Although shmem_fault() has been careful to count a major fault to vm_mm,
shmem_getpage_gfp() has been careless in charging a remote access fault
to current->mm owner's memcg instead of to vma->vm_mm owner's memcg:
that is inconsistent with all the mem_cgroup charging on remote access
faults in mm/memory.c.
Fix it by passing fault_mm along with fault_type to
shmem_get_page_gfp(); but in that case, now knowing the right mm, it's
better for it to handle the PGMAJFAULT updates itself.
And let's keep this clutter out of most callers' way: change the common
shmem_getpage() wrapper to hide fault_mm and fault_type as well as gfp.
Signed-off-by: Andres Lagar-Cavilla <andreslc@google.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Yang Shi <yang.shi@linaro.org>
Cc: Ning Qu <quning@gmail.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Konstantin Khlebnikov <koct9i@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Make a few cleanups in mm/shmem.c, before going on to complicate it.
shmem_alloc_page() will become more complicated: we can't afford to to
have that complication duplicated between a CONFIG_NUMA version and a
!CONFIG_NUMA version, so rearrange the #ifdef'ery there to yield a
single shmem_swapin() and a single shmem_alloc_page().
Yes, it's a shame to inflict the horrid pseudo-vma on non-NUMA
configurations, but eliminating it is a larger cleanup: I have an
alloc_pages_mpol() patchset not yet ready - mpol handling is subtle and
bug-prone, and changed yet again since my last version.
Move __SetPageLocked, __SetPageSwapBacked from shmem_getpage_gfp() to
shmem_alloc_page(): that SwapBacked flag will be useful in future, to
help to distinguish different cases appropriately.
And the SGP_DIRTY variant of SGP_CACHE is hard to understand and of
little use (IIRC it dates back to when shmem_getpage() returned the page
unlocked): kill it and do the necessary in shmem_file_read_iter().
But an arm64 build then complained that info may be uninitialized (where
shmem_getpage_gfp() deletes a freshly alloced page beyond eof), and
advancing to an "sgp <= SGP_CACHE" test jogged it back to reality.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Yang Shi <yang.shi@linaro.org>
Cc: Ning Qu <quning@gmail.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Konstantin Khlebnikov <koct9i@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
v3.16 commit 07a4278843 ("mm: shmem: avoid atomic operation during
shmem_getpage_gfp") rightly replaced one instance of SetPageSwapBacked
by __SetPageSwapBacked, pointing out that the newly allocated page is
not yet visible to other users (except speculative get_page_unless_zero-
ers, who may not update page flags before their further checks).
That was part of a series in which Mel was focused on tmpfs profiles:
but almost all SetPageSwapBacked uses can be so optimized, with the same
justification.
Remove ClearPageSwapBacked from __read_swap_cache_async() error path:
it's not an error to free a page with PG_swapbacked set.
Follow a convention of __SetPageLocked, __SetPageSwapBacked instead of
doing it differently in different places; but that's for tidiness - if
the ordering actually mattered, we should not be using the __variants.
There's probably scope for further __SetPageFlags in other places, but
SwapBacked is the one I'm interested in at the moment.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Yang Shi <yang.shi@linaro.org>
Cc: Ning Qu <quning@gmail.com>
Reviewed-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Konstantin Khlebnikov <koct9i@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Konstantin Khlebnikov pointed out (nearly four years ago, when lumpy
reclaim was removed) that lru_size can be updated by -nr_taken once per
call to isolate_lru_pages(), instead of page by page.
Update it inside isolate_lru_pages(), or at its two callsites? I chose
to update it at the callsites, rearranging and grouping the updates by
nr_taken and nr_scanned together in both.
With one exception, mem_cgroup_update_lru_size(,lru,) is then used where
__mod_zone_page_state(,NR_LRU_BASE+lru,) is used; and we shall be adding
some more calls in a future commit. Make the code a little smaller and
simpler by incorporating stat update in lru_size update.
The exception was move_active_pages_to_lru(), which aggregated the
pgmoved stat update separately from the individual lru_size updates; but
I still think this a simplification worth making.
However, the __mod_zone_page_state is not peculiar to mem_cgroups: so
better use the name update_lru_size, calls mem_cgroup_update_lru_size
when CONFIG_MEMCG.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Yang Shi <yang.shi@linaro.org>
Cc: Ning Qu <quning@gmail.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Konstantin Khlebnikov <koct9i@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Though debug kernels have a VM_BUG_ON to help protect from misaccounting
lru_size, non-debug kernels are liable to wrap it around: and then the
vast unsigned long size draws page reclaim into a loop of repeatedly
doing nothing on an empty list, without even a cond_resched().
That soft lockup looks confusingly like an over-busy reclaim scenario,
with lots of contention on the lru_lock in shrink_inactive_list(): yet
has a totally different origin.
Help differentiate with a custom warning in
mem_cgroup_update_lru_size(), even in non-debug kernels; and reset the
size to avoid the lockup. But the particular bug which suggested this
change was mine alone, and since fixed.
Make it a WARN_ONCE: the first occurrence is the most informative, a
flurry may follow, yet even when rate-limited little more is learnt.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Yang Shi <yang.shi@linaro.org>
Cc: Ning Qu <quning@gmail.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Konstantin Khlebnikov <koct9i@gmail.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This patchset deals with some problematic sites that iterate pfn ranges.
There is a system thats node's pfns are overlapped as follows:
-----pfn-------->
N0 N1 N2 N0 N1 N2
Therefore, we need to take care of this overlapping when iterating pfn
range.
I audit many iterating sites that uses pfn_valid(), pfn_valid_within(),
zone_start_pfn and etc. and others looks safe to me. This is a
preparation step for a new CMA implementation, ZONE_CMA
(https://lkml.org/lkml/2015/2/12/95), because it would be easily
overlapped with other zones. But, zone overlap check is also needed for
the general case so I send it separately.
This patch (of 5):
alloc_gigantic_page() uses alloc_contig_range() and this requires that
the requested range is in a single zone. To satisfy this requirement,
add this check to pfn_range_valid_gigantic().
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Rik van Riel <riel@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Laura Abbott <lauraa@codeaurora.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Marek Szyprowski <m.szyprowski@samsung.com>
Cc: Michal Nazarewicz <mina86@mina86.com>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
Cc: "Rafael J. Wysocki" <rjw@rjwysocki.net>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The goal of direct compaction is to quickly make a high-order page
available for the pending allocation. Within an aligned block of pages
of desired order, a single allocated page that cannot be isolated for
migration means that the block cannot fully merge to a buddy page that
would satisfy the allocation request. Therefore we can reduce the
allocation stall by skipping the rest of the block immediately on
isolation failure. For async compaction, this also means a higher
chance of succeeding until it detects contention.
We however shouldn't completely sacrifice the second objective of
compaction, which is to reduce overal long-term memory fragmentation.
As a compromise, perform the eager skipping only in direct async
compaction, while sync compaction (including kcompactd) remains
thorough.
Testing was done using stress-highalloc from mmtests, configured for
order-4 GFP_KERNEL allocations:
4.6-rc1 4.6-rc1
before after
Success 1 Min 24.00 ( 0.00%) 27.00 (-12.50%)
Success 1 Mean 30.20 ( 0.00%) 31.60 ( -4.64%)
Success 1 Max 37.00 ( 0.00%) 35.00 ( 5.41%)
Success 2 Min 42.00 ( 0.00%) 32.00 ( 23.81%)
Success 2 Mean 44.00 ( 0.00%) 44.80 ( -1.82%)
Success 2 Max 48.00 ( 0.00%) 52.00 ( -8.33%)
Success 3 Min 91.00 ( 0.00%) 92.00 ( -1.10%)
Success 3 Mean 92.20 ( 0.00%) 92.80 ( -0.65%)
Success 3 Max 94.00 ( 0.00%) 93.00 ( 1.06%)
We can see that success rates are unaffected by the skipping.
4.6-rc1 4.6-rc1
before after
User 2587.42 2566.53
System 482.89 471.20
Elapsed 1395.68 1382.00
Times are not so useful metric for this benchmark as main portion is the
interfering kernel builds, but results do hint at reduced system times.
4.6-rc1 4.6-rc1
before after
Direct pages scanned 163614 159608
Kswapd pages scanned 2070139 2078790
Kswapd pages reclaimed 2061707 2069757
Direct pages reclaimed 163354 159505
Reduced direct reclaim was unintended, but could be explained by more
successful first attempt at (async) direct compaction, which is
attempted before the first reclaim attempt in __alloc_pages_slowpath().
Compaction stalls 33052 39853
Compaction success 12121 19773
Compaction failures 20931 20079
Compaction is indeed more successful, and thus less likely to get
deferred, so there are also more direct compaction stalls.
Page migrate success 3781876 3326819
Page migrate failure 45817 41774
Compaction pages isolated 7868232 6941457
Compaction migrate scanned 168160492 127269354
Compaction migrate prescanned 0 0
Compaction free scanned 2522142582 2326342620
Compaction free direct alloc 0 0
Compaction free dir. all. miss 0 0
Compaction cost 5252 4476
The patch reduces migration scanned pages by 25% thanks to the eager
skipping.
[hughd@google.com: prevent nr_isolated_* from going negative]
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Rik van Riel <riel@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Compaction drains the local pcplists each time migration scanner moves
away from a cc->order aligned block where it isolated pages for
migration, so that the pages freed by migrations can merge into higher
orders.
The detection is currently coarser than it could be. The
cc->last_migrated_pfn variable should track the lowest pfn that was
isolated for migration. But it is set to the pfn where
isolate_migratepages_block() starts scanning, which is typically the
first pfn of the pageblock. There, the scanner might fail to isolate
several order-aligned blocks, and then isolate COMPACT_CLUSTER_MAX in
another block. This would cause the pcplists drain to be performed,
although the scanner didn't yet finish the block where it isolated from.
This patch thus makes cc->last_migrated_pfn handling more accurate by
setting it to the pfn of an actually isolated page in
isolate_migratepages_block(). Although practical effects of this patch
are likely low, it arguably makes the intent of the code more obvious.
Also the next patch will make async direct compaction skip blocks more
aggressively, and draining pcplists due to skipped blocks is wasteful.
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Rik van Riel <riel@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Compaction code has accumulated numerous instances of manual
calculations of the first (inclusive) and last (exclusive) pfn of a
pageblock (or a smaller block of given order), given a pfn within the
pageblock.
Wrap these calculations by introducing pageblock_start_pfn(pfn) and
pageblock_end_pfn(pfn) macros.
[vbabka@suse.cz: fix crash in get_pfnblock_flags_mask() from isolate_freepages():]
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Rik van Riel <riel@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This check effectively catches anon vma hierarchy inconsistence and some
vma corruptions. It was effective for catching corner cases in anon vma
reusing logic. For now this code seems stable so check could be hidden
under CONFIG_DEBUG_VM and replaced with WARN because it's not so fatal.
Signed-off-by: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
Suggested-by: Vasily Averin <vvs@virtuozzo.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
> The comment seems to have not much to do with the code?
I guess the comment tries to say that the code path is triggered when we
charge the page which happens _before_ it is added to the LRU list and
so last_scanned_node might contain the stale data.
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
When any unsupported hugepage size is specified, 'hugepagesz=' and
'hugepages=' should be ignored during command line parsing until any
supported hugepage size is found. But currently incorrect number of
hugepages are allocated when unsupported size is specified as it fails
to ignore the 'hugepages=' command.
Test case:
Note that this is specific to x86 architecture.
Boot the kernel with command line option 'hugepagesz=256M hugepages=X'.
After boot, dmesg output shows that X number of hugepages of the size 2M
is pre-allocated instead of 0.
So, to handle such command line options, introduce new routine
hugetlb_bad_size. The routine hugetlb_bad_size sets the global variable
parsed_valid_hugepagesz. We are using parsed_valid_hugepagesz to save
the state when unsupported hugepagesize is found so that we can ignore
the 'hugepages=' parameters after that and then reset the variable when
supported hugepage size is found.
The routine hugetlb_bad_size can be called while setting 'hugepagesz='
parameter in an architecture specific code.
Signed-off-by: Vaishali Thakkar <vaishali.thakkar@oracle.com>
Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
Reviewed-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: Yaowei Bai <baiyaowei@cmss.chinamobile.com>
Cc: Dominik Dingel <dingel@linux.vnet.ibm.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Paul Gortmaker <paul.gortmaker@windriver.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: James Hogan <james.hogan@imgtec.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
It was observed that minimum size accounting associated with the
hugetlbfs min_size mount option may not perform optimally and as
expected. As huge pages/reservations are released from the filesystem
and given back to the global pools, they are reserved for subsequent
filesystem use as long as the subpool reserved count is less than
subpool minimum size. It does not take into account used pages within
the filesystem. The filesystem size limits are not exceeded and this is
technically not a bug. However, better behavior would be to wait for
the number of used pages/reservations associated with the filesystem to
drop below the minimum size before taking reservations to satisfy
minimum size.
An optimization is also made to the hugepage_subpool_get_pages() routine
which is called when pages/reservations are allocated. This does not
change behavior, but simply avoids the accounting if all reservations
have already been taken (subpool reserved count == 0).
Signed-off-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Paul Gortmaker <paul.gortmaker@windriver.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Lots of code does
node = next_node(node, XXX);
if (node == MAX_NUMNODES)
node = first_node(XXX);
so create next_node_in() to do this and use it in various places.
[mhocko@suse.com: use next_node_in() helper]
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Michal Hocko <mhocko@suse.com>
Cc: Xishi Qiu <qiuxishi@huawei.com>
Cc: Joonsoo Kim <js1304@gmail.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Laura Abbott <lauraa@codeaurora.org>
Cc: Hui Zhu <zhuhui@xiaomi.com>
Cc: Wang Xiaoqiang <wangxq10@lzu.edu.cn>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Many developers already know that field for reference count of the
struct page is _count and atomic type. They would try to handle it
directly and this could break the purpose of page reference count
tracepoint. To prevent direct _count modification, this patch rename it
to _refcount and add warning message on the code. After that, developer
who need to handle reference count will find that field should not be
accessed directly.
[akpm@linux-foundation.org: fix comments, per Vlastimil]
[akpm@linux-foundation.org: Documentation/vm/transhuge.txt too]
[sfr@canb.auug.org.au: sync ethernet driver changes]
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Hugh Dickins <hughd@google.com>
Cc: Johannes Berg <johannes@sipsolutions.net>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Sunil Goutham <sgoutham@cavium.com>
Cc: Chris Metcalf <cmetcalf@mellanox.com>
Cc: Manish Chopra <manish.chopra@qlogic.com>
Cc: Yuval Mintz <yuval.mintz@qlogic.com>
Cc: Tariq Toukan <tariqt@mellanox.com>
Cc: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Provides an optional config (CONFIG_SLAB_FREELIST_RANDOM) to randomize
the SLAB freelist. The list is randomized during initialization of a
new set of pages. The order on different freelist sizes is pre-computed
at boot for performance. Each kmem_cache has its own randomized
freelist. Before pre-computed lists are available freelists are
generated dynamically. This security feature reduces the predictability
of the kernel SLAB allocator against heap overflows rendering attacks
much less stable.
For example this attack against SLUB (also applicable against SLAB)
would be affected:
https://jon.oberheide.org/blog/2010/09/10/linux-kernel-can-slub-overflow/
Also, since v4.6 the freelist was moved at the end of the SLAB. It
means a controllable heap is opened to new attacks not yet publicly
discussed. A kernel heap overflow can be transformed to multiple
use-after-free. This feature makes this type of attack harder too.
To generate entropy, we use get_random_bytes_arch because 0 bits of
entropy is available in the boot stage. In the worse case this function
will fallback to the get_random_bytes sub API. We also generate a shift
random number to shift pre-computed freelist for each new set of pages.
The config option name is not specific to the SLAB as this approach will
be extended to other allocators like SLUB.
Performance results highlighted no major changes:
Hackbench (running 90 10 times):
Before average: 0.0698
After average: 0.0663 (-5.01%)
slab_test 1 run on boot. Difference only seen on the 2048 size test
being the worse case scenario covered by freelist randomization. New
slab pages are constantly being created on the 10000 allocations.
Variance should be mainly due to getting new pages every few
allocations.
Before:
Single thread testing
=====================
1. Kmalloc: Repeatedly allocate then free test
10000 times kmalloc(8) -> 99 cycles kfree -> 112 cycles
10000 times kmalloc(16) -> 109 cycles kfree -> 140 cycles
10000 times kmalloc(32) -> 129 cycles kfree -> 137 cycles
10000 times kmalloc(64) -> 141 cycles kfree -> 141 cycles
10000 times kmalloc(128) -> 152 cycles kfree -> 148 cycles
10000 times kmalloc(256) -> 195 cycles kfree -> 167 cycles
10000 times kmalloc(512) -> 257 cycles kfree -> 199 cycles
10000 times kmalloc(1024) -> 393 cycles kfree -> 251 cycles
10000 times kmalloc(2048) -> 649 cycles kfree -> 228 cycles
10000 times kmalloc(4096) -> 806 cycles kfree -> 370 cycles
10000 times kmalloc(8192) -> 814 cycles kfree -> 411 cycles
10000 times kmalloc(16384) -> 892 cycles kfree -> 455 cycles
2. Kmalloc: alloc/free test
10000 times kmalloc(8)/kfree -> 121 cycles
10000 times kmalloc(16)/kfree -> 121 cycles
10000 times kmalloc(32)/kfree -> 121 cycles
10000 times kmalloc(64)/kfree -> 121 cycles
10000 times kmalloc(128)/kfree -> 121 cycles
10000 times kmalloc(256)/kfree -> 119 cycles
10000 times kmalloc(512)/kfree -> 119 cycles
10000 times kmalloc(1024)/kfree -> 119 cycles
10000 times kmalloc(2048)/kfree -> 119 cycles
10000 times kmalloc(4096)/kfree -> 121 cycles
10000 times kmalloc(8192)/kfree -> 119 cycles
10000 times kmalloc(16384)/kfree -> 119 cycles
After:
Single thread testing
=====================
1. Kmalloc: Repeatedly allocate then free test
10000 times kmalloc(8) -> 130 cycles kfree -> 86 cycles
10000 times kmalloc(16) -> 118 cycles kfree -> 86 cycles
10000 times kmalloc(32) -> 121 cycles kfree -> 85 cycles
10000 times kmalloc(64) -> 176 cycles kfree -> 102 cycles
10000 times kmalloc(128) -> 178 cycles kfree -> 100 cycles
10000 times kmalloc(256) -> 205 cycles kfree -> 109 cycles
10000 times kmalloc(512) -> 262 cycles kfree -> 136 cycles
10000 times kmalloc(1024) -> 342 cycles kfree -> 157 cycles
10000 times kmalloc(2048) -> 701 cycles kfree -> 238 cycles
10000 times kmalloc(4096) -> 803 cycles kfree -> 364 cycles
10000 times kmalloc(8192) -> 835 cycles kfree -> 404 cycles
10000 times kmalloc(16384) -> 896 cycles kfree -> 441 cycles
2. Kmalloc: alloc/free test
10000 times kmalloc(8)/kfree -> 121 cycles
10000 times kmalloc(16)/kfree -> 121 cycles
10000 times kmalloc(32)/kfree -> 123 cycles
10000 times kmalloc(64)/kfree -> 142 cycles
10000 times kmalloc(128)/kfree -> 121 cycles
10000 times kmalloc(256)/kfree -> 119 cycles
10000 times kmalloc(512)/kfree -> 119 cycles
10000 times kmalloc(1024)/kfree -> 119 cycles
10000 times kmalloc(2048)/kfree -> 119 cycles
10000 times kmalloc(4096)/kfree -> 119 cycles
10000 times kmalloc(8192)/kfree -> 119 cycles
10000 times kmalloc(16384)/kfree -> 119 cycles
[akpm@linux-foundation.org: propagate gfp_t into cache_random_seq_create()]
Signed-off-by: Thomas Garnier <thgarnie@google.com>
Acked-by: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Kees Cook <keescook@chromium.org>
Cc: Greg Thelen <gthelen@google.com>
Cc: Laura Abbott <labbott@fedoraproject.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
When we call __kmem_cache_shrink on memory cgroup removal, we need to
synchronize kmem_cache->cpu_partial update with put_cpu_partial that
might be running on other cpus. Currently, we achieve that by using
kick_all_cpus_sync, which works as a system wide memory barrier. Though
fast it is, this method has a flaw - it issues a lot of IPIs, which
might hurt high performance or real-time workloads.
To fix this, let's replace kick_all_cpus_sync with synchronize_sched.
Although the latter one may take much longer to finish, it shouldn't be
a problem in this particular case, because memory cgroups are destroyed
asynchronously from a workqueue so that no user visible effects should
be introduced. OTOH, it will save us from excessive IPIs when someone
removes a cgroup.
Anyway, even if using synchronize_sched turns out to take too long, we
can always introduce a kind of __kmem_cache_shrink batching so that this
method would only be called once per one cgroup destruction (not per
each per memcg kmem cache as it is now).
Signed-off-by: Vladimir Davydov <vdavydov@virtuozzo.com>
Reported-by: Peter Zijlstra <peterz@infradead.org>
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
To check whether free objects exist or not precisely, we need to grab a
lock. But, accuracy isn't that important because race window would be
even small and if there is too much free object, cache reaper would reap
it. So, this patch makes the check for free object exisistence not to
hold a lock. This will reduce lock contention in heavily allocation
case.
Note that until now, n->shared can be freed during the processing by
writing slabinfo, but, with some trick in this patch, we can access it
freely within interrupt disabled period.
Below is the result of concurrent allocation/free in slab allocation
benchmark made by Christoph a long time ago. I make the output simpler.
The number shows cycle count during alloc/free respectively so less is
better.
* Before
Kmalloc N*alloc N*free(32): Average=248/966
Kmalloc N*alloc N*free(64): Average=261/949
Kmalloc N*alloc N*free(128): Average=314/1016
Kmalloc N*alloc N*free(256): Average=741/1061
Kmalloc N*alloc N*free(512): Average=1246/1152
Kmalloc N*alloc N*free(1024): Average=2437/1259
Kmalloc N*alloc N*free(2048): Average=4980/1800
Kmalloc N*alloc N*free(4096): Average=9000/2078
* After
Kmalloc N*alloc N*free(32): Average=344/792
Kmalloc N*alloc N*free(64): Average=347/882
Kmalloc N*alloc N*free(128): Average=390/959
Kmalloc N*alloc N*free(256): Average=393/1067
Kmalloc N*alloc N*free(512): Average=683/1229
Kmalloc N*alloc N*free(1024): Average=1295/1325
Kmalloc N*alloc N*free(2048): Average=2513/1664
Kmalloc N*alloc N*free(4096): Average=4742/2172
It shows that allocation performance decreases for the object size up to
128 and it may be due to extra checks in cache_alloc_refill(). But,
with considering improvement of free performance, net result looks the
same. Result for other size class looks very promising, roughly, 50%
performance improvement.
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Until now, cache growing makes a free slab on node's slab list and then
we can allocate free objects from it. This necessarily requires to hold
a node lock which is very contended. If we refill cpu cache before
attaching it to node's slab list, we can avoid holding a node lock as
much as possible because this newly allocated slab is only visible to
the current task. This will reduce lock contention.
Below is the result of concurrent allocation/free in slab allocation
benchmark made by Christoph a long time ago. I make the output simpler.
The number shows cycle count during alloc/free respectively so less is
better.
* Before
Kmalloc N*alloc N*free(32): Average=355/750
Kmalloc N*alloc N*free(64): Average=452/812
Kmalloc N*alloc N*free(128): Average=559/1070
Kmalloc N*alloc N*free(256): Average=1176/980
Kmalloc N*alloc N*free(512): Average=1939/1189
Kmalloc N*alloc N*free(1024): Average=3521/1278
Kmalloc N*alloc N*free(2048): Average=7152/1838
Kmalloc N*alloc N*free(4096): Average=13438/2013
* After
Kmalloc N*alloc N*free(32): Average=248/966
Kmalloc N*alloc N*free(64): Average=261/949
Kmalloc N*alloc N*free(128): Average=314/1016
Kmalloc N*alloc N*free(256): Average=741/1061
Kmalloc N*alloc N*free(512): Average=1246/1152
Kmalloc N*alloc N*free(1024): Average=2437/1259
Kmalloc N*alloc N*free(2048): Average=4980/1800
Kmalloc N*alloc N*free(4096): Average=9000/2078
It shows that contention is reduced for all the object sizes and
performance increases by 30 ~ 40%.
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This is a preparation step to implement lockless allocation path when
there is no free objects in kmem_cache.
What we'd like to do here is to refill cpu cache without holding a node
lock. To accomplish this purpose, refill should be done after new slab
allocation but before attaching the slab to the management list. So,
this patch separates cache_grow() to two parts, allocation and attaching
to the list in order to add some code inbetween them in the following
patch.
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Currently, cache_grow() assumes that allocated page's nodeid would be
same with parameter nodeid which is used for allocation request. If we
discard this assumption, we can handle fallback_alloc() case gracefully.
So, this patch makes cache_grow() handle the page allocated on arbitrary
node and clean-up relevant code.
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Slab color isn't needed to be changed strictly. Because locking for
changing slab color could cause more lock contention so this patch
implements racy access/modify the slab color. This is a preparation
step to implement lockless allocation path when there is no free objects
in the kmem_cache.
Below is the result of concurrent allocation/free in slab allocation
benchmark made by Christoph a long time ago. I make the output simpler.
The number shows cycle count during alloc/free respectively so less is
better.
* Before
Kmalloc N*alloc N*free(32): Average=365/806
Kmalloc N*alloc N*free(64): Average=452/690
Kmalloc N*alloc N*free(128): Average=736/886
Kmalloc N*alloc N*free(256): Average=1167/985
Kmalloc N*alloc N*free(512): Average=2088/1125
Kmalloc N*alloc N*free(1024): Average=4115/1184
Kmalloc N*alloc N*free(2048): Average=8451/1748
Kmalloc N*alloc N*free(4096): Average=16024/2048
* After
Kmalloc N*alloc N*free(32): Average=355/750
Kmalloc N*alloc N*free(64): Average=452/812
Kmalloc N*alloc N*free(128): Average=559/1070
Kmalloc N*alloc N*free(256): Average=1176/980
Kmalloc N*alloc N*free(512): Average=1939/1189
Kmalloc N*alloc N*free(1024): Average=3521/1278
Kmalloc N*alloc N*free(2048): Average=7152/1838
Kmalloc N*alloc N*free(4096): Average=13438/2013
It shows that contention is reduced for object size >= 1024 and
performance increases by roughly 15%.
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Acked-by: Christoph Lameter <cl@linux.com>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Currently, determination to free a slab is done whenever each freed
object is put into the slab. This has a following problem.
Assume free_limit = 10 and nr_free = 9.
Free happens as following sequence and nr_free changes as following.
free(become a free slab) free(not become a free slab) nr_free: 9 -> 10
(at first free) -> 11 (at second free)
If we try to check if we can free current slab or not on each object
free, we can't free any slab in this situation because current slab
isn't a free slab when nr_free exceed free_limit (at second free) even
if there is a free slab.
However, if we check it lastly, we can free 1 free slab.
This problem would cause to keep too much memory in the slab subsystem.
This patch try to fix it by checking number of free object after all
free work is done. If there is free slab at that time, we can free slab
as much as possible so we keep free slab as minimal.
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Initial attemp to remove BAD_ALIEN_MAGIC is once reverted by 'commit
edcad25095 ("Revert "slab: remove BAD_ALIEN_MAGIC"")' because it
causes a problem on m68k which has many node but !CONFIG_NUMA. In this
case, although alien cache isn't used at all but to cope with some
initialization path, garbage value is used and that is BAD_ALIEN_MAGIC.
Now, this patch set use_alien_caches to 0 when !CONFIG_NUMA, there is no
initialization path problem so we don't need BAD_ALIEN_MAGIC at all. So
remove it.
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Tested-by: Geert Uytterhoeven <geert@linux-m68k.org>
Acked-by: Christoph Lameter <cl@linux.com>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
While processing concurrent allocation, SLAB could be contended a lot
because it did a lots of work with holding a lock. This patchset try to
reduce the number of critical section to reduce lock contention. Major
changes are lockless decision to allocate more slab and lockless cpu
cache refill from the newly allocated slab.
Below is the result of concurrent allocation/free in slab allocation
benchmark made by Christoph a long time ago. I make the output simpler.
The number shows cycle count during alloc/free respectively so less is
better.
* Before
Kmalloc N*alloc N*free(32): Average=365/806
Kmalloc N*alloc N*free(64): Average=452/690
Kmalloc N*alloc N*free(128): Average=736/886
Kmalloc N*alloc N*free(256): Average=1167/985
Kmalloc N*alloc N*free(512): Average=2088/1125
Kmalloc N*alloc N*free(1024): Average=4115/1184
Kmalloc N*alloc N*free(2048): Average=8451/1748
Kmalloc N*alloc N*free(4096): Average=16024/2048
* After
Kmalloc N*alloc N*free(32): Average=344/792
Kmalloc N*alloc N*free(64): Average=347/882
Kmalloc N*alloc N*free(128): Average=390/959
Kmalloc N*alloc N*free(256): Average=393/1067
Kmalloc N*alloc N*free(512): Average=683/1229
Kmalloc N*alloc N*free(1024): Average=1295/1325
Kmalloc N*alloc N*free(2048): Average=2513/1664
Kmalloc N*alloc N*free(4096): Average=4742/2172
It shows that performance improves greatly (roughly more than 50%) for
the object class whose size is more than 128 bytes.
This patch (of 11):
If we don't hold neither the slab_mutex nor the node lock, node's shared
array cache could be freed and re-populated. If __kmem_cache_shrink()
is called at the same time, it will call drain_array() with n->shared
without holding node lock so problem can happen. This patch fix the
situation by holding the node lock before trying to drain the shared
array.
In addition, add a debug check to confirm that n->shared access race
doesn't exist.
Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Jesper Dangaard Brouer <brouer@redhat.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Currently faults are protected against truncate by filesystem specific
i_mmap_sem and page lock in case of hole page. Cow faults are protected
DAX radix tree entry locking. So there's no need for i_mmap_lock in DAX
code. Remove it.
Reviewed-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
When doing cow faults, we cannot directly fill in PTE as we do for other
faults as we rely on generic code to do proper accounting of the cowed page.
We also have no page to lock to protect against races with truncate as
other faults have and we need the protection to extend until the moment
generic code inserts cowed page into PTE thus at that point we have no
protection of fs-specific i_mmap_sem. So far we relied on using
i_mmap_lock for the protection however that is completely special to cow
faults. To make fault locking more uniform use DAX entry lock instead.
Reviewed-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Currently DAX page fault locking is racy.
CPU0 (write fault) CPU1 (read fault)
__dax_fault() __dax_fault()
get_block(inode, block, &bh, 0) -> not mapped
get_block(inode, block, &bh, 0)
-> not mapped
if (!buffer_mapped(&bh))
if (vmf->flags & FAULT_FLAG_WRITE)
get_block(inode, block, &bh, 1) -> allocates blocks
if (page) -> no
if (!buffer_mapped(&bh))
if (vmf->flags & FAULT_FLAG_WRITE) {
} else {
dax_load_hole();
}
dax_insert_mapping()
And we are in a situation where we fail in dax_radix_entry() with -EIO.
Another problem with the current DAX page fault locking is that there is
no race-free way to clear dirty tag in the radix tree. We can always
end up with clean radix tree and dirty data in CPU cache.
We fix the first problem by introducing locking of exceptional radix
tree entries in DAX mappings acting very similarly to page lock and thus
synchronizing properly faults against the same mapping index. The same
lock can later be used to avoid races when clearing radix tree dirty
tag.
Reviewed-by: NeilBrown <neilb@suse.com>
Reviewed-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Currently we forbid page_cache_tree_insert() to replace exceptional radix
tree entries for DAX inodes. However to make DAX faults race free we will
lock radix tree entries and when hole is created, we need to replace
such locked radix tree entry with a hole page. So modify
page_cache_tree_insert() to allow that.
Reviewed-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Pull vfs cleanups from Al Viro:
"More cleanups from Christoph"
* 'work.preadv2' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs:
nfsd: use RWF_SYNC
fs: add RWF_DSYNC aand RWF_SYNC
ceph: use generic_write_sync
fs: simplify the generic_write_sync prototype
fs: add IOCB_SYNC and IOCB_DSYNC
direct-io: remove the offset argument to dio_complete
direct-io: eliminate the offset argument to ->direct_IO
xfs: eliminate the pos variable in xfs_file_dio_aio_write
filemap: remove the pos argument to generic_file_direct_write
filemap: remove pos variables in generic_file_read_iter
