git.ipfire.org Git - thirdparty/kernel/stable-queue.git/blob

   1 From 6aaced5abd32e2a57cd94fd64f824514d0361da8 Mon Sep 17 00:00:00 2001
   2 From: Seiji Nishikawa <snishika@redhat.com>
   3 Date: Sun, 1 Dec 2024 01:12:34 +0900
   4 Subject: mm: vmscan: account for free pages to prevent infinite Loop in throttle_direct_reclaim()
   5
   6 From: Seiji Nishikawa <snishika@redhat.com>
   7
   8 commit 6aaced5abd32e2a57cd94fd64f824514d0361da8 upstream.
   9
  10 The task sometimes continues looping in throttle_direct_reclaim() because
  11 allow_direct_reclaim(pgdat) keeps returning false.
  12
  13  #0 [ffff80002cb6f8d0] __switch_to at ffff8000080095ac
  14  #1 [ffff80002cb6f900] __schedule at ffff800008abbd1c
  15  #2 [ffff80002cb6f990] schedule at ffff800008abc50c
  16  #3 [ffff80002cb6f9b0] throttle_direct_reclaim at ffff800008273550
  17  #4 [ffff80002cb6fa20] try_to_free_pages at ffff800008277b68
  18  #5 [ffff80002cb6fae0] __alloc_pages_nodemask at ffff8000082c4660
  19  #6 [ffff80002cb6fc50] alloc_pages_vma at ffff8000082e4a98
  20  #7 [ffff80002cb6fca0] do_anonymous_page at ffff80000829f5a8
  21  #8 [ffff80002cb6fce0] __handle_mm_fault at ffff8000082a5974
  22  #9 [ffff80002cb6fd90] handle_mm_fault at ffff8000082a5bd4
  23
  24 At this point, the pgdat contains the following two zones:
  25
  26         NODE: 4  ZONE: 0  ADDR: ffff00817fffe540  NAME: "DMA32"
  27           SIZE: 20480  MIN/LOW/HIGH: 11/28/45
  28           VM_STAT:
  29                 NR_FREE_PAGES: 359
  30         NR_ZONE_INACTIVE_ANON: 18813
  31           NR_ZONE_ACTIVE_ANON: 0
  32         NR_ZONE_INACTIVE_FILE: 50
  33           NR_ZONE_ACTIVE_FILE: 0
  34           NR_ZONE_UNEVICTABLE: 0
  35         NR_ZONE_WRITE_PENDING: 0
  36                      NR_MLOCK: 0
  37                     NR_BOUNCE: 0
  38                    NR_ZSPAGES: 0
  39             NR_FREE_CMA_PAGES: 0
  40
  41         NODE: 4  ZONE: 1  ADDR: ffff00817fffec00  NAME: "Normal"
  42           SIZE: 8454144  PRESENT: 98304  MIN/LOW/HIGH: 68/166/264
  43           VM_STAT:
  44                 NR_FREE_PAGES: 146
  45         NR_ZONE_INACTIVE_ANON: 94668
  46           NR_ZONE_ACTIVE_ANON: 3
  47         NR_ZONE_INACTIVE_FILE: 735
  48           NR_ZONE_ACTIVE_FILE: 78
  49           NR_ZONE_UNEVICTABLE: 0
  50         NR_ZONE_WRITE_PENDING: 0
  51                      NR_MLOCK: 0
  52                     NR_BOUNCE: 0
  53                    NR_ZSPAGES: 0
  54             NR_FREE_CMA_PAGES: 0
  55
  56 In allow_direct_reclaim(), while processing ZONE_DMA32, the sum of
  57 inactive/active file-backed pages calculated in zone_reclaimable_pages()
  58 based on the result of zone_page_state_snapshot() is zero.
  59
  60 Additionally, since this system lacks swap, the calculation of inactive/
  61 active anonymous pages is skipped.
  62
  63         crash> p nr_swap_pages
  64         nr_swap_pages = $1937 = {
  65           counter = 0
  66         }
  67
  68 As a result, ZONE_DMA32 is deemed unreclaimable and skipped, moving on to
  69 the processing of the next zone, ZONE_NORMAL, despite ZONE_DMA32 having
  70 free pages significantly exceeding the high watermark.
  71
  72 The problem is that the pgdat->kswapd_failures hasn't been incremented.
  73
  74         crash> px ((struct pglist_data *) 0xffff00817fffe540)->kswapd_failures
  75         $1935 = 0x0
  76
  77 This is because the node deemed balanced.  The node balancing logic in
  78 balance_pgdat() evaluates all zones collectively.  If one or more zones
  79 (e.g., ZONE_DMA32) have enough free pages to meet their watermarks, the
  80 entire node is deemed balanced.  This causes balance_pgdat() to exit early
  81 before incrementing the kswapd_failures, as it considers the overall
  82 memory state acceptable, even though some zones (like ZONE_NORMAL) remain
  83 under significant pressure.
  84
  85
  86 The patch ensures that zone_reclaimable_pages() includes free pages
  87 (NR_FREE_PAGES) in its calculation when no other reclaimable pages are
  88 available (e.g., file-backed or anonymous pages).  This change prevents
  89 zones like ZONE_DMA32, which have sufficient free pages, from being
  90 mistakenly deemed unreclaimable.  By doing so, the patch ensures proper
  91 node balancing, avoids masking pressure on other zones like ZONE_NORMAL,
  92 and prevents infinite loops in throttle_direct_reclaim() caused by
  93 allow_direct_reclaim(pgdat) repeatedly returning false.
  94
  95
  96 The kernel hangs due to a task stuck in throttle_direct_reclaim(), caused
  97 by a node being incorrectly deemed balanced despite pressure in certain
  98 zones, such as ZONE_NORMAL.  This issue arises from
  99 zone_reclaimable_pages() returning 0 for zones without reclaimable file-
 100 backed or anonymous pages, causing zones like ZONE_DMA32 with sufficient
 101 free pages to be skipped.
 102
 103 The lack of swap or reclaimable pages results in ZONE_DMA32 being ignored
 104 during reclaim, masking pressure in other zones.  Consequently,
 105 pgdat->kswapd_failures remains 0 in balance_pgdat(), preventing fallback
 106 mechanisms in allow_direct_reclaim() from being triggered, leading to an
 107 infinite loop in throttle_direct_reclaim().
 108
 109 This patch modifies zone_reclaimable_pages() to account for free pages
 110 (NR_FREE_PAGES) when no other reclaimable pages exist.  This ensures zones
 111 with sufficient free pages are not skipped, enabling proper balancing and
 112 reclaim behavior.
 113
 114 [akpm@linux-foundation.org: coding-style cleanups]
 115 Link: https://lkml.kernel.org/r/20241130164346.436469-1-snishika@redhat.com
 116 Link: https://lkml.kernel.org/r/20241130161236.433747-2-snishika@redhat.com
 117 Fixes: 5a1c84b404a7 ("mm: remove reclaim and compaction retry approximations")
 118 Signed-off-by: Seiji Nishikawa <snishika@redhat.com>
 119 Cc: Mel Gorman <mgorman@techsingularity.net>
 120 Cc: <stable@vger.kernel.org>
 121 Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
 122 Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
 123 ---
 124  mm/vmscan.c |    9 ++++++++-
 125  1 file changed, 8 insertions(+), 1 deletion(-)
 126
 127 --- a/mm/vmscan.c
 128 +++ b/mm/vmscan.c
 129 @@ -374,7 +374,14 @@ unsigned long zone_reclaimable_pages(str
 130         if (can_reclaim_anon_pages(NULL, zone_to_nid(zone), NULL))
 131                 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
 132                         zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
 133 -
 134 +       /*
 135 +        * If there are no reclaimable file-backed or anonymous pages,
 136 +        * ensure zones with sufficient free pages are not skipped.
 137 +        * This prevents zones like DMA32 from being ignored in reclaim
 138 +        * scenarios where they can still help alleviate memory pressure.
 139 +        */
 140 +       if (nr == 0)
 141 +               nr = zone_page_state_snapshot(zone, NR_FREE_PAGES);
 142         return nr;
 143  }
 144