mm: don't steal highatomic pageblock
authorMinchan Kim <minchan@kernel.org>
Tue, 13 Dec 2016 00:42:05 +0000 (16:42 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Tue, 13 Dec 2016 02:55:07 +0000 (18:55 -0800)
commit88ed365ea227aa28841a8d6e196c9a261c76fffd
treee90f853e6806a55f21ffc037e8c0b82d6fe145ac
parent22901c6c9f93058c3803d343db02c14e870e3545
mm: don't steal highatomic pageblock

Patch series "use up highorder free pages before OOM", v3.

I got OOM report from production team with v4.4 kernel.  It had enough
free memory but failed to allocate GFP_KERNEL order-0 page and finally
encountered OOM kill.  It occured during QA process which launches
several apps, switching and so on.  It happned rarely.  IOW, In normal
situation, it was not a problem but if we are unluck so that several
apps uses peak memory at the same time, it can happen.  If we manage to
pass the phase, the system can go working well.

I could reproduce it with my test(memory spike easily.  Look at below.

The reason is free pages(19M) of DMA32 zone are reserved for
HIGHORDERATOMIC and doesn't unreserved before the OOM.

  balloon invoked oom-killer: gfp_mask=0x24280ca(GFP_HIGHUSER_MOVABLE|__GFP_ZERO), order=0, oom_score_adj=0
  balloon cpuset=/ mems_allowed=0
  CPU: 1 PID: 8473 Comm: balloon Tainted: G        W  OE   4.8.0-rc7-00219-g3f74c9559583-dirty #3161
  Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Ubuntu-1.8.2-1ubuntu1 04/01/2014
  Call Trace:
    dump_stack+0x63/0x90
    dump_header+0x5c/0x1ce
    oom_kill_process+0x22e/0x400
    out_of_memory+0x1ac/0x210
    __alloc_pages_nodemask+0x101e/0x1040
    handle_mm_fault+0xa0a/0xbf0
    __do_page_fault+0x1dd/0x4d0
    trace_do_page_fault+0x43/0x130
    do_async_page_fault+0x1a/0xa0
    async_page_fault+0x28/0x30
  Mem-Info:
  active_anon:383949 inactive_anon:106724 isolated_anon:0
   active_file:15 inactive_file:44 isolated_file:0
   unevictable:0 dirty:0 writeback:24 unstable:0
   slab_reclaimable:2483 slab_unreclaimable:3326
   mapped:0 shmem:0 pagetables:1906 bounce:0
   free:6898 free_pcp:291 free_cma:0
  Node 0 active_anon:1535796kB inactive_anon:426896kB active_file:60kB inactive_file:176kB unevictable:0kB isolated(anon):0kB isolated(file):0kB mapped:0kB dirty:0kB writeback:96kB shmem:0kB writeback_tmp:0kB unstable:0kB pages_scanned:1418 all_unreclaimable? no
  DMA free:8188kB min:44kB low:56kB high:68kB active_anon:7648kB inactive_anon:0kB active_file:0kB inactive_file:4kB unevictable:0kB writepending:0kB present:15992kB managed:15908kB mlocked:0kB slab_reclaimable:0kB slab_unreclaimable:20kB kernel_stack:0kB pagetables:0kB bounce:0kB free_pcp:0kB local_pcp:0kB free_cma:0kB
  lowmem_reserve[]: 0 1952 1952 1952
  DMA32 free:19404kB min:5628kB low:7624kB high:9620kB active_anon:1528148kB inactive_anon:426896kB active_file:60kB inactive_file:420kB unevictable:0kB writepending:96kB present:2080640kB managed:2030092kB mlocked:0kB slab_reclaimable:9932kB slab_unreclaimable:13284kB kernel_stack:2496kB pagetables:7624kB bounce:0kB free_pcp:900kB local_pcp:112kB free_cma:0kB
  lowmem_reserve[]: 0 0 0 0
  DMA: 0*4kB 0*8kB 0*16kB 0*32kB 0*64kB 0*128kB 0*256kB 0*512kB 0*1024kB 0*2048kB 2*4096kB (H) = 8192kB
  DMA32: 7*4kB (H) 8*8kB (H) 30*16kB (H) 31*32kB (H) 14*64kB (H) 9*128kB (H) 2*256kB (H) 2*512kB (H) 4*1024kB (H) 5*2048kB (H) 0*4096kB = 19484kB
  51131 total pagecache pages
  50795 pages in swap cache
  Swap cache stats: add 3532405601, delete 3532354806, find 124289150/1822712228
  Free swap  = 8kB
  Total swap = 255996kB
  524158 pages RAM
  0 pages HighMem/MovableOnly
  12658 pages reserved
  0 pages cma reserved
  0 pages hwpoisoned

Another example exceeded the limit by the race is

  in:imklog: page allocation failure: order:0, mode:0x2280020(GFP_ATOMIC|__GFP_NOTRACK)
  CPU: 0 PID: 476 Comm: in:imklog Tainted: G            E   4.8.0-rc7-00217-g266ef83c51e5-dirty #3135
  Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Ubuntu-1.8.2-1ubuntu1 04/01/2014
  Call Trace:
    dump_stack+0x63/0x90
    warn_alloc_failed+0xdb/0x130
    __alloc_pages_nodemask+0x4d6/0xdb0
    new_slab+0x339/0x490
    ___slab_alloc.constprop.74+0x367/0x480
    __slab_alloc.constprop.73+0x20/0x40
    __kmalloc+0x1a4/0x1e0
    alloc_indirect.isra.14+0x1d/0x50
    virtqueue_add_sgs+0x1c4/0x470
    __virtblk_add_req+0xae/0x1f0
    virtio_queue_rq+0x12d/0x290
    __blk_mq_run_hw_queue+0x239/0x370
    blk_mq_run_hw_queue+0x8f/0xb0
    blk_mq_insert_requests+0x18c/0x1a0
    blk_mq_flush_plug_list+0x125/0x140
    blk_flush_plug_list+0xc7/0x220
    blk_finish_plug+0x2c/0x40
    __do_page_cache_readahead+0x196/0x230
    filemap_fault+0x448/0x4f0
    ext4_filemap_fault+0x36/0x50
    __do_fault+0x75/0x140
    handle_mm_fault+0x84d/0xbe0
    __do_page_fault+0x1dd/0x4d0
    trace_do_page_fault+0x43/0x130
    do_async_page_fault+0x1a/0xa0
    async_page_fault+0x28/0x30
  Mem-Info:
  active_anon:363826 inactive_anon:121283 isolated_anon:32
   active_file:65 inactive_file:152 isolated_file:0
   unevictable:0 dirty:0 writeback:46 unstable:0
   slab_reclaimable:2778 slab_unreclaimable:3070
   mapped:112 shmem:0 pagetables:1822 bounce:0
   free:9469 free_pcp:231 free_cma:0
  Node 0 active_anon:1455304kB inactive_anon:485132kB active_file:260kB inactive_file:608kB unevictable:0kB isolated(anon):128kB isolated(file):0kB mapped:448kB dirty:0kB writeback:184kB shmem:0kB writeback_tmp:0kB unstable:0kB pages_scanned:13641 all_unreclaimable? no
  DMA free:7748kB min:44kB low:56kB high:68kB active_anon:7944kB inactive_anon:104kB active_file:0kB inactive_file:0kB unevictable:0kB writepending:0kB present:15992kB managed:15908kB mlocked:0kB slab_reclaimable:0kB slab_unreclaimable:108kB kernel_stack:0kB pagetables:4kB bounce:0kB free_pcp:0kB local_pcp:0kB free_cma:0kB
  lowmem_reserve[]: 0 1952 1952 1952
  DMA32 free:30128kB min:5628kB low:7624kB high:9620kB active_anon:1447360kB inactive_anon:485028kB active_file:260kB inactive_file:608kB unevictable:0kB writepending:184kB present:2080640kB managed:2030132kB mlocked:0kB slab_reclaimable:11112kB slab_unreclaimable:12172kB kernel_stack:2400kB pagetables:7284kB bounce:0kB free_pcp:924kB local_pcp:72kB free_cma:0kB
  lowmem_reserve[]: 0 0 0 0
  DMA: 7*4kB (UE) 3*8kB (UH) 1*16kB (M) 0*32kB 2*64kB (U) 1*128kB (M) 1*256kB (U) 0*512kB 1*1024kB (U) 1*2048kB (U) 1*4096kB (H) = 7748kB
  DMA32: 10*4kB (H) 3*8kB (H) 47*16kB (H) 38*32kB (H) 5*64kB (H) 1*128kB (H) 2*256kB (H) 3*512kB (H) 3*1024kB (H) 3*2048kB (H) 4*4096kB (H) = 30128kB
  2775 total pagecache pages
  2536 pages in swap cache
  Swap cache stats: add 206786828, delete 206784292, find 7323106/106686077
  Free swap  = 108744kB
  Total swap = 255996kB
  524158 pages RAM
  0 pages HighMem/MovableOnly
  12648 pages reserved
  0 pages cma reserved
  0 pages hwpoisoned

During the investigation, I found some problems with highatomic so this
patch aims to solve the problems and the final goal is to unreserve
every highatomic free pages before the OOM kill.

This patch (of 4):

In page freeing path, migratetype is racy so that a highorderatomic page
could free into non-highorderatomic free list.  If that page is
allocated, VM can change the pageblock from higorderatomic to something.
In that case, highatomic pageblock accounting is broken so it doesn't
work(e.g., VM cannot reserve highorderatomic pageblocks any more
although it doesn't reach 1% limit).

So, this patch prohibits the changing from highatomic to other type.
It's no problem because MIGRATE_HIGHATOMIC is not listed in fallback
array so stealing will only happen due to unexpected races which is
really rare.  Also, such prohibiting keeps highatomic pageblock more
longer so it would be better for highorderatomic page allocation.

Link: http://lkml.kernel.org/r/1476259429-18279-2-git-send-email-minchan@kernel.org
Signed-off-by: Minchan Kim <minchan@kernel.org>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Sangseok Lee <sangseok.lee@lge.com>
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/page_alloc.c