- 21 May, 2016 40 commits
-
-
Christoph Lameter authored
The cpu_stat_off variable is unecessary since we can check if a workqueue request is pending otherwise. Removal of cpu_stat_off makes it pretty easy for the vmstat shepherd to ensure that the proper things happen. Removing the state also removes all races related to it. Should a workqueue not be scheduled as needed for vmstat_update then the shepherd will notice and schedule it as needed. Should a workqueue be unecessarily scheduled then the vmstat updater will disable it. [akpm@linux-foundation.org: fix indentation, per Michal] Link: http://lkml.kernel.org/r/alpine.DEB.2.20.1605061306460.17934@east.gentwo.orgSigned-off-by:
Christoph Lameter <cl@linux.com> Cc: Tejun Heo <htejun@gmail.com> Acked-by:
Michal Hocko <mhocko@suse.com> Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp> Signed-off-by:
Andrew Morton <akpm@linux-foundation.org> Signed-off-by:
Linus Torvalds <torvalds@linux-foundation.org>
-
Greg Thelen authored
Commit f61c42a7 ("memcg: remove tasks/children test from mem_cgroup_force_empty()") removed memory reparenting from the function. Fix the function's comment. Link: http://lkml.kernel.org/r/1462569810-54496-1-git-send-email-gthelen@google.comSigned-off-by:
Greg Thelen <gthelen@google.com> Acked-by:
Johannes Weiner <hannes@cmpxchg.org> Acked-by:
-
Yu Zhao authored
struct page->flags is unsigned long, so when shifting bits we should use UL suffix to match it. Found this problem after I added 64-bit CPU specific page flags and failed to compile the kernel: mm/page_alloc.c: In function '__free_one_page': mm/page_alloc.c:672:2: error: integer overflow in expression [-Werror=overflow] Link: http://lkml.kernel.org/r/1461971723-16187-1-git-send-email-yuzhao@google.comSigned-off-by:
Yu Zhao <yuzhao@google.com> Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com> Cc: Jerome Marchand <jmarchan@redhat.com> Cc: Denys Vlasenko <dvlasenk@redhat.com> Signed-off-by:
-
Minfei Huang authored
It's more convenient to use existing function helper to convert string "on/off" to boolean. Link: http://lkml.kernel.org/r/1461908824-16129-1-git-send-email-mnghuan@gmail.comSigned-off-by:
Minfei Huang <mnghuan@gmail.com> Acked-by:
-
Tetsuo Handa authored
When writeback operation cannot make forward progress because memory allocation requests needed for doing I/O cannot be satisfied (e.g. under OOM-livelock situation), we can observe flood of order-0 page allocation failure messages caused by complete depletion of memory reserves. This is caused by unconditionally allocating "struct wb_writeback_work" objects using GFP_ATOMIC from PF_MEMALLOC context. __alloc_pages_nodemask() { __alloc_pages_slowpath() { __alloc_pages_direct_reclaim() { __perform_reclaim() { current->flags |= PF_MEMALLOC; try_to_free_pages() { do_try_to_free_pages() { wakeup_flusher_threads() { wb_start_writeback() { kzalloc(sizeof(*work), GFP_ATOMIC) { /* ALLOC_NO_WATERMARKS via PF_MEMALLOC */ } } } } } current->flags &= ~PF_MEMALLOC; } } } } Since I/O is stalling, allocating writeback requests forever shall deplete memory reserves. Fortunately, since wb_start_writeback() can fall back to wb_wakeup() when allocating "struct wb_writeback_work" failed, we don't need to allow wb_start_writeback() to use memory reserves. Mem-Info: active_anon:289393 inactive_anon:2093 isolated_anon:29 active_file:10838 inactive_file:113013 isolated_file:859 unevictable:0 dirty:108531 writeback:5308 unstable:0 slab_reclaimable:5526 slab_unreclaimable:7077 mapped:9970 shmem:2159 pagetables:2387 bounce:0 free:3042 free_pcp:0 free_cma:0 Node 0 DMA free:6968kB min:44kB low:52kB high:64kB active_anon:6056kB inactive_anon:176kB active_file:712kB inactive_file:744kB unevictable:0kB isolated(anon):0kB isolated(file):0kB present:15988kB managed:15904kB mlocked:0kB dirty:756kB writeback:0kB mapped:736kB shmem:184kB slab_reclaimable:48kB slab_unreclaimable:208kB kernel_stack:160kB pagetables:144kB unstable:0kB bounce:0kB free_pcp:0kB local_pcp:0kB free_cma:0kB writeback_tmp:0kB pages_scanned:9708 all_unreclaimable? yes lowmem_reserve[]: 0 1732 1732 1732 Node 0 DMA32 free:5200kB min:5200kB low:6500kB high:7800kB active_anon:1151516kB inactive_anon:8196kB active_file:42640kB inactive_file:451076kB unevictable:0kB isolated(anon):116kB isolated(file):3564kB present:2080640kB managed:1775332kB mlocked:0kB dirty:433368kB writeback:21232kB mapped:39144kB shmem:8452kB slab_reclaimable:22056kB slab_unreclaimable:28100kB kernel_stack:20976kB pagetables:9404kB unstable:0kB bounce:0kB free_pcp:120kB local_pcp:0kB free_cma:0kB writeback_tmp:0kB pages_scanned:2701604 all_unreclaimable? no lowmem_reserve[]: 0 0 0 0 Node 0 DMA: 25*4kB (UME) 16*8kB (UME) 3*16kB (UE) 5*32kB (UME) 2*64kB (UM) 2*128kB (ME) 2*256kB (ME) 1*512kB (E) 1*1024kB (E) 2*2048kB (ME) 0*4096kB = 6964kB Node 0 DMA32: 925*4kB (UME) 140*8kB (UME) 5*16kB (ME) 5*32kB (M) 0*64kB 0*128kB 0*256kB 0*512kB 0*1024kB 0*2048kB 0*4096kB = 5060kB Node 0 hugepages_total=0 hugepages_free=0 hugepages_surp=0 hugepages_size=1048576kB Node 0 hugepages_total=0 hugepages_free=0 hugepages_surp=0 hugepages_size=2048kB 126847 total pagecache pages 0 pages in swap cache Swap cache stats: add 0, delete 0, find 0/0 Free swap = 0kB Total swap = 0kB 524157 pages RAM 0 pages HighMem/MovableOnly 76348 pages reserved 0 pages hwpoisoned Out of memory: Kill process 4450 (file_io.00) score 998 or sacrifice child Killed process 4450 (file_io.00) total-vm:4308kB, anon-rss:100kB, file-rss:1184kB, shmem-rss:0kB kthreadd: page allocation failure: order:0, mode:0x2200020 file_io.00: page allocation failure: order:0, mode:0x2200020 CPU: 0 PID: 4457 Comm: file_io.00 Not tainted 4.5.0-rc7+ #45 Hardware name: VMware, Inc. VMware Virtual Platform/440BX Desktop Reference Platform, BIOS 6.00 07/31/2013 Call Trace: warn_alloc_failed+0xf7/0x150 __alloc_pages_nodemask+0x23f/0xa60 alloc_pages_current+0x87/0x110 new_slab+0x3a1/0x440 ___slab_alloc+0x3cf/0x590 __slab_alloc.isra.64+0x18/0x1d kmem_cache_alloc+0x11c/0x150 wb_start_writeback+0x39/0x90 wakeup_flusher_threads+0x7f/0xf0 do_try_to_free_pages+0x1f9/0x410 try_to_free_pages+0x94/0xc0 __alloc_pages_nodemask+0x566/0xa60 alloc_pages_current+0x87/0x110 __page_cache_alloc+0xaf/0xc0 pagecache_get_page+0x88/0x260 grab_cache_page_write_begin+0x21/0x40 xfs_vm_write_begin+0x2f/0xf0 generic_perform_write+0xca/0x1c0 xfs_file_buffered_aio_write+0xcc/0x1f0 xfs_file_write_iter+0x84/0x140 __vfs_write+0xc7/0x100 vfs_write+0x9d/0x190 SyS_write+0x50/0xc0 entry_SYSCALL_64_fastpath+0x12/0x6a Mem-Info: active_anon:293335 inactive_anon:2093 isolated_anon:0 active_file:10829 inactive_file:110045 isolated_file:32 unevictable:0 dirty:109275 writeback:822 unstable:0 slab_reclaimable:5489 slab_unreclaimable:10070 mapped:9999 shmem:2159 pagetables:2420 bounce:0 free:3 free_pcp:0 free_cma:0 Node 0 DMA free:12kB min:44kB low:52kB high:64kB active_anon:6060kB inactive_anon:176kB active_file:708kB inactive_file:756kB unevictable:0kB isolated(anon):0kB isolated(file):0kB present:15988kB managed:15904kB mlocked:0kB dirty:756kB writeback:0kB mapped:736kB shmem:184kB slab_reclaimable:48kB slab_unreclaimable:7160kB kernel_stack:160kB pagetables:144kB unstable:0kB bounce:0kB free_pcp:0kB local_pcp:0kB free_cma:0kB writeback_tmp:0kB pages_scanned:9844 all_unreclaimable? yes lowmem_reserve[]: 0 1732 1732 1732 Node 0 DMA32 free:0kB min:5200kB low:6500kB high:7800kB active_anon:1167280kB inactive_anon:8196kB active_file:42608kB inactive_file:439424kB unevictable:0kB isolated(anon):0kB isolated(file):128kB present:2080640kB managed:1775332kB mlocked:0kB dirty:436344kB writeback:3288kB mapped:39260kB shmem:8452kB slab_reclaimable:21908kB slab_unreclaimable:33120kB kernel_stack:20976kB pagetables:9536kB unstable:0kB bounce:0kB free_pcp:0kB local_pcp:0kB free_cma:0kB writeback_tmp:0kB pages_scanned:11073180 all_unreclaimable? yes lowmem_reserve[]: 0 0 0 0 Node 0 DMA: 0*4kB 0*8kB 0*16kB 0*32kB 0*64kB 0*128kB 0*256kB 0*512kB 0*1024kB 0*2048kB 0*4096kB = 0kB Node 0 DMA32: 0*4kB 0*8kB 0*16kB 0*32kB 0*64kB 0*128kB 0*256kB 0*512kB 0*1024kB 0*2048kB 0*4096kB = 0kB Node 0 hugepages_total=0 hugepages_free=0 hugepages_surp=0 hugepages_size=1048576kB Node 0 hugepages_total=0 hugepages_free=0 hugepages_surp=0 hugepages_size=2048kB 123086 total pagecache pages 0 pages in swap cache Swap cache stats: add 0, delete 0, find 0/0 Free swap = 0kB Total swap = 0kB 524157 pages RAM 0 pages HighMem/MovableOnly 76348 pages reserved 0 pages hwpoisoned SLUB: Unable to allocate memory on node -1 (gfp=0x2088020) cache: kmalloc-64, object size: 64, buffer size: 64, default order: 0, min order: 0 node 0: slabs: 3218, objs: 205952, free: 0 file_io.00: page allocation failure: order:0, mode:0x2200020 CPU: 0 PID: 4457 Comm: file_io.00 Not tainted 4.5.0-rc7+ #45 Assuming that somebody will find a better solution, let's apply this patch for now to stop bleeding, for this problem frequently prevents me from testing OOM livelock condition. Link: http://lkml.kernel.org/r/20160318131136.GE7152@quack.suse.czSigned-off-by:Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp> Acked-by:
-
Eric Engestrom authored
Signed-off-by:
Eric Engestrom <eric@engestrom.ch> Cc: Jonathan Corbet <corbet@lwn.net> Signed-off-by:
-
Weijie Yang authored
If SPARSEMEM, use page_ext in mem_section if !SPARSEMEM, use page_ext in pgdata Signed-off-by:
Weijie Yang <weijie.yang@samsung.com> Signed-off-by:
-
Chen Gang authored
It is used as a pure bool function within kernel source wide. Signed-off-by:
Chen Gang <gang.chen.5i5j@gmail.com> Signed-off-by:
-
Chen Gang authored
Macro HUGETLBFS_SB is clear enough, so one statement is clearer than 3 lines statements. Remove redundant return statements for non-return functions, which can save lines, at least. Signed-off-by:
-
Ming Li authored
Put the activate_page_pvecs definition next to those of the other pagevecs, for clarity. Signed-off-by:
Ming Li <mingli199x@qq.com> Signed-off-by:
-
Eric Dumazet authored
copy_page_to_iter_iovec() is currently the only user of fault_in_pages_writeable(), and it definitely can use fragments from high order pages. Make sure fault_in_pages_writeable() is only touching two adjacent pages at most, as claimed. Signed-off-by:
Eric Dumazet <edumazet@google.com> Cc: Al Viro <viro@zeniv.linux.org.uk> Signed-off-by:
-
David Rientjes authored
The page_counter rounds limits down to page size values. This makes sense, except in the case of hugetlb_cgroup where it's not possible to charge partial hugepages. If the hugetlb_cgroup margin is less than the hugepage size being charged, it will fail as expected. Round the hugetlb_cgroup limit down to hugepage size, since it is the effective limit of the cgroup. For consistency, round down PAGE_COUNTER_MAX as well when a hugetlb_cgroup is created: this prevents error reports when a user cannot restore the value to the kernel default. Signed-off-by:
David Rientjes <rientjes@google.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Nikolay Borisov <kernel@kyup.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com> Signed-off-by:
-
Rich Felker authored
The nommu do_mmap expects f_op->get_unmapped_area to either succeed or return -ENOSYS for VM_MAYSHARE (e.g. private read-only) mappings. Returning addr in the non-MAP_SHARED case was completely wrong, and only happened to work because addr was 0. However, it prevented VM_MAYSHARE mappings from sharing backing with the fs cache, and forced such mappings (including shareable program text) to be copied whenever the number of mappings transitioned from 0 to 1, impacting performance and memory usage. Subsequent mappings beyond the first still correctly shared memory with the first. Instead, treat VM_MAYSHARE identically to VM_SHARED at the file ops level; do_mmap already handles the semantic differences between them. Signed-off-by:
Rich Felker <dalias@libc.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Greg Ungerer <gerg@uclinux.org> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Yoshinori Sato <ysato@users.sourceforge.jp> Signed-off-by:
-
Konstantin Khlebnikov authored
Since commit 84638335 ("mm: rework virtual memory accounting") RLIMIT_DATA limits both brk() and private mmap() but this's disabled by default because of incompatibility with older versions of valgrind. Valgrind always set limit to zero and fails if RLIMIT_DATA is enabled. Fortunately it changes only rlim_cur and keeps rlim_max for reverting limit back when needed. This patch checks current usage also against rlim_max if rlim_cur is zero. This is safe because task anyway can increase rlim_cur up to rlim_max. Size of brk is still checked against rlim_cur, so this part is completely compatible - zero rlim_cur forbids brk() but allows private mmap(). Link: http://lkml.kernel.org/r/56A28613.5070104@de.ibm.comSigned-off-by:
Konstantin Khlebnikov <koct9i@gmail.com> Acked-by:
-
Yongji Xie authored
We use generic hooks in remap_pfn_range() to help archs to track pfnmap regions. The code is something like: int remap_pfn_range() { ... track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size)); ... pfn -= addr >> PAGE_SHIFT; ... untrack_pfn(vma, pfn, PAGE_ALIGN(size)); ... } Here we can easily find the pfn is changed but not recovered before untrack_pfn() is called. That's incorrect. There are no known runtime effects - this is from inspection. Signed-off-by:Yongji Xie <xyjxie@linux.vnet.ibm.com> Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Jerome Marchand <jmarchan@redhat.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Matthew Wilcox <matthew.r.wilcox@intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Andy Lutomirski <luto@kernel.org> Cc: David Hildenbrand <dahi@linux.vnet.ibm.com> Signed-off-by:
-
Chris Wilson authored
When mixing lots of vmallocs and set_memory_*() (which calls vm_unmap_aliases()) I encountered situations where the performance degraded severely due to the walking of the entire vmap_area list each invocation. One simple improvement is to add the lazily freed vmap_area to a separate lockless free list, such that we then avoid having to walk the full list on each purge. Signed-off-by:
Chris Wilson <chris@chris-wilson.co.uk> Reviewed-by:
Roman Pen <r.peniaev@gmail.com> Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com> Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Cc: Daniel Vetter <daniel.vetter@ffwll.ch> Cc: David Rientjes <rientjes@google.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Roman Pen <r.peniaev@gmail.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Toshi Kani <toshi.kani@hp.com> Cc: Shawn Lin <shawn.lin@rock-chips.com> Signed-off-by:
-
Alexander Kuleshov authored
memblock_add_region() and memblock_reserve_region() do nothing specific before the call of memblock_add_range(), only print debug output. We can do the same in memblock_add() and memblock_reserve() since both memblock_add_region() and memblock_reserve_region() are not used by anybody outside of memblock.c and memblock_{add,reserve}() have the same set of flags and nids. Since memblock_add_region() and memblock_reserve_region() will be inlined, there will not be functional changes, but will improve code readability a little. Signed-off-by:Alexander Kuleshov <kuleshovmail@gmail.com> Acked-by:
Ard Biesheuvel <ard.biesheuvel@linaro.org> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Pekka Enberg <penberg@kernel.org> Cc: Tony Luck <tony.luck@intel.com> Cc: Tang Chen <tangchen@cn.fujitsu.com> Cc: David Gibson <david@gibson.dropbear.id.au> Signed-off-by:
-
Chen Yucong authored
HWPoison was specific to some particular x86 platforms. And it is often seen as high level machine check handler. And therefore, 'MCE' is used for the format prefix of printk(). However, 'PowerNV' has also used HWPoison for handling memory errors[1], so 'MCE' is no longer suitable to memory_failure.c. Additionally, 'MCE' and 'Memory failure' have different context. The former belongs to exception context and the latter belongs to process context. Furthermore, HWPoison can also be used for off-lining those sub-health pages that do not trigger any machine check exception. This patch aims to replace 'MCE' with a more appropriate prefix. [1] commit 75eb3d9b ("powerpc/powernv: Get FSP memory errors and plumb into memory poison infrastructure.") Signed-off-by:
Chen Yucong <slaoub@gmail.com> Acked-by:
Naoya Horiguchi <n-horiguchi@ah.jp.nec.com> Signed-off-by:
-
Yang Shi authored
The implementation of mk_huge_pmd looks verbose, it could be just simplified to one line code. Signed-off-by:
Yang Shi <yang.shi@linaro.org> Signed-off-by:
-
Tetsuo Handa authored
Since commit 3a5dda7a ("oom: prevent unnecessary oom kills or kernel panics"), select_bad_process() is using for_each_process_thread(). Since oom_unkillable_task() scans all threads in the caller's thread group and oom_task_origin() scans signal_struct of the caller's thread group, we don't need to call oom_unkillable_task() and oom_task_origin() on each thread. Also, since !mm test will be done later at oom_badness(), we don't need to do !mm test on each thread. Therefore, we only need to do TIF_MEMDIE test on each thread. Although the original code was correct it was quite inefficient because each thread group was scanned num_threads times which can be a lot especially with processes with many threads. Even though the OOM is extremely cold path it is always good to be as effective as possible when we are inside rcu_read_lock() - aka unpreemptible context. If we track number of TIF_MEMDIE threads inside signal_struct, we don't need to do TIF_MEMDIE test on each thread. This will allow select_bad_process() to use for_each_process(). This patch adds a counter to signal_struct for tracking how many TIF_MEMDIE threads are in a given thread group, and check it at oom_scan_process_thread() so that select_bad_process() can use for_each_process() rather than for_each_process_thread(). [mhocko@suse.com: do not blow the signal_struct size] Link: http://lkml.kernel.org/r/20160520075035.GF19172@dhcp22.suse.cz Link: http://lkml.kernel.org/r/201605182230.IDC73435.MVSOHLFOQFOJtF@I-love.SAKURA.ne.jpSigned-off-by:
-
Michal Hocko authored
task_will_free_mem is a misnomer for a more complex PF_EXITING test for early break out from the oom killer because it is believed that such a task would release its memory shortly and so we do not have to select an oom victim and perform a disruptive action. Currently we make sure that the given task is not participating in the core dumping because it might get blocked for a long time - see commit d003f371 ("oom: don't assume that a coredumping thread will exit soon"). The check can still do better though. We shouldn't consider the task unless the whole thread group is going down. This is rather unlikely but not impossible. A single exiting thread would surely leave all the address space behind. If we are really unlucky it might get stuck on the exit path and keep its TIF_MEMDIE and so block the oom killer. Link: http://lkml.kernel.org/r/1460452756-15491-1-git-send-email-mhocko@kernel.orgSigned-off-by:
-
Michal Hocko authored
Tetsuo has properly noted that mmput slow path might get blocked waiting for another party (e.g. exit_aio waits for an IO). If that happens the oom_reaper would be put out of the way and will not be able to process next oom victim. We should strive for making this context as reliable and independent on other subsystems as much as possible. Introduce mmput_async which will perform the slow path from an async (WQ) context. This will delay the operation but that shouldn't be a problem because the oom_reaper has reclaimed the victim's address space for most cases as much as possible and the remaining context shouldn't bind too much memory anymore. The only exception is when mmap_sem trylock has failed which shouldn't happen too often. The issue is only theoretical but not impossible. Signed-off-by:
-
Michal Hocko authored
Commit 36324a99 ("oom: clear TIF_MEMDIE after oom_reaper managed to unmap the address space") not only clears TIF_MEMDIE for oom reaped task but also set OOM_SCORE_ADJ_MIN for the target task to hide it from the oom killer. This works in simple cases but it is not sufficient for (unlikely) cases where the mm is shared between independent processes (as they do not share signal struct). If the mm had only small amount of memory which could be reaped then another task sharing the mm could be selected and that wouldn't help to move out from the oom situation. Introduce MMF_OOM_REAPED mm flag which is checked in oom_badness (same as OOM_SCORE_ADJ_MIN) and task is skipped if the flag is set. Set the flag after __oom_reap_task is done with a task. This will force the select_bad_process() to ignore all already oom reaped tasks as well as no such task is sacrificed for its parent. Signed-off-by:
-
Michal Hocko authored
Joonsoo has reported that he is able to trigger OOM for !costly high order requests (heavy fork() workload close the OOM) with the new oom detection rework. This is because we rely only on should_reclaim_retry when the compaction is disabled and it only checks watermarks for the requested order and so we might trigger OOM when there is a lot of free memory. It is not very clear what are the usual workloads when the compaction is disabled. Relying on high order allocations heavily without any mechanism to create those orders except for unbound amount of reclaim is certainly not a good idea. To prevent from potential regressions let's help this configuration some. We have to sacrifice the determinsm though because there simply is none here possible. should_compact_retry implementation for !CONFIG_COMPACTION, which was empty so far, will do watermark check for order-0 on all eligible zones. This will cause retrying until either the reclaim cannot make any further progress or all the zones are depleted even for order-0 pages. This means that the number of retries is basically unbounded for !costly orders but that was the case before the rework as well so this shouldn't regress. [akpm@linux-foundation.org: coding-style fixes] Link: http://lkml.kernel.org/r/1463051677-29418-3-git-send-email-mhocko@kernel.orgReported-by:
Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by:
Hillf Danton <hillf.zj@alibaba-inc.com> Signed-off-by:
-
Michal Hocko authored
"mm: consider compaction feedback also for costly allocation" has removed the upper bound for the reclaim/compaction retries based on the number of reclaimed pages for costly orders. While this is desirable the patch did miss a mis interaction between reclaim, compaction and the retry logic. The direct reclaim tries to get zones over min watermark while compaction backs off and returns COMPACT_SKIPPED when all zones are below low watermark + 1<<order gap. If we are getting really close to OOM then __compaction_suitable can keep returning COMPACT_SKIPPED a high order request (e.g. hugetlb order-9) while the reclaim is not able to release enough pages to get us over low watermark. The reclaim is still able to make some progress (usually trashing over few remaining pages) so we are not able to break out from the loop. I have seen this happening with the same test described in "mm: consider compaction feedback also for costly allocation" on a swapless system. The original problem got resolved by "vmscan: consider classzone_idx in compaction_ready" but it shows how things might go wrong when we approach the oom event horizont. The reason why compaction requires being over low rather than min watermark is not clear to me. This check was there essentially since 56de7263 ("mm: compaction: direct compact when a high-order allocation fails"). It is clearly an implementation detail though and we shouldn't pull it into the generic retry logic while we should be able to cope with such eventuality. The only place in should_compact_retry where we retry without any upper bound is for compaction_withdrawn() case. Introduce compaction_zonelist_suitable function which checks the given zonelist and returns true only if there is at least one zone which would would unblock __compaction_suitable if more memory got reclaimed. In this implementation it checks __compaction_suitable with NR_FREE_PAGES plus part of the reclaimable memory as the target for the watermark check. The reclaimable memory is reduced linearly by the allocation order. The idea is that we do not want to reclaim all the remaining memory for a single allocation request just unblock __compaction_suitable which doesn't guarantee we will make a further progress. The new helper is then used if compaction_withdrawn() feedback was provided so we do not retry if there is no outlook for a further progress. !costly requests shouldn't be affected much - e.g. order-2 pages would require to have at least 64kB on the reclaimable LRUs while order-9 would need at least 32M which should be enough to not lock up. [vbabka@suse.cz: fix classzone_idx vs. high_zoneidx usage in compaction_zonelist_suitable] [akpm@linux-foundation.org: fix it for Mel's mm-page_alloc-remove-field-from-alloc_context.patch] Signed-off-by:
Vlastimil Babka <vbabka@suse.cz> Cc: David Rientjes <rientjes@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Joonsoo Kim <js1304@gmail.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp> Cc: Vladimir Davydov <vdavydov@virtuozzo.com> Signed-off-by:
-
Michal Hocko authored
PAGE_ALLOC_COSTLY_ORDER retry logic is mostly handled inside should_reclaim_retry currently where we decide to not retry after at least order worth of pages were reclaimed or the watermark check for at least one zone would succeed after reclaiming all pages if the reclaim hasn't made any progress. Compaction feedback is mostly ignored and we just try to make sure that the compaction did at least something before giving up. The first condition was added by a41f24ea ("page allocator: smarter retry of costly-order allocations) and it assumed that lumpy reclaim could have created a page of the sufficient order. Lumpy reclaim, has been removed quite some time ago so the assumption doesn't hold anymore. Remove the check for the number of reclaimed pages and rely on the compaction feedback solely. should_reclaim_retry now only makes sure that we keep retrying reclaim for high order pages only if they are hidden by watermaks so order-0 reclaim makes really sense. should_compact_retry now keeps retrying even for the costly allocations. The number of retries is reduced wrt. !costly requests because they are less important and harder to grant and so their pressure shouldn't cause contention for other requests or cause an over reclaim. We also do not reset no_progress_loops for costly request to make sure we do not keep reclaiming too agressively. This has been tested by running a process which fragments memory: - compact memory - mmap large portion of the memory (1920M on 2GRAM machine with 2G of swapspace) - MADV_DONTNEED single page in PAGE_SIZE*((1UL<<MAX_ORDER)-1) steps until certain amount of memory is freed (250M in my test) and reduce the step to (step / 2) + 1 after reaching the end of the mapping - then run a script which populates the page cache 2G (MemTotal) from /dev/zero to a new file And then tries to allocate nr_hugepages=$(awk '/MemAvailable/{printf "%d\n", $2/(2*1024)}' /proc/meminfo) huge pages. root@test1:~# echo 1 > /proc/sys/vm/overcommit_memory;echo 1 > /proc/sys/vm/compact_memory; ./fragment-mem-and-run /root/alloc_hugepages.sh 1920M 250M Node 0, zone DMA 31 28 31 10 2 0 2 1 2 3 1 Node 0, zone DMA32 437 319 171 50 28 25 20 16 16 14 437 * This is the /proc/buddyinfo after the compaction Done fragmenting. size=2013265920 freed=262144000 Node 0, zone DMA 165 48 3 1 2 0 2 2 2 2 0 Node 0, zone DMA32 35109 14575 185 51 41 12 6 0 0 0 0 * /proc/buddyinfo after memory got fragmented Executing "/root/alloc_hugepages.sh" Eating some pagecache 508623+0 records in 508623+0 records out 2083319808 bytes (2.1 GB) copied, 11.7292 s, 178 MB/s Node 0, zone DMA 3 5 3 1 2 0 2 2 2 2 0 Node 0, zone DMA32 111 344 153 20 24 10 3 0 0 0 0 * /proc/buddyinfo after page cache got eaten Trying to allocate 129 129 * 129 hugepages requested and all of them granted. Node 0, zone DMA 3 5 3 1 2 0 2 2 2 2 0 Node 0, zone DMA32 127 97 30 99 11 6 2 1 4 0 0 * /proc/buddyinfo after hugetlb allocation. 10 runs will behave as follows: Trying to allocate 130 130 -- Trying to allocate 129 129 -- Trying to allocate 128 128 -- Trying to allocate 129 129 -- Trying to allocate 128 128 -- Trying to allocate 129 129 -- Trying to allocate 132 132 -- Trying to allocate 129 129 -- Trying to allocate 128 128 -- Trying to allocate 129 129 So basically 100% success for all 10 attempts. Without the patch numbers looked much worse: Trying to allocate 128 12 -- Trying to allocate 129 14 -- Trying to allocate 129 7 -- Trying to allocate 129 16 -- Trying to allocate 129 30 -- Trying to allocate 129 38 -- Trying to allocate 129 19 -- Trying to allocate 129 37 -- Trying to allocate 129 28 -- Trying to allocate 129 37 Just for completness the base kernel without oom detection rework looks as follows: Trying to allocate 127 30 -- Trying to allocate 129 12 -- Trying to allocate 129 52 -- Trying to allocate 128 32 -- Trying to allocate 129 12 -- Trying to allocate 129 10 -- Trying to allocate 129 32 -- Trying to allocate 128 14 -- Trying to allocate 128 16 -- Trying to allocate 129 8 As we can see the success rate is much more volatile and smaller without this patch. So the patch not only makes the retry logic for costly requests more sensible the success rate is even higher. Signed-off-by:
-
Michal Hocko authored
should_reclaim_retry will give up retries for higher order allocations if none of the eligible zones has any requested or higher order pages available even if we pass the watermak check for order-0. This is done because there is no guarantee that the reclaimable and currently free pages will form the required order. This can, however, lead to situations where the high-order request (e.g. order-2 required for the stack allocation during fork) will trigger OOM too early - e.g. after the first reclaim/compaction round. Such a system would have to be highly fragmented and there is no guarantee further reclaim/compaction attempts would help but at least make sure that the compaction was active before we go OOM and keep retrying even if should_reclaim_retry tells us to oom if - the last compaction round backed off or - we haven't completed at least MAX_COMPACT_RETRIES active compaction rounds. The first rule ensures that the very last attempt for compaction was not ignored while the second guarantees that the compaction has done some work. Multiple retries might be needed to prevent occasional pigggy backing of other contexts to steal the compacted pages before the current context manages to retry to allocate them. compaction_failed() is taken as a final word from the compaction that the retry doesn't make much sense. We have to be careful though because the first compaction round is MIGRATE_ASYNC which is rather weak as it ignores pages under writeback and gives up too easily in other situations. We therefore have to make sure that MIGRATE_SYNC_LIGHT mode has been used before we give up. With this logic in place we do not have to increase the migration mode unconditionally and rather do it only if the compaction failed for the weaker mode. A nice side effect is that the stronger migration mode is used only when really needed so this has a potential of smaller latencies in some cases. Please note that the compaction doesn't tell us much about how successful it was when returning compaction_made_progress so we just have to blindly trust that another retry is worthwhile and cap the number to something reasonable to guarantee a convergence. If the given number of successful retries is not sufficient for a reasonable workloads we should focus on the collected compaction tracepoints data and try to address the issue in the compaction code. If this is not feasible we can increase the retries limit. [mhocko@suse.com: fix warning] Link: http://lkml.kernel.org/r/20160512061636.GA4200@dhcp22.suse.czSigned-off-by:
-
Michal Hocko authored
wait_iff_congested has been used to throttle allocator before it retried another round of direct reclaim to allow the writeback to make some progress and prevent reclaim from looping over dirty/writeback pages without making any progress. We used to do congestion_wait before commit 0e093d99 ("writeback: do not sleep on the congestion queue if there are no congested BDIs or if significant congestion is not being encountered in the current zone") but that led to undesirable stalls and sleeping for the full timeout even when the BDI wasn't congested. Hence wait_iff_congested was used instead. But it seems that even wait_iff_congested doesn't work as expected. We might have a small file LRU list with all pages dirty/writeback and yet the bdi is not congested so this is just a cond_resched in the end and can end up triggering pre mature OOM. This patch replaces the unconditional wait_iff_congested by congestion_wait which is executed only if we _know_ that the last round of direct reclaim didn't make any progress and dirty+writeback pages are more than a half of the reclaimable pages on the zone which might be usable for our target allocation. This shouldn't reintroduce stalls fixed by 0e093d99 because congestion_wait is called only when we are getting hopeless when sleeping is a better choice than OOM with many pages under IO. We have to preserve logic introduced by commit 373ccbe5 ("mm, vmstat: allow WQ concurrency to discover memory reclaim doesn't make any progress") into the __alloc_pages_slowpath now that wait_iff_congested is not used anymore. As the only remaining user of wait_iff_congested is shrink_inactive_list we can remove the WQ specific short sleep from wait_iff_congested because the sleep is needed to be done only once in the allocation retry cycle. [mhocko@suse.com: high_zoneidx->ac_classzone_idx to evaluate memory reserves properly] Link: http://lkml.kernel.org/r/1463051677-29418-2-git-send-email-mhocko@kernel.orgSigned-off-by:
-
Michal Hocko authored
__alloc_pages_slowpath has traditionally relied on the direct reclaim and did_some_progress as an indicator that it makes sense to retry allocation rather than declaring OOM. shrink_zones had to rely on zone_reclaimable if shrink_zone didn't make any progress to prevent from a premature OOM killer invocation - the LRU might be full of dirty or writeback pages and direct reclaim cannot clean those up. zone_reclaimable allows to rescan the reclaimable lists several times and restart if a page is freed. This is really subtle behavior and it might lead to a livelock when a single freed page keeps allocator looping but the current task will not be able to allocate that single page. OOM killer would be more appropriate than looping without any progress for unbounded amount of time. This patch changes OOM detection logic and pulls it out from shrink_zone which is too low to be appropriate for any high level decisions such as OOM which is per zonelist property. It is __alloc_pages_slowpath which knows how many attempts have been done and what was the progress so far therefore it is more appropriate to implement this logic. The new heuristic is implemented in should_reclaim_retry helper called from __alloc_pages_slowpath. It tries to be more deterministic and easier to follow. It builds on an assumption that retrying makes sense only if the currently reclaimable memory + free pages would allow the current allocation request to succeed (as per __zone_watermark_ok) at least for one zone in the usable zonelist. This alone wouldn't be sufficient, though, because the writeback might get stuck and reclaimable pages might be pinned for a really long time or even depend on the current allocation context. Therefore there is a backoff mechanism implemented which reduces the reclaim target after each reclaim round without any progress. This means that we should eventually converge to only NR_FREE_PAGES as the target and fail on the wmark check and proceed to OOM. The backoff is simple and linear with 1/16 of the reclaimable pages for each round without any progress. We are optimistic and reset counter for successful reclaim rounds. Costly high order pages mostly preserve their semantic and those without __GFP_REPEAT fail right away while those which have the flag set will back off after the amount of reclaimable pages reaches equivalent of the requested order. The only difference is that if there was no progress during the reclaim we rely on zone watermark check. This is more logical thing to do than previous 1<<order attempts which were a result of zone_reclaimable faking the progress. [vdavydov@virtuozzo.com: check classzone_idx for shrink_zone] [hannes@cmpxchg.org: separate the heuristic into should_reclaim_retry] [rientjes@google.com: use zone_page_state_snapshot for NR_FREE_PAGES] [rientjes@google.com: shrink_zones doesn't need to return anything] Signed-off-by:
-
Michal Hocko authored
Compaction can provide a wild variation of feedback to the caller. Many of them are implementation specific and the caller of the compaction (especially the page allocator) shouldn't be bound to specifics of the current implementation. This patch abstracts the feedback into three basic types: - compaction_made_progress - compaction was active and made some progress. - compaction_failed - compaction failed and further attempts to invoke it would most probably fail and therefore it is not worth retrying - compaction_withdrawn - compaction wasn't invoked for an implementation specific reasons. In the current implementation it means that the compaction was deferred, contended or the page scanners met too early without any progress. Retrying is still worthwhile. [vbabka@suse.cz: do not change thp back off behavior] [akpm@linux-foundation.org: fix typo in comment, per Hillf] Signed-off-by: -
Michal Hocko authored
__alloc_pages_direct_compact communicates potential back off by two variables: - deferred_compaction tells that the compaction returned COMPACT_DEFERRED - contended_compaction is set when there is a contention on zone->lock resp. zone->lru_lock locks __alloc_pages_slowpath then backs of for THP allocation requests to prevent from long stalls. This is rather messy and it would be much cleaner to return a single compact result value and hide all the nasty details into __alloc_pages_direct_compact. This patch shouldn't introduce any functional changes. Signed-off-by:
-
Michal Hocko authored
compaction_result will be used as the primary feedback channel for compaction users. At the same time try_to_compact_pages (and potentially others) assume a certain ordering where a more specific feedback takes precendence. This gets a bit awkward when we have conflicting feedback from different zones. E.g one returing COMPACT_COMPLETE meaning the full zone has been scanned without any outcome while other returns with COMPACT_PARTIAL aka made some progress. The caller should get COMPACT_PARTIAL because that means that the compaction still can make some progress. The same applies for COMPACT_PARTIAL vs COMPACT_PARTIAL_SKIPPED. Reorder PARTIAL to be the largest one so the larger the value is the more progress we have done. Signed-off-by:
-
Michal Hocko authored
COMPACT_COMPLETE now means that compaction and free scanner met. This is not very useful information if somebody just wants to use this feedback and make any decisions based on that. The current caller might be a poor guy who just happened to scan tiny portion of the zone and that could be the reason no suitable pages were compacted. Make sure we distinguish the full and partial zone walks. Consumers should treat COMPACT_PARTIAL_SKIPPED as a potential success and be optimistic in retrying. The existing users of COMPACT_COMPLETE are conservatively changed to use COMPACT_PARTIAL_SKIPPED as well but some of them should be probably reconsidered and only defer the compaction only for COMPACT_COMPLETE with the new semantic. This patch shouldn't introduce any functional changes. Signed-off-by:
-
Michal Hocko authored
try_to_compact_pages() can currently return COMPACT_SKIPPED even when the compaction is defered for some zone just because zone DMA is skipped in 99% of cases due to watermark checks. This makes COMPACT_DEFERRED basically unusable for the page allocator as a feedback mechanism. Make sure we distinguish those two states properly and switch their ordering in the enum. This would mean that the COMPACT_SKIPPED will be returned only when all eligible zones are skipped. As a result COMPACT_DEFERRED handling for THP in __alloc_pages_slowpath will be more precise and we would bail out rather than reclaim. Signed-off-by:
-
Michal Hocko authored
The compiler is complaining after "mm, compaction: change COMPACT_ constants into enum" mm/compaction.c: In function `compact_zone': mm/compaction.c:1350:2: warning: enumeration value `COMPACT_DEFERRED' not handled in switch [-Wswitch] switch (ret) { ^ mm/compaction.c:1350:2: warning: enumeration value `COMPACT_COMPLETE' not handled in switch [-Wswitch] mm/compaction.c:1350:2: warning: enumeration value `COMPACT_NO_SUITABLE_PAGE' not handled in switch [-Wswitch] mm/compaction.c:1350:2: warning: enumeration value `COMPACT_NOT_SUITABLE_ZONE' not handled in switch [-Wswitch] mm/compaction.c:1350:2: warning: enumeration value `COMPACT_CONTENDED' not handled in switch [-Wswitch] compaction_suitable is allowed to return only COMPACT_PARTIAL, COMPACT_SKIPPED and COMPACT_CONTINUE so other cases are simply impossible. Put a VM_BUG_ON to catch an impossible return value. Signed-off-by: -
Michal Hocko authored
Compaction code is doing weird dances between COMPACT_FOO -> int -> unsigned long But there doesn't seem to be any reason for that. All functions which return/use one of those constants are not expecting any other value so it really makes sense to define an enum for them and make it clear that no other values are expected. This is a pure cleanup and shouldn't introduce any functional changes. Signed-off-by:
-
Michal Hocko authored
Motivation: As pointed out by Linus [2][3] relying on zone_reclaimable as a way to communicate the reclaim progress is rater dubious. I tend to agree, not only it is really obscure, it is not hard to imagine cases where a single page freed in the loop keeps all the reclaimers looping without getting any progress because their gfp_mask wouldn't allow to get that page anyway (e.g. single GFP_ATOMIC alloc and free loop). This is rather rare so it doesn't happen in the practice but the current logic which we have is rather obscure and hard to follow a also non-deterministic. This is an attempt to make the OOM detection more deterministic and easier to follow because each reclaimer basically tracks its own progress which is implemented at the page allocator layer rather spread out between the allocator and the reclaim. The more on the implementation is described in the first patch. I have tested several different scenarios but it should be clear that testing OOM killer is quite hard to be representative. There is usually a tiny gap between almost OOM and full blown OOM which is often time sensitive. Anyway, I have tested the following 2 scenarios and I would appreciate if there are more to test. Testing environment: a virtual machine with 2G of RAM and 2CPUs without any swap to make the OOM more deterministic. 1) 2 writers (each doing dd with 4M blocks to an xfs partition with 1G file size, removes the files and starts over again) running in parallel for 10s to build up a lot of dirty pages when 100 parallel mem_eaters (anon private populated mmap which waits until it gets signal) with 80M each. This causes an OOM flood of course and I have compared both patched and unpatched kernels. The test is considered finished after there are no OOM conditions detected. This should tell us whether there are any excessive kills or some of them premature (e.g. due to dirty pages): I have performed two runs this time each after a fresh boot. * base kernel $ grep "Out of memory:" base-oom-run1.log | wc -l 78 $ grep "Out of memory:" base-oom-run2.log | wc -l 78 $ grep "Kill process" base-oom-run1.log | tail -n1 [ 91.391203] Out of memory: Kill process 3061 (mem_eater) score 39 or sacrifice child $ grep "Kill process" base-oom-run2.log | tail -n1 [ 82.141919] Out of memory: Kill process 3086 (mem_eater) score 39 or sacrifice child $ grep "DMA32 free:" base-oom-run1.log | sed 's@.*free:\([0-9]*\)kB.*@\1@' | calc_min_max.awk min: 5376.00 max: 6776.00 avg: 5530.75 std: 166.50 nr: 61 $ grep "DMA32 free:" base-oom-run2.log | sed 's@.*free:\([0-9]*\)kB.*@\1@' | calc_min_max.awk min: 5416.00 max: 5608.00 avg: 5514.15 std: 42.94 nr: 52 $ grep "DMA32.*all_unreclaimable? no" base-oom-run1.log | wc -l 1 $ grep "DMA32.*all_unreclaimable? no" base-oom-run2.log | wc -l 3 * patched kernel $ grep "Out of memory:" patched-oom-run1.log | wc -l 78 miso@tiehlicka /mnt/share/devel/miso/kvm $ grep "Out of memory:" patched-oom-run2.log | wc -l 77 e grep "Kill process" patched-oom-run1.log | tail -n1 [ 497.317732] Out of memory: Kill process 3108 (mem_eater) score 39 or sacrifice child $ grep "Kill process" patched-oom-run2.log | tail -n1 [ 316.169920] Out of memory: Kill process 3093 (mem_eater) score 39 or sacrifice child $ grep "DMA32 free:" patched-oom-run1.log | sed 's@.*free:\([0-9]*\)kB.*@\1@' | calc_min_max.awk min: 5420.00 max: 5808.00 avg: 5513.90 std: 60.45 nr: 78 $ grep "DMA32 free:" patched-oom-run2.log | sed 's@.*free:\([0-9]*\)kB.*@\1@' | calc_min_max.awk min: 5380.00 max: 6384.00 avg: 5520.94 std: 136.84 nr: 77 e grep "DMA32.*all_unreclaimable? no" patched-oom-run1.log | wc -l 2 $ grep "DMA32.*all_unreclaimable? no" patched-oom-run2.log | wc -l 3 The patched kernel run noticeably longer while invoking OOM killer same number of times. This means that the original implementation is much more aggressive and triggers the OOM killer sooner. free pages stats show that neither kernels went OOM too early most of the time, though. I guess the difference is in the backoff when retries without any progress do sleep for a while if there is memory under writeback or dirty which is highly likely considering the parallel IO. Both kernels have seen races where zone wasn't marked unreclaimable and we still hit the OOM killer. This is most likely a race where a task managed to exit between the last allocation attempt and the oom killer invocation. 2) 2 writers again with 10s of run and then 10 mem_eaters to consume as much memory as possible without triggering the OOM killer. This required a lot of tuning but I've considered 3 consecutive runs in three different boots without OOM as a success. * base kernel size=$(awk '/MemFree/{printf "%dK", ($2/10)-(16*1024)}' /proc/meminfo) * patched kernel size=$(awk '/MemFree/{printf "%dK", ($2/10)-(12*1024)}' /proc/meminfo) That means 40M more memory was usable without triggering OOM killer. The base kernel sometimes managed to handle the same as patched but it wasn't consistent and failed in at least on of the 3 runs. This seems like a minor improvement. I was testing also GPF_REPEAT costly requests (hughetlb) with fragmented memory and under memory pressure. The results are in patch 11 where the logic is implemented. In short I can see huge improvement there. I am certainly interested in other usecases as well as well as any feedback. Especially those which require higher order requests. This patch (of 14): While playing with the oom detection rework [1] I have noticed that my heavy order-9 (hugetlb) load close to OOM ended up in an endless loop where the reclaim hasn't made any progress but did_some_progress didn't reflect that and compaction_suitable was backing off because no zone is above low wmark + 1 << order. It turned out that this is in fact an old standing bug in compaction_ready which ignores the requested_highidx and did the watermark check for 0 classzone_idx. This succeeds for zone DMA most of the time as the zone is mostly unused because of lowmem protection. As a result costly high order allocatios always report a successfull progress even when there was none. This wasn't a problem so far because these allocations usually fail quite early or retry only few times with __GFP_REPEAT but this will change after later patch in this series so make sure to not lie about the progress and propagate requested_highidx down to compaction_ready and use it for both the watermak check and compaction_suitable to fix this issue. [1] http://lkml.kernel.org/r/1459855533-4600-1-git-send-email-mhocko@kernel.org [2] https://lkml.org/lkml/2015/10/12/808 [3] https://lkml.org/lkml/2015/10/13/597Signed-off-by: -
Rik van Riel authored
The inactive file list should still be large enough to contain readahead windows and freshly written file data, but it no longer is the only source for detecting multiple accesses to file pages. The workingset refault measurement code causes recently evicted file pages that get accessed again after a shorter interval to be promoted directly to the active list. With that mechanism in place, we can afford to (on a larger system) dedicate more memory to the active file list, so we can actually cache more of the frequently used file pages in memory, and not have them pushed out by streaming writes, once-used streaming file reads, etc. This can help things like database workloads, where only half the page cache can currently be used to cache the database working set. This patch automatically increases that fraction on larger systems, using the same ratio that has already been used for anonymous memory. [hannes@cmpxchg.org: cgroup-awareness] Signed-off-by:
Rik van Riel <riel@redhat.com> Signed-off-by:
Andres Freund <andres@anarazel.de> Signed-off-by:
-
Johannes Weiner authored
Andres observed that his database workload is struggling with the transaction journal creating pressure on frequently read pages. Access patterns like transaction journals frequently write the same pages over and over, but in the majority of cases those pages are never read back. There are no caching benefits to be had for those pages, so activating them and having them put pressure on pages that do benefit from caching is a bad choice. Leave page activations to read accesses and don't promote pages based on writes alone. It could be said that partially written pages do contain cache-worthy data, because even if *userspace* does not access the unwritten part, the kernel still has to read it from the filesystem for correctness. However, a counter argument is that these pages enjoy at least *some* protection over other inactive file pages through the writeback cache, in the sense that dirty pages are written back with a delay and cache reclaim leaves them alone until they have been written back to disk. Should that turn out to be insufficient and we see increased read IO from partial writes under memory pressure, we can always go back and update grab_cache_page_write_begin() to take (pos, len) so that it can tell partial writes from pages that don't need partial reads. But for now, keep it simple. Signed-off-by:
-
Rik van Riel authored
This is a follow-up to http://www.spinics.net/lists/linux-mm/msg101739.html where Andres reported his database workingset being pushed out by the minimum size enforcement of the inactive file list - currently 50% of cache - as well as repeatedly written file pages that are never actually read. Two changes fell out of the discussions. The first change observes that pages that are only ever written don't benefit from caching beyond what the writeback cache does for partial page writes, and so we shouldn't promote them to the active file list where they compete with pages whose cached data is actually accessed repeatedly. This change comes in two patches - one for in-cache write accesses and one for refaults triggered by writes, neither of which should promote a cache page. Second, with the refault detection we don't need to set 50% of the cache aside for used-once cache anymore since we can detect frequently used pages even when they are evicted between accesses. We can allow the active list to be bigger and thus protect a bigger workingset that isn't challenged by streamers. Depending on the access patterns, this can increase major faults during workingset transitions for better performance during stable phases. This patch (of 3): When rewriting a page, the data in that page is replaced with new data. This means that evicting something else from the active file list, in order to cache data that will be replaced by something else, is likely to be a waste of memory. It is better to save the active list for frequently read pages, because reads actually use the data that is in the page. This patch ignores partial writes, because it is unclear whether the complexity of identifying those is worth any potential performance gain obtained from better caching pages that see repeated partial writes at large enough intervals to not get caught by the use-twice promotion code used for the inactive file list. Signed-off-by:
-
