|
|
|
* vwrite() - write vmalloc area in a safe way.
|
|
|
|
* @buf: buffer for source data
|
|
|
|
* @addr: vm address.
|
|
|
|
* @count: number of bytes to be read.
|
|
|
|
*
|
|
|
|
* This function checks that addr is a valid vmalloc'ed area, and
|
|
|
|
* copy data from a buffer to the given addr. If specified range of
|
|
|
|
* [addr...addr+count) includes some valid address, data is copied from
|
|
|
|
* proper area of @buf. If there are memory holes, no copy to hole.
|
|
|
|
* IOREMAP area is treated as memory hole and no copy is done.
|
|
|
|
*
|
|
|
|
* If [addr...addr+count) doesn't includes any intersects with alive
|
|
|
|
* vm_struct area, returns 0. @buf should be kernel's buffer.
|
|
|
|
*
|
|
|
|
* Note: In usual ops, vwrite() is never necessary because the caller
|
|
|
|
* should know vmalloc() area is valid and can use memcpy().
|
|
|
|
* This is for routines which have to access vmalloc area without
|
2019-07-12 11:59:06 +08:00
|
|
|
* any information, as /dev/kmem.
|
2019-03-06 07:48:42 +08:00
|
|
|
*
|
|
|
|
* Return: number of bytes for which addr and buf should be
|
|
|
|
* increased (same number as @count) or %0 if [addr...addr+count)
|
|
|
|
* doesn't include any intersection with valid vmalloc area
|
2009-09-22 08:02:34 +08:00
|
|
|
*/
|
2005-04-17 06:20:36 +08:00
|
|
|
long vwrite(char *buf, char *addr, unsigned long count)
|
|
|
|
{
|
mm, vmalloc: iterate vmap_area_list, instead of vmlist in vread/vwrite()
Now, when we hold a vmap_area_lock, va->vm can't be discarded. So we can
safely access to va->vm when iterating a vmap_area_list with holding a
vmap_area_lock. With this property, change iterating vmlist codes in
vread/vwrite() to iterating vmap_area_list.
There is a little difference relate to lock, because vmlist_lock is mutex,
but, vmap_area_lock is spin_lock. It may introduce a spinning overhead
during vread/vwrite() is executing. But, these are debug-oriented
functions, so this overhead is not real problem for common case.
Signed-off-by: Joonsoo Kim <js1304@gmail.com>
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Atsushi Kumagai <kumagai-atsushi@mxc.nes.nec.co.jp>
Cc: Chris Metcalf <cmetcalf@tilera.com>
Cc: Dave Anderson <anderson@redhat.com>
Cc: Eric Biederman <ebiederm@xmission.com>
Cc: Guan Xuetao <gxt@mprc.pku.edu.cn>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Vivek Goyal <vgoyal@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-04-30 06:07:32 +08:00
|
|
|
struct vmap_area *va;
|
|
|
|
struct vm_struct *vm;
|
2009-09-22 08:02:34 +08:00
|
|
|
char *vaddr;
|
|
|
|
unsigned long n, buflen;
|
|
|
|
int copied = 0;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
|
|
|
/* Don't allow overflow */
|
|
|
|
if ((unsigned long) addr + count < count)
|
|
|
|
count = -(unsigned long) addr;
|
2009-09-22 08:02:34 +08:00
|
|
|
buflen = count;
|
2005-04-17 06:20:36 +08:00
|
|
|
|
mm, vmalloc: iterate vmap_area_list, instead of vmlist in vread/vwrite()
Now, when we hold a vmap_area_lock, va->vm can't be discarded. So we can
safely access to va->vm when iterating a vmap_area_list with holding a
vmap_area_lock. With this property, change iterating vmlist codes in
vread/vwrite() to iterating vmap_area_list.
There is a little difference relate to lock, because vmlist_lock is mutex,
but, vmap_area_lock is spin_lock. It may introduce a spinning overhead
during vread/vwrite() is executing. But, these are debug-oriented
functions, so this overhead is not real problem for common case.
Signed-off-by: Joonsoo Kim <js1304@gmail.com>
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Atsushi Kumagai <kumagai-atsushi@mxc.nes.nec.co.jp>
Cc: Chris Metcalf <cmetcalf@tilera.com>
Cc: Dave Anderson <anderson@redhat.com>
Cc: Eric Biederman <ebiederm@xmission.com>
Cc: Guan Xuetao <gxt@mprc.pku.edu.cn>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Vivek Goyal <vgoyal@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-04-30 06:07:32 +08:00
|
|
|
spin_lock(&vmap_area_lock);
|
|
|
|
list_for_each_entry(va, &vmap_area_list, list) {
|
|
|
|
if (!count)
|
|
|
|
break;
|
|
|
|
|
2019-09-24 06:36:39 +08:00
|
|
|
if (!va->vm)
|
mm, vmalloc: iterate vmap_area_list, instead of vmlist in vread/vwrite()
Now, when we hold a vmap_area_lock, va->vm can't be discarded. So we can
safely access to va->vm when iterating a vmap_area_list with holding a
vmap_area_lock. With this property, change iterating vmlist codes in
vread/vwrite() to iterating vmap_area_list.
There is a little difference relate to lock, because vmlist_lock is mutex,
but, vmap_area_lock is spin_lock. It may introduce a spinning overhead
during vread/vwrite() is executing. But, these are debug-oriented
functions, so this overhead is not real problem for common case.
Signed-off-by: Joonsoo Kim <js1304@gmail.com>
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Atsushi Kumagai <kumagai-atsushi@mxc.nes.nec.co.jp>
Cc: Chris Metcalf <cmetcalf@tilera.com>
Cc: Dave Anderson <anderson@redhat.com>
Cc: Eric Biederman <ebiederm@xmission.com>
Cc: Guan Xuetao <gxt@mprc.pku.edu.cn>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Vivek Goyal <vgoyal@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-04-30 06:07:32 +08:00
|
|
|
continue;
|
|
|
|
|
|
|
|
vm = va->vm;
|
|
|
|
vaddr = (char *) vm->addr;
|
2013-09-12 05:22:42 +08:00
|
|
|
if (addr >= vaddr + get_vm_area_size(vm))
|
2005-04-17 06:20:36 +08:00
|
|
|
continue;
|
|
|
|
while (addr < vaddr) {
|
|
|
|
if (count == 0)
|
|
|
|
goto finished;
|
|
|
|
buf++;
|
|
|
|
addr++;
|
|
|
|
count--;
|
|
|
|
}
|
2013-09-12 05:22:42 +08:00
|
|
|
n = vaddr + get_vm_area_size(vm) - addr;
|
2009-09-22 08:02:34 +08:00
|
|
|
if (n > count)
|
|
|
|
n = count;
|
mm, vmalloc: iterate vmap_area_list, instead of vmlist in vread/vwrite()
Now, when we hold a vmap_area_lock, va->vm can't be discarded. So we can
safely access to va->vm when iterating a vmap_area_list with holding a
vmap_area_lock. With this property, change iterating vmlist codes in
vread/vwrite() to iterating vmap_area_list.
There is a little difference relate to lock, because vmlist_lock is mutex,
but, vmap_area_lock is spin_lock. It may introduce a spinning overhead
during vread/vwrite() is executing. But, these are debug-oriented
functions, so this overhead is not real problem for common case.
Signed-off-by: Joonsoo Kim <js1304@gmail.com>
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Atsushi Kumagai <kumagai-atsushi@mxc.nes.nec.co.jp>
Cc: Chris Metcalf <cmetcalf@tilera.com>
Cc: Dave Anderson <anderson@redhat.com>
Cc: Eric Biederman <ebiederm@xmission.com>
Cc: Guan Xuetao <gxt@mprc.pku.edu.cn>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Vivek Goyal <vgoyal@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-04-30 06:07:32 +08:00
|
|
|
if (!(vm->flags & VM_IOREMAP)) {
|
2009-09-22 08:02:34 +08:00
|
|
|
aligned_vwrite(buf, addr, n);
|
|
|
|
copied++;
|
|
|
|
}
|
|
|
|
buf += n;
|
|
|
|
addr += n;
|
|
|
|
count -= n;
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
|
|
|
finished:
|
mm, vmalloc: iterate vmap_area_list, instead of vmlist in vread/vwrite()
Now, when we hold a vmap_area_lock, va->vm can't be discarded. So we can
safely access to va->vm when iterating a vmap_area_list with holding a
vmap_area_lock. With this property, change iterating vmlist codes in
vread/vwrite() to iterating vmap_area_list.
There is a little difference relate to lock, because vmlist_lock is mutex,
but, vmap_area_lock is spin_lock. It may introduce a spinning overhead
during vread/vwrite() is executing. But, these are debug-oriented
functions, so this overhead is not real problem for common case.
Signed-off-by: Joonsoo Kim <js1304@gmail.com>
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Atsushi Kumagai <kumagai-atsushi@mxc.nes.nec.co.jp>
Cc: Chris Metcalf <cmetcalf@tilera.com>
Cc: Dave Anderson <anderson@redhat.com>
Cc: Eric Biederman <ebiederm@xmission.com>
Cc: Guan Xuetao <gxt@mprc.pku.edu.cn>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Vivek Goyal <vgoyal@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-04-30 06:07:32 +08:00
|
|
|
spin_unlock(&vmap_area_lock);
|
2009-09-22 08:02:34 +08:00
|
|
|
if (!copied)
|
|
|
|
return 0;
|
|
|
|
return buflen;
|
2005-04-17 06:20:36 +08:00
|
|
|
}
|
2006-06-23 17:03:20 +08:00
|
|
|
|
|
|
|
/**
|
2019-03-06 07:48:36 +08:00
|
|
|
* remap_vmalloc_range_partial - map vmalloc pages to userspace
|
|
|
|
* @vma: vma to cover
|
|
|
|
* @uaddr: target user address to start at
|
|
|
|
* @kaddr: virtual address of vmalloc kernel memory
|
|
|
|
* @size: size of map area
|
2008-03-20 08:00:40 +08:00
|
|
|
*
|
2019-03-06 07:48:36 +08:00
|
|
|
* Returns: 0 for success, -Exxx on failure
|
2006-06-23 17:03:20 +08:00
|
|
|
*
|
2019-03-06 07:48:36 +08:00
|
|
|
* This function checks that @kaddr is a valid vmalloc'ed area,
|
|
|
|
* and that it is big enough to cover the range starting at
|
|
|
|
* @uaddr in @vma. Will return failure if that criteria isn't
|
|
|
|
* met.
|
2006-06-23 17:03:20 +08:00
|
|
|
*
|
2019-03-06 07:48:36 +08:00
|
|
|
* Similar to remap_pfn_range() (see mm/memory.c)
|
2006-06-23 17:03:20 +08:00
|
|
|
*/
|
2013-07-04 06:02:18 +08:00
|
|
|
int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
|
|
|
|
void *kaddr, unsigned long size)
|
2006-06-23 17:03:20 +08:00
|
|
|
{
|
|
|
|
struct vm_struct *area;
|
|
|
|
|
2013-07-04 06:02:18 +08:00
|
|
|
size = PAGE_ALIGN(size);
|
|
|
|
|
|
|
|
if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
|
2006-06-23 17:03:20 +08:00
|
|
|
return -EINVAL;
|
|
|
|
|
2013-07-04 06:02:18 +08:00
|
|
|
area = find_vm_area(kaddr);
|
2006-06-23 17:03:20 +08:00
|
|
|
if (!area)
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
return -EINVAL;
|
2006-06-23 17:03:20 +08:00
|
|
|
|
2019-06-03 14:55:13 +08:00
|
|
|
if (!(area->flags & (VM_USERMAP | VM_DMA_COHERENT)))
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
return -EINVAL;
|
2006-06-23 17:03:20 +08:00
|
|
|
|
mm/vmalloc: fix size check for remap_vmalloc_range_partial()
When VM_NO_GUARD is not set area->size includes adjacent guard page,
thus for correct size checking get_vm_area_size() should be used, but
not area->size.
This fixes possible kernel oops when userspace tries to mmap an area on
1 page bigger than was allocated by vmalloc_user() call: the size check
inside remap_vmalloc_range_partial() accounts non-existing guard page
also, so check successfully passes but vmalloc_to_page() returns NULL
(guard page does not physically exist).
The following code pattern example should trigger an oops:
static int oops_mmap(struct file *file, struct vm_area_struct *vma)
{
void *mem;
mem = vmalloc_user(4096);
BUG_ON(!mem);
/* Do not care about mem leak */
return remap_vmalloc_range(vma, mem, 0);
}
And userspace simply mmaps size + PAGE_SIZE:
mmap(NULL, 8192, PROT_WRITE|PROT_READ, MAP_PRIVATE, fd, 0);
Possible candidates for oops which do not have any explicit size
checks:
*** drivers/media/usb/stkwebcam/stk-webcam.c:
v4l_stk_mmap[789] ret = remap_vmalloc_range(vma, sbuf->buffer, 0);
Or the following one:
*** drivers/video/fbdev/core/fbmem.c
static int
fb_mmap(struct file *file, struct vm_area_struct * vma)
...
res = fb->fb_mmap(info, vma);
Where fb_mmap callback calls remap_vmalloc_range() directly without any
explicit checks:
*** drivers/video/fbdev/vfb.c
static int vfb_mmap(struct fb_info *info,
struct vm_area_struct *vma)
{
return remap_vmalloc_range(vma, (void *)info->fix.smem_start, vma->vm_pgoff);
}
Link: http://lkml.kernel.org/r/20190103145954.16942-2-rpenyaev@suse.de
Signed-off-by: Roman Penyaev <rpenyaev@suse.de>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
Cc: Joe Perches <joe@perches.com>
Cc: "Luis R. Rodriguez" <mcgrof@kernel.org>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-06 07:43:20 +08:00
|
|
|
if (kaddr + size > area->addr + get_vm_area_size(area))
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
return -EINVAL;
|
2006-06-23 17:03:20 +08:00
|
|
|
|
|
|
|
do {
|
2013-07-04 06:02:18 +08:00
|
|
|
struct page *page = vmalloc_to_page(kaddr);
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
int ret;
|
|
|
|
|
2006-06-23 17:03:20 +08:00
|
|
|
ret = vm_insert_page(vma, uaddr, page);
|
|
|
|
if (ret)
|
|
|
|
return ret;
|
|
|
|
|
|
|
|
uaddr += PAGE_SIZE;
|
2013-07-04 06:02:18 +08:00
|
|
|
kaddr += PAGE_SIZE;
|
|
|
|
size -= PAGE_SIZE;
|
|
|
|
} while (size > 0);
|
2006-06-23 17:03:20 +08:00
|
|
|
|
mm: kill vma flag VM_RESERVED and mm->reserved_vm counter
A long time ago, in v2.4, VM_RESERVED kept swapout process off VMA,
currently it lost original meaning but still has some effects:
| effect | alternative flags
-+------------------------+---------------------------------------------
1| account as reserved_vm | VM_IO
2| skip in core dump | VM_IO, VM_DONTDUMP
3| do not merge or expand | VM_IO, VM_DONTEXPAND, VM_HUGETLB, VM_PFNMAP
4| do not mlock | VM_IO, VM_DONTEXPAND, VM_HUGETLB, VM_PFNMAP
This patch removes reserved_vm counter from mm_struct. Seems like nobody
cares about it, it does not exported into userspace directly, it only
reduces total_vm showed in proc.
Thus VM_RESERVED can be replaced with VM_IO or pair VM_DONTEXPAND | VM_DONTDUMP.
remap_pfn_range() and io_remap_pfn_range() set VM_IO|VM_DONTEXPAND|VM_DONTDUMP.
remap_vmalloc_range() set VM_DONTEXPAND | VM_DONTDUMP.
[akpm@linux-foundation.org: drivers/vfio/pci/vfio_pci.c fixup]
Signed-off-by: Konstantin Khlebnikov <khlebnikov@openvz.org>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Carsten Otte <cotte@de.ibm.com>
Cc: Chris Metcalf <cmetcalf@tilera.com>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Eric Paris <eparis@redhat.com>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: James Morris <james.l.morris@oracle.com>
Cc: Jason Baron <jbaron@redhat.com>
Cc: Kentaro Takeda <takedakn@nttdata.co.jp>
Cc: Matt Helsley <matthltc@us.ibm.com>
Cc: Nick Piggin <npiggin@kernel.dk>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Robert Richter <robert.richter@amd.com>
Cc: Suresh Siddha <suresh.b.siddha@intel.com>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Venkatesh Pallipadi <venki@google.com>
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-10-09 07:29:02 +08:00
|
|
|
vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
|
2006-06-23 17:03:20 +08:00
|
|
|
|
mm: rewrite vmap layer
Rewrite the vmap allocator to use rbtrees and lazy tlb flushing, and
provide a fast, scalable percpu frontend for small vmaps (requires a
slightly different API, though).
The biggest problem with vmap is actually vunmap. Presently this requires
a global kernel TLB flush, which on most architectures is a broadcast IPI
to all CPUs to flush the cache. This is all done under a global lock. As
the number of CPUs increases, so will the number of vunmaps a scaled
workload will want to perform, and so will the cost of a global TLB flush.
This gives terrible quadratic scalability characteristics.
Another problem is that the entire vmap subsystem works under a single
lock. It is a rwlock, but it is actually taken for write in all the fast
paths, and the read locking would likely never be run concurrently anyway,
so it's just pointless.
This is a rewrite of vmap subsystem to solve those problems. The existing
vmalloc API is implemented on top of the rewritten subsystem.
The TLB flushing problem is solved by using lazy TLB unmapping. vmap
addresses do not have to be flushed immediately when they are vunmapped,
because the kernel will not reuse them again (would be a use-after-free)
until they are reallocated. So the addresses aren't allocated again until
a subsequent TLB flush. A single TLB flush then can flush multiple
vunmaps from each CPU.
XEN and PAT and such do not like deferred TLB flushing because they can't
always handle multiple aliasing virtual addresses to a physical address.
They now call vm_unmap_aliases() in order to flush any deferred mappings.
That call is very expensive (well, actually not a lot more expensive than
a single vunmap under the old scheme), however it should be OK if not
called too often.
The virtual memory extent information is stored in an rbtree rather than a
linked list to improve the algorithmic scalability.
There is a per-CPU allocator for small vmaps, which amortizes or avoids
global locking.
To use the per-CPU interface, the vm_map_ram / vm_unmap_ram interfaces
must be used in place of vmap and vunmap. Vmalloc does not use these
interfaces at the moment, so it will not be quite so scalable (although it
will use lazy TLB flushing).
As a quick test of performance, I ran a test that loops in the kernel,
linearly mapping then touching then unmapping 4 pages. Different numbers
of tests were run in parallel on an 4 core, 2 socket opteron. Results are
in nanoseconds per map+touch+unmap.
threads vanilla vmap rewrite
1 14700 2900
2 33600 3000
4 49500 2800
8 70631 2900
So with a 8 cores, the rewritten version is already 25x faster.
In a slightly more realistic test (although with an older and less
scalable version of the patch), I ripped the not-very-good vunmap batching
code out of XFS, and implemented the large buffer mapping with vm_map_ram
and vm_unmap_ram... along with a couple of other tricks, I was able to
speed up a large directory workload by 20x on a 64 CPU system. I believe
vmap/vunmap is actually sped up a lot more than 20x on such a system, but
I'm running into other locks now. vmap is pretty well blown off the
profiles.
Before:
1352059 total 0.1401
798784 _write_lock 8320.6667 <- vmlist_lock
529313 default_idle 1181.5022
15242 smp_call_function 15.8771 <- vmap tlb flushing
2472 __get_vm_area_node 1.9312 <- vmap
1762 remove_vm_area 4.5885 <- vunmap
316 map_vm_area 0.2297 <- vmap
312 kfree 0.1950
300 _spin_lock 3.1250
252 sn_send_IPI_phys 0.4375 <- tlb flushing
238 vmap 0.8264 <- vmap
216 find_lock_page 0.5192
196 find_next_bit 0.3603
136 sn2_send_IPI 0.2024
130 pio_phys_write_mmr 2.0312
118 unmap_kernel_range 0.1229
After:
78406 total 0.0081
40053 default_idle 89.4040
33576 ia64_spinlock_contention 349.7500
1650 _spin_lock 17.1875
319 __reg_op 0.5538
281 _atomic_dec_and_lock 1.0977
153 mutex_unlock 1.5938
123 iget_locked 0.1671
117 xfs_dir_lookup 0.1662
117 dput 0.1406
114 xfs_iget_core 0.0268
92 xfs_da_hashname 0.1917
75 d_alloc 0.0670
68 vmap_page_range 0.0462 <- vmap
58 kmem_cache_alloc 0.0604
57 memset 0.0540
52 rb_next 0.1625
50 __copy_user 0.0208
49 bitmap_find_free_region 0.2188 <- vmap
46 ia64_sn_udelay 0.1106
45 find_inode_fast 0.1406
42 memcmp 0.2188
42 finish_task_switch 0.1094
42 __d_lookup 0.0410
40 radix_tree_lookup_slot 0.1250
37 _spin_unlock_irqrestore 0.3854
36 xfs_bmapi 0.0050
36 kmem_cache_free 0.0256
35 xfs_vn_getattr 0.0322
34 radix_tree_lookup 0.1062
33 __link_path_walk 0.0035
31 xfs_da_do_buf 0.0091
30 _xfs_buf_find 0.0204
28 find_get_page 0.0875
27 xfs_iread 0.0241
27 __strncpy_from_user 0.2812
26 _xfs_buf_initialize 0.0406
24 _xfs_buf_lookup_pages 0.0179
24 vunmap_page_range 0.0250 <- vunmap
23 find_lock_page 0.0799
22 vm_map_ram 0.0087 <- vmap
20 kfree 0.0125
19 put_page 0.0330
18 __kmalloc 0.0176
17 xfs_da_node_lookup_int 0.0086
17 _read_lock 0.0885
17 page_waitqueue 0.0664
vmap has gone from being the top 5 on the profiles and flushing the crap
out of all TLBs, to using less than 1% of kernel time.
[akpm@linux-foundation.org: cleanups, section fix]
[akpm@linux-foundation.org: fix build on alpha]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Cc: Krzysztof Helt <krzysztof.h1@poczta.fm>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 11:27:03 +08:00
|
|
|
return 0;
|
2006-06-23 17:03:20 +08:00
|
|
|
}
|
2013-07-04 06:02:18 +08:00
|
|
|
EXPORT_SYMBOL(remap_vmalloc_range_partial);
|
|
|
|
|
|
|
|
/**
|
2019-03-06 07:48:36 +08:00
|
|
|
* remap_vmalloc_range - map vmalloc pages to userspace
|
|
|
|
* @vma: vma to cover (map full range of vma)
|
|
|
|
* @addr: vmalloc memory
|
|
|
|
* @pgoff: number of pages into addr before first page to map
|
2013-07-04 06:02:18 +08:00
|
|
|
*
|
2019-03-06 07:48:36 +08:00
|
|
|
* Returns: 0 for success, -Exxx on failure
|
2013-07-04 06:02:18 +08:00
|
|
|
*
|
2019-03-06 07:48:36 +08:00
|
|
|
* This function checks that addr is a valid vmalloc'ed area, and
|
|
|
|
* that it is big enough to cover the vma. Will return failure if
|
|
|
|
* that criteria isn't met.
|
2013-07-04 06:02:18 +08:00
|
|
|
*
|
2019-03-06 07:48:36 +08:00
|
|
|
* Similar to remap_pfn_range() (see mm/memory.c)
|
2013-07-04 06:02:18 +08:00
|
|
|
*/
|
|
|
|
int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
|
|
|
|
unsigned long pgoff)
|
|
|
|
{
|
|
|
|
return remap_vmalloc_range_partial(vma, vma->vm_start,
|
|
|
|
addr + (pgoff << PAGE_SHIFT),
|
|
|
|
vma->vm_end - vma->vm_start);
|
|
|
|
}
|
2006-06-23 17:03:20 +08:00
|
|
|
EXPORT_SYMBOL(remap_vmalloc_range);
|
|
|
|
|
2007-05-08 15:27:03 +08:00
|
|
|
/*
|
|
|
|
* Implement a stub for vmalloc_sync_all() if the architecture chose not to
|
|
|
|
* have one.
|
2019-07-20 02:46:52 +08:00
|
|
|
*
|
|
|
|
* The purpose of this function is to make sure the vmalloc area
|
|
|
|
* mappings are identical in all page-tables in the system.
|
2007-05-08 15:27:03 +08:00
|
|
|
*/
|
2014-04-08 06:37:26 +08:00
|
|
|
void __weak vmalloc_sync_all(void)
|
2007-05-08 15:27:03 +08:00
|
|
|
{
|
|
|
|
}
|
2007-07-18 09:37:04 +08:00
|
|
|
|
|
|
|
|
2019-07-12 11:58:43 +08:00
|
|
|
static int f(pte_t *pte, unsigned long addr, void *data)
|
2007-07-18 09:37:04 +08:00
|
|
|
{
|
2011-09-29 23:53:32 +08:00
|
|
|
pte_t ***p = data;
|
|
|
|
|
|
|
|
if (p) {
|
|
|
|
*(*p) = pte;
|
|
|
|
(*p)++;
|
|
|
|
}
|
2007-07-18 09:37:04 +08:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
2019-03-06 07:48:36 +08:00
|
|
|
* alloc_vm_area - allocate a range of kernel address space
|
|
|
|
* @size: size of the area
|
|
|
|
* @ptes: returns the PTEs for the address space
|
2008-03-20 08:00:40 +08:00
|
|
|
*
|
2019-03-06 07:48:36 +08:00
|
|
|
* Returns: NULL on failure, vm_struct on success
|
2007-07-18 09:37:04 +08:00
|
|
|
*
|
2019-03-06 07:48:36 +08:00
|
|
|
* This function reserves a range of kernel address space, and
|
|
|
|
* allocates pagetables to map that range. No actual mappings
|
|
|
|
* are created.
|
2011-09-29 23:53:32 +08:00
|
|
|
*
|
2019-03-06 07:48:36 +08:00
|
|
|
* If @ptes is non-NULL, pointers to the PTEs (in init_mm)
|
|
|
|
* allocated for the VM area are returned.
|
2007-07-18 09:37:04 +08:00
|
|
|
*/
|
2011-09-29 23:53:32 +08:00
|
|
|
struct vm_struct *alloc_vm_area(size_t size, pte_t **ptes)
|
2007-07-18 09:37:04 +08:00
|
|
|
{
|
|
|
|
struct vm_struct *area;
|
|
|
|
|
2008-04-28 17:12:42 +08:00
|
|
|
area = get_vm_area_caller(size, VM_IOREMAP,
|
|
|
|
__builtin_return_address(0));
|
2007-07-18 09:37:04 +08:00
|
|
|
if (area == NULL)
|
|
|
|
return NULL;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* This ensures that page tables are constructed for this region
|
|
|
|
* of kernel virtual address space and mapped into init_mm.
|
|
|
|
*/
|
|
|
|
if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
|
2011-09-29 23:53:32 +08:00
|
|
|
size, f, ptes ? &ptes : NULL)) {
|
2007-07-18 09:37:04 +08:00
|
|
|
free_vm_area(area);
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
return area;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(alloc_vm_area);
|
|
|
|
|
|
|
|
void free_vm_area(struct vm_struct *area)
|
|
|
|
{
|
|
|
|
struct vm_struct *ret;
|
|
|
|
ret = remove_vm_area(area->addr);
|
|
|
|
BUG_ON(ret != area);
|
|
|
|
kfree(area);
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(free_vm_area);
|
2008-04-28 17:12:40 +08:00
|
|
|
|
2010-09-04 00:22:47 +08:00
|
|
|
#ifdef CONFIG_SMP
|
2009-08-14 14:00:52 +08:00
|
|
|
static struct vmap_area *node_to_va(struct rb_node *n)
|
|
|
|
{
|
2017-02-23 07:41:54 +08:00
|
|
|
return rb_entry_safe(n, struct vmap_area, rb_node);
|
2009-08-14 14:00:52 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
* pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
|
|
|
|
* @addr: target address
|
2009-08-14 14:00:52 +08:00
|
|
|
*
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
* Returns: vmap_area if it is found. If there is no such area
|
|
|
|
* the first highest(reverse order) vmap_area is returned
|
|
|
|
* i.e. va->va_start < addr && va->va_end < addr or NULL
|
|
|
|
* if there are no any areas before @addr.
|
2009-08-14 14:00:52 +08:00
|
|
|
*/
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
static struct vmap_area *
|
|
|
|
pvm_find_va_enclose_addr(unsigned long addr)
|
2009-08-14 14:00:52 +08:00
|
|
|
{
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
struct vmap_area *va, *tmp;
|
|
|
|
struct rb_node *n;
|
|
|
|
|
|
|
|
n = free_vmap_area_root.rb_node;
|
|
|
|
va = NULL;
|
2009-08-14 14:00:52 +08:00
|
|
|
|
|
|
|
while (n) {
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
tmp = rb_entry(n, struct vmap_area, rb_node);
|
|
|
|
if (tmp->va_start <= addr) {
|
|
|
|
va = tmp;
|
|
|
|
if (tmp->va_end >= addr)
|
|
|
|
break;
|
|
|
|
|
2009-08-14 14:00:52 +08:00
|
|
|
n = n->rb_right;
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
} else {
|
|
|
|
n = n->rb_left;
|
|
|
|
}
|
2009-08-14 14:00:52 +08:00
|
|
|
}
|
|
|
|
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
return va;
|
2009-08-14 14:00:52 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
* pvm_determine_end_from_reverse - find the highest aligned address
|
|
|
|
* of free block below VMALLOC_END
|
|
|
|
* @va:
|
|
|
|
* in - the VA we start the search(reverse order);
|
|
|
|
* out - the VA with the highest aligned end address.
|
2009-08-14 14:00:52 +08:00
|
|
|
*
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
* Returns: determined end address within vmap_area
|
2009-08-14 14:00:52 +08:00
|
|
|
*/
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
static unsigned long
|
|
|
|
pvm_determine_end_from_reverse(struct vmap_area **va, unsigned long align)
|
2009-08-14 14:00:52 +08:00
|
|
|
{
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
|
2009-08-14 14:00:52 +08:00
|
|
|
unsigned long addr;
|
|
|
|
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
if (likely(*va)) {
|
|
|
|
list_for_each_entry_from_reverse((*va),
|
|
|
|
&free_vmap_area_list, list) {
|
|
|
|
addr = min((*va)->va_end & ~(align - 1), vmalloc_end);
|
|
|
|
if ((*va)->va_start < addr)
|
|
|
|
return addr;
|
|
|
|
}
|
2009-08-14 14:00:52 +08:00
|
|
|
}
|
|
|
|
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
return 0;
|
2009-08-14 14:00:52 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
|
|
|
|
* @offsets: array containing offset of each area
|
|
|
|
* @sizes: array containing size of each area
|
|
|
|
* @nr_vms: the number of areas to allocate
|
|
|
|
* @align: alignment, all entries in @offsets and @sizes must be aligned to this
|
|
|
|
*
|
|
|
|
* Returns: kmalloc'd vm_struct pointer array pointing to allocated
|
|
|
|
* vm_structs on success, %NULL on failure
|
|
|
|
*
|
|
|
|
* Percpu allocator wants to use congruent vm areas so that it can
|
|
|
|
* maintain the offsets among percpu areas. This function allocates
|
2011-01-14 07:46:01 +08:00
|
|
|
* congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
|
|
|
|
* be scattered pretty far, distance between two areas easily going up
|
|
|
|
* to gigabytes. To avoid interacting with regular vmallocs, these
|
|
|
|
* areas are allocated from top.
|
2009-08-14 14:00:52 +08:00
|
|
|
*
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
* Despite its complicated look, this allocator is rather simple. It
|
|
|
|
* does everything top-down and scans free blocks from the end looking
|
|
|
|
* for matching base. While scanning, if any of the areas do not fit the
|
|
|
|
* base address is pulled down to fit the area. Scanning is repeated till
|
|
|
|
* all the areas fit and then all necessary data structures are inserted
|
|
|
|
* and the result is returned.
|
2009-08-14 14:00:52 +08:00
|
|
|
*/
|
|
|
|
struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
|
|
|
|
const size_t *sizes, int nr_vms,
|
2011-01-14 07:46:01 +08:00
|
|
|
size_t align)
|
2009-08-14 14:00:52 +08:00
|
|
|
{
|
|
|
|
const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
|
|
|
|
const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
struct vmap_area **vas, *va;
|
2009-08-14 14:00:52 +08:00
|
|
|
struct vm_struct **vms;
|
|
|
|
int area, area2, last_area, term_area;
|
2019-12-18 12:51:49 +08:00
|
|
|
unsigned long base, start, size, end, last_end, orig_start, orig_end;
|
2009-08-14 14:00:52 +08:00
|
|
|
bool purged = false;
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
enum fit_type type;
|
2009-08-14 14:00:52 +08:00
|
|
|
|
|
|
|
/* verify parameters and allocate data structures */
|
2015-11-06 10:46:51 +08:00
|
|
|
BUG_ON(offset_in_page(align) || !is_power_of_2(align));
|
2009-08-14 14:00:52 +08:00
|
|
|
for (last_area = 0, area = 0; area < nr_vms; area++) {
|
|
|
|
start = offsets[area];
|
|
|
|
end = start + sizes[area];
|
|
|
|
|
|
|
|
/* is everything aligned properly? */
|
|
|
|
BUG_ON(!IS_ALIGNED(offsets[area], align));
|
|
|
|
BUG_ON(!IS_ALIGNED(sizes[area], align));
|
|
|
|
|
|
|
|
/* detect the area with the highest address */
|
|
|
|
if (start > offsets[last_area])
|
|
|
|
last_area = area;
|
|
|
|
|
2017-09-07 07:24:09 +08:00
|
|
|
for (area2 = area + 1; area2 < nr_vms; area2++) {
|
2009-08-14 14:00:52 +08:00
|
|
|
unsigned long start2 = offsets[area2];
|
|
|
|
unsigned long end2 = start2 + sizes[area2];
|
|
|
|
|
2017-09-07 07:24:09 +08:00
|
|
|
BUG_ON(start2 < end && start < end2);
|
2009-08-14 14:00:52 +08:00
|
|
|
}
|
|
|
|
}
|
|
|
|
last_end = offsets[last_area] + sizes[last_area];
|
|
|
|
|
|
|
|
if (vmalloc_end - vmalloc_start < last_end) {
|
|
|
|
WARN_ON(true);
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
2012-05-30 06:06:21 +08:00
|
|
|
vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
|
|
|
|
vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
|
2009-08-14 14:00:52 +08:00
|
|
|
if (!vas || !vms)
|
2012-01-13 09:20:08 +08:00
|
|
|
goto err_free2;
|
2009-08-14 14:00:52 +08:00
|
|
|
|
|
|
|
for (area = 0; area < nr_vms; area++) {
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
vas[area] = kmem_cache_zalloc(vmap_area_cachep, GFP_KERNEL);
|
2011-01-14 07:46:01 +08:00
|
|
|
vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
|
2009-08-14 14:00:52 +08:00
|
|
|
if (!vas[area] || !vms[area])
|
|
|
|
goto err_free;
|
|
|
|
}
|
|
|
|
retry:
|
mm/vmalloc: rework vmap_area_lock
With the new allocation approach introduced in the 5.2 kernel, it
becomes possible to get rid of one global spinlock. By doing that we
can further improve the KVA from the performance point of view.
Basically we can have two independent locks, one for allocation part and
another one for deallocation, because of two different entities: "free
data structures" and "busy data structures".
As a result, allocation/deallocation operations can still interfere
between each other in case of running simultaneously on different CPUs,
it means there is still dependency, but with two locks it becomes lower.
Summarizing:
- it reduces the high lock contention
- it allows to perform operations on "free" and "busy"
trees in parallel on different CPUs. Please note it
does not solve scalability issue.
Test results:
In order to evaluate this patch, we can run "vmalloc test driver" to see
how many CPU cycles it takes to complete all test cases running
sequentially. All online CPUs run it so it will cause a high lock
contention.
HiKey 960, ARM64, 8xCPUs, big.LITTLE:
<snip>
sudo ./test_vmalloc.sh sequential_test_order=1
<snip>
<default>
[ 390.950557] All test took CPU0=457126382 cycles
[ 391.046690] All test took CPU1=454763452 cycles
[ 391.128586] All test took CPU2=454539334 cycles
[ 391.222669] All test took CPU3=455649517 cycles
[ 391.313946] All test took CPU4=388272196 cycles
[ 391.410425] All test took CPU5=384036264 cycles
[ 391.492219] All test took CPU6=387432964 cycles
[ 391.578433] All test took CPU7=387201996 cycles
<default>
<patched>
[ 304.721224] All test took CPU0=391521310 cycles
[ 304.821219] All test took CPU1=393533002 cycles
[ 304.917120] All test took CPU2=392243032 cycles
[ 305.008986] All test took CPU3=392353853 cycles
[ 305.108944] All test took CPU4=297630721 cycles
[ 305.196406] All test took CPU5=297548736 cycles
[ 305.288602] All test took CPU6=297092392 cycles
[ 305.381088] All test took CPU7=297293597 cycles
<patched>
~14%-23% patched variant is better.
Link: http://lkml.kernel.org/r/20191022155800.20468-1-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Acked-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:47 +08:00
|
|
|
spin_lock(&free_vmap_area_lock);
|
2009-08-14 14:00:52 +08:00
|
|
|
|
|
|
|
/* start scanning - we scan from the top, begin with the last area */
|
|
|
|
area = term_area = last_area;
|
|
|
|
start = offsets[area];
|
|
|
|
end = start + sizes[area];
|
|
|
|
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
va = pvm_find_va_enclose_addr(vmalloc_end);
|
|
|
|
base = pvm_determine_end_from_reverse(&va, align) - end;
|
2009-08-14 14:00:52 +08:00
|
|
|
|
|
|
|
while (true) {
|
|
|
|
/*
|
|
|
|
* base might have underflowed, add last_end before
|
|
|
|
* comparing.
|
|
|
|
*/
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
if (base + last_end < vmalloc_start + last_end)
|
|
|
|
goto overflow;
|
2009-08-14 14:00:52 +08:00
|
|
|
|
|
|
|
/*
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
* Fitting base has not been found.
|
2009-08-14 14:00:52 +08:00
|
|
|
*/
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
if (va == NULL)
|
|
|
|
goto overflow;
|
2009-08-14 14:00:52 +08:00
|
|
|
|
mm/vmalloc.c: fix percpu free VM area search criteria
Recent changes to the vmalloc code by commit 68ad4a330433
("mm/vmalloc.c: keep track of free blocks for vmap allocation") can
cause spurious percpu allocation failures. These, in turn, can result
in panic()s in the slub code. One such possible panic was reported by
Dave Hansen in following link https://lkml.org/lkml/2019/6/19/939.
Another related panic observed is,
RIP: 0033:0x7f46f7441b9b
Call Trace:
dump_stack+0x61/0x80
pcpu_alloc.cold.30+0x22/0x4f
mem_cgroup_css_alloc+0x110/0x650
cgroup_apply_control_enable+0x133/0x330
cgroup_mkdir+0x41b/0x500
kernfs_iop_mkdir+0x5a/0x90
vfs_mkdir+0x102/0x1b0
do_mkdirat+0x7d/0xf0
do_syscall_64+0x5b/0x180
entry_SYSCALL_64_after_hwframe+0x44/0xa9
VMALLOC memory manager divides the entire VMALLOC space (VMALLOC_START
to VMALLOC_END) into multiple VM areas (struct vm_areas), and it mainly
uses two lists (vmap_area_list & free_vmap_area_list) to track the used
and free VM areas in VMALLOC space. And pcpu_get_vm_areas(offsets[],
sizes[], nr_vms, align) function is used for allocating congruent VM
areas for percpu memory allocator. In order to not conflict with
VMALLOC users, pcpu_get_vm_areas allocates VM areas near the end of the
VMALLOC space. So the search for free vm_area for the given requirement
starts near VMALLOC_END and moves upwards towards VMALLOC_START.
Prior to commit 68ad4a330433, the search for free vm_area in
pcpu_get_vm_areas() involves following two main steps.
Step 1:
Find a aligned "base" adress near VMALLOC_END.
va = free vm area near VMALLOC_END
Step 2:
Loop through number of requested vm_areas and check,
Step 2.1:
if (base < VMALLOC_START)
1. fail with error
Step 2.2:
// end is offsets[area] + sizes[area]
if (base + end > va->vm_end)
1. Move the base downwards and repeat Step 2
Step 2.3:
if (base + start < va->vm_start)
1. Move to previous free vm_area node, find aligned
base address and repeat Step 2
But Commit 68ad4a330433 removed Step 2.2 and modified Step 2.3 as below:
Step 2.3:
if (base + start < va->vm_start || base + end > va->vm_end)
1. Move to previous free vm_area node, find aligned
base address and repeat Step 2
Above change is the root cause of spurious percpu memory allocation
failures. For example, consider a case where a relatively large vm_area
(~ 30 TB) was ignored in free vm_area search because it did not pass the
base + end < vm->vm_end boundary check. Ignoring such large free
vm_area's would lead to not finding free vm_area within boundary of
VMALLOC_start to VMALLOC_END which in turn leads to allocation failures.
So modify the search algorithm to include Step 2.2.
Link: http://lkml.kernel.org/r/20190729232139.91131-1-sathyanarayanan.kuppuswamy@linux.intel.com
Fixes: 68ad4a330433 ("mm/vmalloc.c: keep track of free blocks for vmap allocation")
Signed-off-by: Kuppuswamy Sathyanarayanan <sathyanarayanan.kuppuswamy@linux.intel.com>
Reported-by: Dave Hansen <dave.hansen@intel.com>
Acked-by: Dennis Zhou <dennis@kernel.org>
Reviewed-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: sathyanarayanan kuppuswamy <sathyanarayanan.kuppuswamy@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-08-14 06:37:31 +08:00
|
|
|
/*
|
|
|
|
* If required width exeeds current VA block, move
|
|
|
|
* base downwards and then recheck.
|
|
|
|
*/
|
|
|
|
if (base + end > va->va_end) {
|
|
|
|
base = pvm_determine_end_from_reverse(&va, align) - end;
|
|
|
|
term_area = area;
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
2009-08-14 14:00:52 +08:00
|
|
|
/*
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
* If this VA does not fit, move base downwards and recheck.
|
2009-08-14 14:00:52 +08:00
|
|
|
*/
|
mm/vmalloc.c: fix percpu free VM area search criteria
Recent changes to the vmalloc code by commit 68ad4a330433
("mm/vmalloc.c: keep track of free blocks for vmap allocation") can
cause spurious percpu allocation failures. These, in turn, can result
in panic()s in the slub code. One such possible panic was reported by
Dave Hansen in following link https://lkml.org/lkml/2019/6/19/939.
Another related panic observed is,
RIP: 0033:0x7f46f7441b9b
Call Trace:
dump_stack+0x61/0x80
pcpu_alloc.cold.30+0x22/0x4f
mem_cgroup_css_alloc+0x110/0x650
cgroup_apply_control_enable+0x133/0x330
cgroup_mkdir+0x41b/0x500
kernfs_iop_mkdir+0x5a/0x90
vfs_mkdir+0x102/0x1b0
do_mkdirat+0x7d/0xf0
do_syscall_64+0x5b/0x180
entry_SYSCALL_64_after_hwframe+0x44/0xa9
VMALLOC memory manager divides the entire VMALLOC space (VMALLOC_START
to VMALLOC_END) into multiple VM areas (struct vm_areas), and it mainly
uses two lists (vmap_area_list & free_vmap_area_list) to track the used
and free VM areas in VMALLOC space. And pcpu_get_vm_areas(offsets[],
sizes[], nr_vms, align) function is used for allocating congruent VM
areas for percpu memory allocator. In order to not conflict with
VMALLOC users, pcpu_get_vm_areas allocates VM areas near the end of the
VMALLOC space. So the search for free vm_area for the given requirement
starts near VMALLOC_END and moves upwards towards VMALLOC_START.
Prior to commit 68ad4a330433, the search for free vm_area in
pcpu_get_vm_areas() involves following two main steps.
Step 1:
Find a aligned "base" adress near VMALLOC_END.
va = free vm area near VMALLOC_END
Step 2:
Loop through number of requested vm_areas and check,
Step 2.1:
if (base < VMALLOC_START)
1. fail with error
Step 2.2:
// end is offsets[area] + sizes[area]
if (base + end > va->vm_end)
1. Move the base downwards and repeat Step 2
Step 2.3:
if (base + start < va->vm_start)
1. Move to previous free vm_area node, find aligned
base address and repeat Step 2
But Commit 68ad4a330433 removed Step 2.2 and modified Step 2.3 as below:
Step 2.3:
if (base + start < va->vm_start || base + end > va->vm_end)
1. Move to previous free vm_area node, find aligned
base address and repeat Step 2
Above change is the root cause of spurious percpu memory allocation
failures. For example, consider a case where a relatively large vm_area
(~ 30 TB) was ignored in free vm_area search because it did not pass the
base + end < vm->vm_end boundary check. Ignoring such large free
vm_area's would lead to not finding free vm_area within boundary of
VMALLOC_start to VMALLOC_END which in turn leads to allocation failures.
So modify the search algorithm to include Step 2.2.
Link: http://lkml.kernel.org/r/20190729232139.91131-1-sathyanarayanan.kuppuswamy@linux.intel.com
Fixes: 68ad4a330433 ("mm/vmalloc.c: keep track of free blocks for vmap allocation")
Signed-off-by: Kuppuswamy Sathyanarayanan <sathyanarayanan.kuppuswamy@linux.intel.com>
Reported-by: Dave Hansen <dave.hansen@intel.com>
Acked-by: Dennis Zhou <dennis@kernel.org>
Reviewed-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: sathyanarayanan kuppuswamy <sathyanarayanan.kuppuswamy@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-08-14 06:37:31 +08:00
|
|
|
if (base + start < va->va_start) {
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
va = node_to_va(rb_prev(&va->rb_node));
|
|
|
|
base = pvm_determine_end_from_reverse(&va, align) - end;
|
2009-08-14 14:00:52 +08:00
|
|
|
term_area = area;
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* This area fits, move on to the previous one. If
|
|
|
|
* the previous one is the terminal one, we're done.
|
|
|
|
*/
|
|
|
|
area = (area + nr_vms - 1) % nr_vms;
|
|
|
|
if (area == term_area)
|
|
|
|
break;
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
|
2009-08-14 14:00:52 +08:00
|
|
|
start = offsets[area];
|
|
|
|
end = start + sizes[area];
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
va = pvm_find_va_enclose_addr(base + end);
|
2009-08-14 14:00:52 +08:00
|
|
|
}
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
|
2009-08-14 14:00:52 +08:00
|
|
|
/* we've found a fitting base, insert all va's */
|
|
|
|
for (area = 0; area < nr_vms; area++) {
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
int ret;
|
2009-08-14 14:00:52 +08:00
|
|
|
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
start = base + offsets[area];
|
|
|
|
size = sizes[area];
|
2009-08-14 14:00:52 +08:00
|
|
|
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
va = pvm_find_va_enclose_addr(start);
|
|
|
|
if (WARN_ON_ONCE(va == NULL))
|
|
|
|
/* It is a BUG(), but trigger recovery instead. */
|
|
|
|
goto recovery;
|
|
|
|
|
|
|
|
type = classify_va_fit_type(va, start, size);
|
|
|
|
if (WARN_ON_ONCE(type == NOTHING_FIT))
|
|
|
|
/* It is a BUG(), but trigger recovery instead. */
|
|
|
|
goto recovery;
|
|
|
|
|
|
|
|
ret = adjust_va_to_fit_type(va, start, size, type);
|
|
|
|
if (unlikely(ret))
|
|
|
|
goto recovery;
|
|
|
|
|
|
|
|
/* Allocated area. */
|
|
|
|
va = vas[area];
|
|
|
|
va->va_start = start;
|
|
|
|
va->va_end = start + size;
|
|
|
|
}
|
2009-08-14 14:00:52 +08:00
|
|
|
|
mm/vmalloc: rework vmap_area_lock
With the new allocation approach introduced in the 5.2 kernel, it
becomes possible to get rid of one global spinlock. By doing that we
can further improve the KVA from the performance point of view.
Basically we can have two independent locks, one for allocation part and
another one for deallocation, because of two different entities: "free
data structures" and "busy data structures".
As a result, allocation/deallocation operations can still interfere
between each other in case of running simultaneously on different CPUs,
it means there is still dependency, but with two locks it becomes lower.
Summarizing:
- it reduces the high lock contention
- it allows to perform operations on "free" and "busy"
trees in parallel on different CPUs. Please note it
does not solve scalability issue.
Test results:
In order to evaluate this patch, we can run "vmalloc test driver" to see
how many CPU cycles it takes to complete all test cases running
sequentially. All online CPUs run it so it will cause a high lock
contention.
HiKey 960, ARM64, 8xCPUs, big.LITTLE:
<snip>
sudo ./test_vmalloc.sh sequential_test_order=1
<snip>
<default>
[ 390.950557] All test took CPU0=457126382 cycles
[ 391.046690] All test took CPU1=454763452 cycles
[ 391.128586] All test took CPU2=454539334 cycles
[ 391.222669] All test took CPU3=455649517 cycles
[ 391.313946] All test took CPU4=388272196 cycles
[ 391.410425] All test took CPU5=384036264 cycles
[ 391.492219] All test took CPU6=387432964 cycles
[ 391.578433] All test took CPU7=387201996 cycles
<default>
<patched>
[ 304.721224] All test took CPU0=391521310 cycles
[ 304.821219] All test took CPU1=393533002 cycles
[ 304.917120] All test took CPU2=392243032 cycles
[ 305.008986] All test took CPU3=392353853 cycles
[ 305.108944] All test took CPU4=297630721 cycles
[ 305.196406] All test took CPU5=297548736 cycles
[ 305.288602] All test took CPU6=297092392 cycles
[ 305.381088] All test took CPU7=297293597 cycles
<patched>
~14%-23% patched variant is better.
Link: http://lkml.kernel.org/r/20191022155800.20468-1-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Acked-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:47 +08:00
|
|
|
spin_unlock(&free_vmap_area_lock);
|
2009-08-14 14:00:52 +08:00
|
|
|
|
2019-12-18 12:51:49 +08:00
|
|
|
/* populate the kasan shadow space */
|
|
|
|
for (area = 0; area < nr_vms; area++) {
|
|
|
|
if (kasan_populate_vmalloc(vas[area]->va_start, sizes[area]))
|
|
|
|
goto err_free_shadow;
|
|
|
|
|
|
|
|
kasan_unpoison_vmalloc((void *)vas[area]->va_start,
|
|
|
|
sizes[area]);
|
|
|
|
}
|
|
|
|
|
2009-08-14 14:00:52 +08:00
|
|
|
/* insert all vm's */
|
mm/vmalloc: rework vmap_area_lock
With the new allocation approach introduced in the 5.2 kernel, it
becomes possible to get rid of one global spinlock. By doing that we
can further improve the KVA from the performance point of view.
Basically we can have two independent locks, one for allocation part and
another one for deallocation, because of two different entities: "free
data structures" and "busy data structures".
As a result, allocation/deallocation operations can still interfere
between each other in case of running simultaneously on different CPUs,
it means there is still dependency, but with two locks it becomes lower.
Summarizing:
- it reduces the high lock contention
- it allows to perform operations on "free" and "busy"
trees in parallel on different CPUs. Please note it
does not solve scalability issue.
Test results:
In order to evaluate this patch, we can run "vmalloc test driver" to see
how many CPU cycles it takes to complete all test cases running
sequentially. All online CPUs run it so it will cause a high lock
contention.
HiKey 960, ARM64, 8xCPUs, big.LITTLE:
<snip>
sudo ./test_vmalloc.sh sequential_test_order=1
<snip>
<default>
[ 390.950557] All test took CPU0=457126382 cycles
[ 391.046690] All test took CPU1=454763452 cycles
[ 391.128586] All test took CPU2=454539334 cycles
[ 391.222669] All test took CPU3=455649517 cycles
[ 391.313946] All test took CPU4=388272196 cycles
[ 391.410425] All test took CPU5=384036264 cycles
[ 391.492219] All test took CPU6=387432964 cycles
[ 391.578433] All test took CPU7=387201996 cycles
<default>
<patched>
[ 304.721224] All test took CPU0=391521310 cycles
[ 304.821219] All test took CPU1=393533002 cycles
[ 304.917120] All test took CPU2=392243032 cycles
[ 305.008986] All test took CPU3=392353853 cycles
[ 305.108944] All test took CPU4=297630721 cycles
[ 305.196406] All test took CPU5=297548736 cycles
[ 305.288602] All test took CPU6=297092392 cycles
[ 305.381088] All test took CPU7=297293597 cycles
<patched>
~14%-23% patched variant is better.
Link: http://lkml.kernel.org/r/20191022155800.20468-1-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Acked-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:47 +08:00
|
|
|
spin_lock(&vmap_area_lock);
|
|
|
|
for (area = 0; area < nr_vms; area++) {
|
|
|
|
insert_vmap_area(vas[area], &vmap_area_root, &vmap_area_list);
|
|
|
|
|
|
|
|
setup_vmalloc_vm_locked(vms[area], vas[area], VM_ALLOC,
|
2013-07-04 06:04:48 +08:00
|
|
|
pcpu_get_vm_areas);
|
mm/vmalloc: rework vmap_area_lock
With the new allocation approach introduced in the 5.2 kernel, it
becomes possible to get rid of one global spinlock. By doing that we
can further improve the KVA from the performance point of view.
Basically we can have two independent locks, one for allocation part and
another one for deallocation, because of two different entities: "free
data structures" and "busy data structures".
As a result, allocation/deallocation operations can still interfere
between each other in case of running simultaneously on different CPUs,
it means there is still dependency, but with two locks it becomes lower.
Summarizing:
- it reduces the high lock contention
- it allows to perform operations on "free" and "busy"
trees in parallel on different CPUs. Please note it
does not solve scalability issue.
Test results:
In order to evaluate this patch, we can run "vmalloc test driver" to see
how many CPU cycles it takes to complete all test cases running
sequentially. All online CPUs run it so it will cause a high lock
contention.
HiKey 960, ARM64, 8xCPUs, big.LITTLE:
<snip>
sudo ./test_vmalloc.sh sequential_test_order=1
<snip>
<default>
[ 390.950557] All test took CPU0=457126382 cycles
[ 391.046690] All test took CPU1=454763452 cycles
[ 391.128586] All test took CPU2=454539334 cycles
[ 391.222669] All test took CPU3=455649517 cycles
[ 391.313946] All test took CPU4=388272196 cycles
[ 391.410425] All test took CPU5=384036264 cycles
[ 391.492219] All test took CPU6=387432964 cycles
[ 391.578433] All test took CPU7=387201996 cycles
<default>
<patched>
[ 304.721224] All test took CPU0=391521310 cycles
[ 304.821219] All test took CPU1=393533002 cycles
[ 304.917120] All test took CPU2=392243032 cycles
[ 305.008986] All test took CPU3=392353853 cycles
[ 305.108944] All test took CPU4=297630721 cycles
[ 305.196406] All test took CPU5=297548736 cycles
[ 305.288602] All test took CPU6=297092392 cycles
[ 305.381088] All test took CPU7=297293597 cycles
<patched>
~14%-23% patched variant is better.
Link: http://lkml.kernel.org/r/20191022155800.20468-1-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Acked-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:47 +08:00
|
|
|
}
|
|
|
|
spin_unlock(&vmap_area_lock);
|
2009-08-14 14:00:52 +08:00
|
|
|
|
|
|
|
kfree(vas);
|
|
|
|
return vms;
|
|
|
|
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
recovery:
|
mm/vmalloc: rework vmap_area_lock
With the new allocation approach introduced in the 5.2 kernel, it
becomes possible to get rid of one global spinlock. By doing that we
can further improve the KVA from the performance point of view.
Basically we can have two independent locks, one for allocation part and
another one for deallocation, because of two different entities: "free
data structures" and "busy data structures".
As a result, allocation/deallocation operations can still interfere
between each other in case of running simultaneously on different CPUs,
it means there is still dependency, but with two locks it becomes lower.
Summarizing:
- it reduces the high lock contention
- it allows to perform operations on "free" and "busy"
trees in parallel on different CPUs. Please note it
does not solve scalability issue.
Test results:
In order to evaluate this patch, we can run "vmalloc test driver" to see
how many CPU cycles it takes to complete all test cases running
sequentially. All online CPUs run it so it will cause a high lock
contention.
HiKey 960, ARM64, 8xCPUs, big.LITTLE:
<snip>
sudo ./test_vmalloc.sh sequential_test_order=1
<snip>
<default>
[ 390.950557] All test took CPU0=457126382 cycles
[ 391.046690] All test took CPU1=454763452 cycles
[ 391.128586] All test took CPU2=454539334 cycles
[ 391.222669] All test took CPU3=455649517 cycles
[ 391.313946] All test took CPU4=388272196 cycles
[ 391.410425] All test took CPU5=384036264 cycles
[ 391.492219] All test took CPU6=387432964 cycles
[ 391.578433] All test took CPU7=387201996 cycles
<default>
<patched>
[ 304.721224] All test took CPU0=391521310 cycles
[ 304.821219] All test took CPU1=393533002 cycles
[ 304.917120] All test took CPU2=392243032 cycles
[ 305.008986] All test took CPU3=392353853 cycles
[ 305.108944] All test took CPU4=297630721 cycles
[ 305.196406] All test took CPU5=297548736 cycles
[ 305.288602] All test took CPU6=297092392 cycles
[ 305.381088] All test took CPU7=297293597 cycles
<patched>
~14%-23% patched variant is better.
Link: http://lkml.kernel.org/r/20191022155800.20468-1-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Acked-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:47 +08:00
|
|
|
/*
|
|
|
|
* Remove previously allocated areas. There is no
|
|
|
|
* need in removing these areas from the busy tree,
|
|
|
|
* because they are inserted only on the final step
|
|
|
|
* and when pcpu_get_vm_areas() is success.
|
|
|
|
*/
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
while (area--) {
|
2019-12-18 12:51:49 +08:00
|
|
|
orig_start = vas[area]->va_start;
|
|
|
|
orig_end = vas[area]->va_end;
|
|
|
|
va = merge_or_add_vmap_area(vas[area], &free_vmap_area_root,
|
|
|
|
&free_vmap_area_list);
|
|
|
|
kasan_release_vmalloc(orig_start, orig_end,
|
|
|
|
va->va_start, va->va_end);
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
vas[area] = NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
overflow:
|
mm/vmalloc: rework vmap_area_lock
With the new allocation approach introduced in the 5.2 kernel, it
becomes possible to get rid of one global spinlock. By doing that we
can further improve the KVA from the performance point of view.
Basically we can have two independent locks, one for allocation part and
another one for deallocation, because of two different entities: "free
data structures" and "busy data structures".
As a result, allocation/deallocation operations can still interfere
between each other in case of running simultaneously on different CPUs,
it means there is still dependency, but with two locks it becomes lower.
Summarizing:
- it reduces the high lock contention
- it allows to perform operations on "free" and "busy"
trees in parallel on different CPUs. Please note it
does not solve scalability issue.
Test results:
In order to evaluate this patch, we can run "vmalloc test driver" to see
how many CPU cycles it takes to complete all test cases running
sequentially. All online CPUs run it so it will cause a high lock
contention.
HiKey 960, ARM64, 8xCPUs, big.LITTLE:
<snip>
sudo ./test_vmalloc.sh sequential_test_order=1
<snip>
<default>
[ 390.950557] All test took CPU0=457126382 cycles
[ 391.046690] All test took CPU1=454763452 cycles
[ 391.128586] All test took CPU2=454539334 cycles
[ 391.222669] All test took CPU3=455649517 cycles
[ 391.313946] All test took CPU4=388272196 cycles
[ 391.410425] All test took CPU5=384036264 cycles
[ 391.492219] All test took CPU6=387432964 cycles
[ 391.578433] All test took CPU7=387201996 cycles
<default>
<patched>
[ 304.721224] All test took CPU0=391521310 cycles
[ 304.821219] All test took CPU1=393533002 cycles
[ 304.917120] All test took CPU2=392243032 cycles
[ 305.008986] All test took CPU3=392353853 cycles
[ 305.108944] All test took CPU4=297630721 cycles
[ 305.196406] All test took CPU5=297548736 cycles
[ 305.288602] All test took CPU6=297092392 cycles
[ 305.381088] All test took CPU7=297293597 cycles
<patched>
~14%-23% patched variant is better.
Link: http://lkml.kernel.org/r/20191022155800.20468-1-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Acked-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:47 +08:00
|
|
|
spin_unlock(&free_vmap_area_lock);
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
if (!purged) {
|
|
|
|
purge_vmap_area_lazy();
|
|
|
|
purged = true;
|
|
|
|
|
|
|
|
/* Before "retry", check if we recover. */
|
|
|
|
for (area = 0; area < nr_vms; area++) {
|
|
|
|
if (vas[area])
|
|
|
|
continue;
|
|
|
|
|
|
|
|
vas[area] = kmem_cache_zalloc(
|
|
|
|
vmap_area_cachep, GFP_KERNEL);
|
|
|
|
if (!vas[area])
|
|
|
|
goto err_free;
|
|
|
|
}
|
|
|
|
|
|
|
|
goto retry;
|
|
|
|
}
|
|
|
|
|
2009-08-14 14:00:52 +08:00
|
|
|
err_free:
|
|
|
|
for (area = 0; area < nr_vms; area++) {
|
mm/vmalloc.c: keep track of free blocks for vmap allocation
Patch series "improve vmap allocation", v3.
Objective
---------
Please have a look for the description at:
https://lkml.org/lkml/2018/10/19/786
but let me also summarize it a bit here as well.
The current implementation has O(N) complexity. Requests with different
permissive parameters can lead to long allocation time. When i say
"long" i mean milliseconds.
Description
-----------
This approach organizes the KVA memory layout into free areas of the
1-ULONG_MAX range, i.e. an allocation is done over free areas lookups,
instead of finding a hole between two busy blocks. It allows to have
lower number of objects which represent the free space, therefore to have
less fragmented memory allocator. Because free blocks are always as large
as possible.
It uses the augment tree where all free areas are sorted in ascending
order of va->va_start address in pair with linked list that provides
O(1) access to prev/next elements.
Since the tree is augment, we also maintain the "subtree_max_size" of VA
that reflects a maximum available free block in its left or right
sub-tree. Knowing that, we can easily traversal toward the lowest (left
most path) free area.
Allocation: ~O(log(N)) complexity. It is sequential allocation method
therefore tends to maximize locality. The search is done until a first
suitable block is large enough to encompass the requested parameters.
Bigger areas are split.
I copy paste here the description of how the area is split, since i
described it in https://lkml.org/lkml/2018/10/19/786
<snip>
A free block can be split by three different ways. Their names are
FL_FIT_TYPE, LE_FIT_TYPE/RE_FIT_TYPE and NE_FIT_TYPE, i.e. they
correspond to how requested size and alignment fit to a free block.
FL_FIT_TYPE - in this case a free block is just removed from the free
list/tree because it fully fits. Comparing with current design there is
an extra work with rb-tree updating.
LE_FIT_TYPE/RE_FIT_TYPE - left/right edges fit. In this case what we do
is just cutting a free block. It is as fast as a current design. Most of
the vmalloc allocations just end up with this case, because the edge is
always aligned to 1.
NE_FIT_TYPE - Is much less common case. Basically it happens when
requested size and alignment does not fit left nor right edges, i.e. it
is between them. In this case during splitting we have to build a
remaining left free area and place it back to the free list/tree.
Comparing with current design there are two extra steps. First one is we
have to allocate a new vmap_area structure. Second one we have to insert
that remaining free block to the address sorted list/tree.
In order to optimize a first case there is a cache with free_vmap objects.
Instead of allocating from slab we just take an object from the cache and
reuse it.
Second one is pretty optimized. Since we know a start point in the tree
we do not do a search from the top. Instead a traversal begins from a
rb-tree node we split.
<snip>
De-allocation. ~O(log(N)) complexity. An area is not inserted straight
away to the tree/list, instead we identify the spot first, checking if it
can be merged around neighbors. The list provides O(1) access to
prev/next, so it is pretty fast to check it. Summarizing. If merged then
large coalesced areas are created, if not the area is just linked making
more fragments.
There is one more thing that i should mention here. After modification of
VA node, its subtree_max_size is updated if it was/is the biggest area in
its left or right sub-tree. Apart of that it can also be populated back
to upper levels to fix the tree. For more details please have a look at
the __augment_tree_propagate_from() function and the description.
Tests and stressing
-------------------
I use the "test_vmalloc.sh" test driver available under
"tools/testing/selftests/vm/" since 5.1-rc1 kernel. Just trigger "sudo
./test_vmalloc.sh" to find out how to deal with it.
Tested on different platforms including x86_64/i686/ARM64/x86_64_NUMA.
Regarding last one, i do not have any physical access to NUMA system,
therefore i emulated it. The time of stressing is days.
If you run the test driver in "stress mode", you also need the patch that
is in Andrew's tree but not in Linux 5.1-rc1. So, please apply it:
http://git.cmpxchg.org/cgit.cgi/linux-mmotm.git/commit/?id=e0cf7749bade6da318e98e934a24d8b62fab512c
After massive testing, i have not identified any problems like memory
leaks, crashes or kernel panics. I find it stable, but more testing would
be good.
Performance analysis
--------------------
I have used two systems to test. One is i5-3320M CPU @ 2.60GHz and
another is HiKey960(arm64) board. i5-3320M runs on 4.20 kernel, whereas
Hikey960 uses 4.15 kernel. I have both system which could run on 5.1-rc1
as well, but the results have not been ready by time i an writing this.
Currently it consist of 8 tests. There are three of them which correspond
to different types of splitting(to compare with default). We have 3
ones(see above). Another 5 do allocations in different conditions.
a) sudo ./test_vmalloc.sh performance
When the test driver is run in "performance" mode, it runs all available
tests pinned to first online CPU with sequential execution test order. We
do it in order to get stable and repeatable results. Take a look at time
difference in "long_busy_list_alloc_test". It is not surprising because
the worst case is O(N).
# i5-3320M
How many cycles all tests took:
CPU0=646919905370(default) cycles vs CPU0=193290498550(patched) cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_performance_patched.txt
# Hikey960 8x CPUs
How many cycles all tests took:
CPU0=3478683207 cycles vs CPU0=463767978 cycles
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/HiKey960_performance_patched.txt
b) time sudo ./test_vmalloc.sh test_repeat_count=1
With this configuration, all tests are run on all available online CPUs.
Before running each CPU shuffles its tests execution order. It gives
random allocation behaviour. So it is rough comparison, but it puts in
the picture for sure.
# i5-3320M
<default> vs <patched>
real 101m22.813s real 0m56.805s
user 0m0.011s user 0m0.015s
sys 0m5.076s sys 0m0.023s
# See detailed table with results here:
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_default.txt
ftp://vps418301.ovh.net/incoming/vmap_test_results_v2/i5-3320M_test_repeat_count_1_patched.txt
# Hikey960 8x CPUs
<default> vs <patched>
real unknown real 4m25.214s
user unknown user 0m0.011s
sys unknown sys 0m0.670s
I did not manage to complete this test on "default Hikey960" kernel
version. After 24 hours it was still running, therefore i had to cancel
it. That is why real/user/sys are "unknown".
This patch (of 3):
Currently an allocation of the new vmap area is done over busy list
iteration(complexity O(n)) until a suitable hole is found between two busy
areas. Therefore each new allocation causes the list being grown. Due to
over fragmented list and different permissive parameters an allocation can
take a long time. For example on embedded devices it is milliseconds.
This patch organizes the KVA memory layout into free areas of the
1-ULONG_MAX range. It uses an augment red-black tree that keeps blocks
sorted by their offsets in pair with linked list keeping the free space in
order of increasing addresses.
Nodes are augmented with the size of the maximum available free block in
its left or right sub-tree. Thus, that allows to take a decision and
traversal toward the block that will fit and will have the lowest start
address, i.e. it is sequential allocation.
Allocation: to allocate a new block a search is done over the tree until a
suitable lowest(left most) block is large enough to encompass: the
requested size, alignment and vstart point. If the block is bigger than
requested size - it is split.
De-allocation: when a busy vmap area is freed it can either be merged or
inserted to the tree. Red-black tree allows efficiently find a spot
whereas a linked list provides a constant-time access to previous and next
blocks to check if merging can be done. In case of merging of
de-allocated memory chunk a large coalesced area is created.
Complexity: ~O(log(N))
[urezki@gmail.com: v3]
Link: http://lkml.kernel.org/r/20190402162531.10888-2-urezki@gmail.com
[urezki@gmail.com: v4]
Link: http://lkml.kernel.org/r/20190406183508.25273-2-urezki@gmail.com
Link: http://lkml.kernel.org/r/20190321190327.11813-2-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Roman Gushchin <guro@fb.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Thomas Garnier <thgarnie@google.com>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Joel Fernandes <joelaf@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-05-18 05:31:31 +08:00
|
|
|
if (vas[area])
|
|
|
|
kmem_cache_free(vmap_area_cachep, vas[area]);
|
|
|
|
|
2012-01-13 09:20:08 +08:00
|
|
|
kfree(vms[area]);
|
2009-08-14 14:00:52 +08:00
|
|
|
}
|
2012-01-13 09:20:08 +08:00
|
|
|
err_free2:
|
2009-08-14 14:00:52 +08:00
|
|
|
kfree(vas);
|
|
|
|
kfree(vms);
|
|
|
|
return NULL;
|
2019-12-18 12:51:49 +08:00
|
|
|
|
|
|
|
err_free_shadow:
|
|
|
|
spin_lock(&free_vmap_area_lock);
|
|
|
|
/*
|
|
|
|
* We release all the vmalloc shadows, even the ones for regions that
|
|
|
|
* hadn't been successfully added. This relies on kasan_release_vmalloc
|
|
|
|
* being able to tolerate this case.
|
|
|
|
*/
|
|
|
|
for (area = 0; area < nr_vms; area++) {
|
|
|
|
orig_start = vas[area]->va_start;
|
|
|
|
orig_end = vas[area]->va_end;
|
|
|
|
va = merge_or_add_vmap_area(vas[area], &free_vmap_area_root,
|
|
|
|
&free_vmap_area_list);
|
|
|
|
kasan_release_vmalloc(orig_start, orig_end,
|
|
|
|
va->va_start, va->va_end);
|
|
|
|
vas[area] = NULL;
|
|
|
|
kfree(vms[area]);
|
|
|
|
}
|
|
|
|
spin_unlock(&free_vmap_area_lock);
|
|
|
|
kfree(vas);
|
|
|
|
kfree(vms);
|
|
|
|
return NULL;
|
2009-08-14 14:00:52 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* pcpu_free_vm_areas - free vmalloc areas for percpu allocator
|
|
|
|
* @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
|
|
|
|
* @nr_vms: the number of allocated areas
|
|
|
|
*
|
|
|
|
* Free vm_structs and the array allocated by pcpu_get_vm_areas().
|
|
|
|
*/
|
|
|
|
void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
|
|
|
|
{
|
|
|
|
int i;
|
|
|
|
|
|
|
|
for (i = 0; i < nr_vms; i++)
|
|
|
|
free_vm_area(vms[i]);
|
|
|
|
kfree(vms);
|
|
|
|
}
|
2010-09-04 00:22:47 +08:00
|
|
|
#endif /* CONFIG_SMP */
|
2008-04-28 17:12:40 +08:00
|
|
|
|
|
|
|
#ifdef CONFIG_PROC_FS
|
|
|
|
static void *s_start(struct seq_file *m, loff_t *pos)
|
mm/vmalloc: rework vmap_area_lock
With the new allocation approach introduced in the 5.2 kernel, it
becomes possible to get rid of one global spinlock. By doing that we
can further improve the KVA from the performance point of view.
Basically we can have two independent locks, one for allocation part and
another one for deallocation, because of two different entities: "free
data structures" and "busy data structures".
As a result, allocation/deallocation operations can still interfere
between each other in case of running simultaneously on different CPUs,
it means there is still dependency, but with two locks it becomes lower.
Summarizing:
- it reduces the high lock contention
- it allows to perform operations on "free" and "busy"
trees in parallel on different CPUs. Please note it
does not solve scalability issue.
Test results:
In order to evaluate this patch, we can run "vmalloc test driver" to see
how many CPU cycles it takes to complete all test cases running
sequentially. All online CPUs run it so it will cause a high lock
contention.
HiKey 960, ARM64, 8xCPUs, big.LITTLE:
<snip>
sudo ./test_vmalloc.sh sequential_test_order=1
<snip>
<default>
[ 390.950557] All test took CPU0=457126382 cycles
[ 391.046690] All test took CPU1=454763452 cycles
[ 391.128586] All test took CPU2=454539334 cycles
[ 391.222669] All test took CPU3=455649517 cycles
[ 391.313946] All test took CPU4=388272196 cycles
[ 391.410425] All test took CPU5=384036264 cycles
[ 391.492219] All test took CPU6=387432964 cycles
[ 391.578433] All test took CPU7=387201996 cycles
<default>
<patched>
[ 304.721224] All test took CPU0=391521310 cycles
[ 304.821219] All test took CPU1=393533002 cycles
[ 304.917120] All test took CPU2=392243032 cycles
[ 305.008986] All test took CPU3=392353853 cycles
[ 305.108944] All test took CPU4=297630721 cycles
[ 305.196406] All test took CPU5=297548736 cycles
[ 305.288602] All test took CPU6=297092392 cycles
[ 305.381088] All test took CPU7=297293597 cycles
<patched>
~14%-23% patched variant is better.
Link: http://lkml.kernel.org/r/20191022155800.20468-1-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Acked-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:47 +08:00
|
|
|
__acquires(&vmap_purge_lock)
|
2013-04-30 06:07:35 +08:00
|
|
|
__acquires(&vmap_area_lock)
|
2008-04-28 17:12:40 +08:00
|
|
|
{
|
mm/vmalloc: rework vmap_area_lock
With the new allocation approach introduced in the 5.2 kernel, it
becomes possible to get rid of one global spinlock. By doing that we
can further improve the KVA from the performance point of view.
Basically we can have two independent locks, one for allocation part and
another one for deallocation, because of two different entities: "free
data structures" and "busy data structures".
As a result, allocation/deallocation operations can still interfere
between each other in case of running simultaneously on different CPUs,
it means there is still dependency, but with two locks it becomes lower.
Summarizing:
- it reduces the high lock contention
- it allows to perform operations on "free" and "busy"
trees in parallel on different CPUs. Please note it
does not solve scalability issue.
Test results:
In order to evaluate this patch, we can run "vmalloc test driver" to see
how many CPU cycles it takes to complete all test cases running
sequentially. All online CPUs run it so it will cause a high lock
contention.
HiKey 960, ARM64, 8xCPUs, big.LITTLE:
<snip>
sudo ./test_vmalloc.sh sequential_test_order=1
<snip>
<default>
[ 390.950557] All test took CPU0=457126382 cycles
[ 391.046690] All test took CPU1=454763452 cycles
[ 391.128586] All test took CPU2=454539334 cycles
[ 391.222669] All test took CPU3=455649517 cycles
[ 391.313946] All test took CPU4=388272196 cycles
[ 391.410425] All test took CPU5=384036264 cycles
[ 391.492219] All test took CPU6=387432964 cycles
[ 391.578433] All test took CPU7=387201996 cycles
<default>
<patched>
[ 304.721224] All test took CPU0=391521310 cycles
[ 304.821219] All test took CPU1=393533002 cycles
[ 304.917120] All test took CPU2=392243032 cycles
[ 305.008986] All test took CPU3=392353853 cycles
[ 305.108944] All test took CPU4=297630721 cycles
[ 305.196406] All test took CPU5=297548736 cycles
[ 305.288602] All test took CPU6=297092392 cycles
[ 305.381088] All test took CPU7=297293597 cycles
<patched>
~14%-23% patched variant is better.
Link: http://lkml.kernel.org/r/20191022155800.20468-1-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Acked-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:47 +08:00
|
|
|
mutex_lock(&vmap_purge_lock);
|
2013-04-30 06:07:35 +08:00
|
|
|
spin_lock(&vmap_area_lock);
|
mm/vmalloc: rework vmap_area_lock
With the new allocation approach introduced in the 5.2 kernel, it
becomes possible to get rid of one global spinlock. By doing that we
can further improve the KVA from the performance point of view.
Basically we can have two independent locks, one for allocation part and
another one for deallocation, because of two different entities: "free
data structures" and "busy data structures".
As a result, allocation/deallocation operations can still interfere
between each other in case of running simultaneously on different CPUs,
it means there is still dependency, but with two locks it becomes lower.
Summarizing:
- it reduces the high lock contention
- it allows to perform operations on "free" and "busy"
trees in parallel on different CPUs. Please note it
does not solve scalability issue.
Test results:
In order to evaluate this patch, we can run "vmalloc test driver" to see
how many CPU cycles it takes to complete all test cases running
sequentially. All online CPUs run it so it will cause a high lock
contention.
HiKey 960, ARM64, 8xCPUs, big.LITTLE:
<snip>
sudo ./test_vmalloc.sh sequential_test_order=1
<snip>
<default>
[ 390.950557] All test took CPU0=457126382 cycles
[ 391.046690] All test took CPU1=454763452 cycles
[ 391.128586] All test took CPU2=454539334 cycles
[ 391.222669] All test took CPU3=455649517 cycles
[ 391.313946] All test took CPU4=388272196 cycles
[ 391.410425] All test took CPU5=384036264 cycles
[ 391.492219] All test took CPU6=387432964 cycles
[ 391.578433] All test took CPU7=387201996 cycles
<default>
<patched>
[ 304.721224] All test took CPU0=391521310 cycles
[ 304.821219] All test took CPU1=393533002 cycles
[ 304.917120] All test took CPU2=392243032 cycles
[ 305.008986] All test took CPU3=392353853 cycles
[ 305.108944] All test took CPU4=297630721 cycles
[ 305.196406] All test took CPU5=297548736 cycles
[ 305.288602] All test took CPU6=297092392 cycles
[ 305.381088] All test took CPU7=297293597 cycles
<patched>
~14%-23% patched variant is better.
Link: http://lkml.kernel.org/r/20191022155800.20468-1-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Acked-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:47 +08:00
|
|
|
|
2016-12-13 08:42:17 +08:00
|
|
|
return seq_list_start(&vmap_area_list, *pos);
|
2008-04-28 17:12:40 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
static void *s_next(struct seq_file *m, void *p, loff_t *pos)
|
|
|
|
{
|
2016-12-13 08:42:17 +08:00
|
|
|
return seq_list_next(p, &vmap_area_list, pos);
|
2008-04-28 17:12:40 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
static void s_stop(struct seq_file *m, void *p)
|
mm/vmalloc: rework vmap_area_lock
With the new allocation approach introduced in the 5.2 kernel, it
becomes possible to get rid of one global spinlock. By doing that we
can further improve the KVA from the performance point of view.
Basically we can have two independent locks, one for allocation part and
another one for deallocation, because of two different entities: "free
data structures" and "busy data structures".
As a result, allocation/deallocation operations can still interfere
between each other in case of running simultaneously on different CPUs,
it means there is still dependency, but with two locks it becomes lower.
Summarizing:
- it reduces the high lock contention
- it allows to perform operations on "free" and "busy"
trees in parallel on different CPUs. Please note it
does not solve scalability issue.
Test results:
In order to evaluate this patch, we can run "vmalloc test driver" to see
how many CPU cycles it takes to complete all test cases running
sequentially. All online CPUs run it so it will cause a high lock
contention.
HiKey 960, ARM64, 8xCPUs, big.LITTLE:
<snip>
sudo ./test_vmalloc.sh sequential_test_order=1
<snip>
<default>
[ 390.950557] All test took CPU0=457126382 cycles
[ 391.046690] All test took CPU1=454763452 cycles
[ 391.128586] All test took CPU2=454539334 cycles
[ 391.222669] All test took CPU3=455649517 cycles
[ 391.313946] All test took CPU4=388272196 cycles
[ 391.410425] All test took CPU5=384036264 cycles
[ 391.492219] All test took CPU6=387432964 cycles
[ 391.578433] All test took CPU7=387201996 cycles
<default>
<patched>
[ 304.721224] All test took CPU0=391521310 cycles
[ 304.821219] All test took CPU1=393533002 cycles
[ 304.917120] All test took CPU2=392243032 cycles
[ 305.008986] All test took CPU3=392353853 cycles
[ 305.108944] All test took CPU4=297630721 cycles
[ 305.196406] All test took CPU5=297548736 cycles
[ 305.288602] All test took CPU6=297092392 cycles
[ 305.381088] All test took CPU7=297293597 cycles
<patched>
~14%-23% patched variant is better.
Link: http://lkml.kernel.org/r/20191022155800.20468-1-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Acked-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:47 +08:00
|
|
|
__releases(&vmap_purge_lock)
|
2013-04-30 06:07:35 +08:00
|
|
|
__releases(&vmap_area_lock)
|
2008-04-28 17:12:40 +08:00
|
|
|
{
|
mm/vmalloc: rework vmap_area_lock
With the new allocation approach introduced in the 5.2 kernel, it
becomes possible to get rid of one global spinlock. By doing that we
can further improve the KVA from the performance point of view.
Basically we can have two independent locks, one for allocation part and
another one for deallocation, because of two different entities: "free
data structures" and "busy data structures".
As a result, allocation/deallocation operations can still interfere
between each other in case of running simultaneously on different CPUs,
it means there is still dependency, but with two locks it becomes lower.
Summarizing:
- it reduces the high lock contention
- it allows to perform operations on "free" and "busy"
trees in parallel on different CPUs. Please note it
does not solve scalability issue.
Test results:
In order to evaluate this patch, we can run "vmalloc test driver" to see
how many CPU cycles it takes to complete all test cases running
sequentially. All online CPUs run it so it will cause a high lock
contention.
HiKey 960, ARM64, 8xCPUs, big.LITTLE:
<snip>
sudo ./test_vmalloc.sh sequential_test_order=1
<snip>
<default>
[ 390.950557] All test took CPU0=457126382 cycles
[ 391.046690] All test took CPU1=454763452 cycles
[ 391.128586] All test took CPU2=454539334 cycles
[ 391.222669] All test took CPU3=455649517 cycles
[ 391.313946] All test took CPU4=388272196 cycles
[ 391.410425] All test took CPU5=384036264 cycles
[ 391.492219] All test took CPU6=387432964 cycles
[ 391.578433] All test took CPU7=387201996 cycles
<default>
<patched>
[ 304.721224] All test took CPU0=391521310 cycles
[ 304.821219] All test took CPU1=393533002 cycles
[ 304.917120] All test took CPU2=392243032 cycles
[ 305.008986] All test took CPU3=392353853 cycles
[ 305.108944] All test took CPU4=297630721 cycles
[ 305.196406] All test took CPU5=297548736 cycles
[ 305.288602] All test took CPU6=297092392 cycles
[ 305.381088] All test took CPU7=297293597 cycles
<patched>
~14%-23% patched variant is better.
Link: http://lkml.kernel.org/r/20191022155800.20468-1-urezki@gmail.com
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Acked-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-12-01 09:54:47 +08:00
|
|
|
mutex_unlock(&vmap_purge_lock);
|
2013-04-30 06:07:35 +08:00
|
|
|
spin_unlock(&vmap_area_lock);
|
2008-04-28 17:12:40 +08:00
|
|
|
}
|
|
|
|
|
2008-07-24 12:27:38 +08:00
|
|
|
static void show_numa_info(struct seq_file *m, struct vm_struct *v)
|
|
|
|
{
|
2012-12-12 08:00:29 +08:00
|
|
|
if (IS_ENABLED(CONFIG_NUMA)) {
|
2008-07-24 12:27:38 +08:00
|
|
|
unsigned int nr, *counters = m->private;
|
|
|
|
|
|
|
|
if (!counters)
|
|
|
|
return;
|
|
|
|
|
2013-11-13 07:07:32 +08:00
|
|
|
if (v->flags & VM_UNINITIALIZED)
|
|
|
|
return;
|
2014-12-13 08:56:30 +08:00
|
|
|
/* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
|
|
|
|
smp_rmb();
|
2013-11-13 07:07:32 +08:00
|
|
|
|
2008-07-24 12:27:38 +08:00
|
|
|
memset(counters, 0, nr_node_ids * sizeof(unsigned int));
|
|
|
|
|
|
|
|
for (nr = 0; nr < v->nr_pages; nr++)
|
|
|
|
counters[page_to_nid(v->pages[nr])]++;
|
|
|
|
|
|
|
|
for_each_node_state(nr, N_HIGH_MEMORY)
|
|
|
|
if (counters[nr])
|
|
|
|
seq_printf(m, " N%u=%u", nr, counters[nr]);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2019-09-24 06:36:36 +08:00
|
|
|
static void show_purge_info(struct seq_file *m)
|
|
|
|
{
|
|
|
|
struct llist_node *head;
|
|
|
|
struct vmap_area *va;
|
|
|
|
|
|
|
|
head = READ_ONCE(vmap_purge_list.first);
|
|
|
|
if (head == NULL)
|
|
|
|
return;
|
|
|
|
|
|
|
|
llist_for_each_entry(va, head, purge_list) {
|
|
|
|
seq_printf(m, "0x%pK-0x%pK %7ld unpurged vm_area\n",
|
|
|
|
(void *)va->va_start, (void *)va->va_end,
|
|
|
|
va->va_end - va->va_start);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2008-04-28 17:12:40 +08:00
|
|
|
static int s_show(struct seq_file *m, void *p)
|
|
|
|
{
|
2016-12-13 08:42:17 +08:00
|
|
|
struct vmap_area *va;
|
2013-04-30 06:07:35 +08:00
|
|
|
struct vm_struct *v;
|
|
|
|
|
2016-12-13 08:42:17 +08:00
|
|
|
va = list_entry(p, struct vmap_area, list);
|
|
|
|
|
2013-11-13 07:07:31 +08:00
|
|
|
/*
|
2019-09-24 06:36:39 +08:00
|
|
|
* s_show can encounter race with remove_vm_area, !vm on behalf
|
|
|
|
* of vmap area is being tear down or vm_map_ram allocation.
|
2013-11-13 07:07:31 +08:00
|
|
|
*/
|
2019-09-24 06:36:39 +08:00
|
|
|
if (!va->vm) {
|
2019-09-24 06:36:36 +08:00
|
|
|
seq_printf(m, "0x%pK-0x%pK %7ld vm_map_ram\n",
|
2017-07-11 06:48:09 +08:00
|
|
|
(void *)va->va_start, (void *)va->va_end,
|
2019-09-24 06:36:36 +08:00
|
|
|
va->va_end - va->va_start);
|
2017-07-11 06:48:09 +08:00
|
|
|
|
2013-04-30 06:07:35 +08:00
|
|
|
return 0;
|
2017-07-11 06:48:09 +08:00
|
|
|
}
|
2013-04-30 06:07:35 +08:00
|
|
|
|
|
|
|
v = va->vm;
|
2008-04-28 17:12:40 +08:00
|
|
|
|
2012-10-09 07:34:09 +08:00
|
|
|
seq_printf(m, "0x%pK-0x%pK %7ld",
|
2008-04-28 17:12:40 +08:00
|
|
|
v->addr, v->addr + v->size, v->size);
|
|
|
|
|
2011-01-14 07:45:52 +08:00
|
|
|
if (v->caller)
|
|
|
|
seq_printf(m, " %pS", v->caller);
|
2008-04-28 17:12:42 +08:00
|
|
|
|
2008-04-28 17:12:40 +08:00
|
|
|
if (v->nr_pages)
|
|
|
|
seq_printf(m, " pages=%d", v->nr_pages);
|
|
|
|
|
|
|
|
if (v->phys_addr)
|
2017-02-25 06:59:51 +08:00
|
|
|
seq_printf(m, " phys=%pa", &v->phys_addr);
|
2008-04-28 17:12:40 +08:00
|
|
|
|
|
|
|
if (v->flags & VM_IOREMAP)
|
2014-06-05 07:08:09 +08:00
|
|
|
seq_puts(m, " ioremap");
|
2008-04-28 17:12:40 +08:00
|
|
|
|
|
|
|
if (v->flags & VM_ALLOC)
|
2014-06-05 07:08:09 +08:00
|
|
|
seq_puts(m, " vmalloc");
|
2008-04-28 17:12:40 +08:00
|
|
|
|
|
|
|
if (v->flags & VM_MAP)
|
2014-06-05 07:08:09 +08:00
|
|
|
seq_puts(m, " vmap");
|
2008-04-28 17:12:40 +08:00
|
|
|
|
|
|
|
if (v->flags & VM_USERMAP)
|
2014-06-05 07:08:09 +08:00
|
|
|
seq_puts(m, " user");
|
2008-04-28 17:12:40 +08:00
|
|
|
|
2019-06-03 14:55:13 +08:00
|
|
|
if (v->flags & VM_DMA_COHERENT)
|
|
|
|
seq_puts(m, " dma-coherent");
|
|
|
|
|
2016-01-15 07:19:35 +08:00
|
|
|
if (is_vmalloc_addr(v->pages))
|
2014-06-05 07:08:09 +08:00
|
|
|
seq_puts(m, " vpages");
|
2008-04-28 17:12:40 +08:00
|
|
|
|
2008-07-24 12:27:38 +08:00
|
|
|
show_numa_info(m, v);
|
2008-04-28 17:12:40 +08:00
|
|
|
seq_putc(m, '\n');
|
2019-09-24 06:36:36 +08:00
|
|
|
|
|
|
|
/*
|
|
|
|
* As a final step, dump "unpurged" areas. Note,
|
|
|
|
* that entire "/proc/vmallocinfo" output will not
|
|
|
|
* be address sorted, because the purge list is not
|
|
|
|
* sorted.
|
|
|
|
*/
|
|
|
|
if (list_is_last(&va->list, &vmap_area_list))
|
|
|
|
show_purge_info(m);
|
|
|
|
|
2008-04-28 17:12:40 +08:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2008-10-06 07:50:47 +08:00
|
|
|
static const struct seq_operations vmalloc_op = {
|
2008-04-28 17:12:40 +08:00
|
|
|
.start = s_start,
|
|
|
|
.next = s_next,
|
|
|
|
.stop = s_stop,
|
|
|
|
.show = s_show,
|
|
|
|
};
|
2008-10-06 07:50:47 +08:00
|
|
|
|
|
|
|
static int __init proc_vmalloc_init(void)
|
|
|
|
{
|
2018-04-14 01:44:18 +08:00
|
|
|
if (IS_ENABLED(CONFIG_NUMA))
|
2018-06-15 06:27:58 +08:00
|
|
|
proc_create_seq_private("vmallocinfo", 0400, NULL,
|
2018-04-24 23:05:17 +08:00
|
|
|
&vmalloc_op,
|
|
|
|
nr_node_ids * sizeof(unsigned int), NULL);
|
2018-04-14 01:44:18 +08:00
|
|
|
else
|
2018-06-15 06:27:58 +08:00
|
|
|
proc_create_seq("vmallocinfo", 0400, NULL, &vmalloc_op);
|
2008-10-06 07:50:47 +08:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
module_init(proc_vmalloc_init);
|
2013-04-30 06:07:28 +08:00
|
|
|
|
2008-04-28 17:12:40 +08:00
|
|
|
#endif
|