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

   1 From b4c46484dc3fa3721d68fdfae85c1d7b1f6b5472 Mon Sep 17 00:00:00 2001
   2 From: Roman Gushchin <guro@fb.com>
   3 Date: Fri, 30 Aug 2019 16:04:39 -0700
   4 Subject: mm, memcg: partially revert "mm/memcontrol.c: keep local VM counters in sync with the hierarchical ones"
   5
   6 From: Roman Gushchin <guro@fb.com>
   7
   8 commit b4c46484dc3fa3721d68fdfae85c1d7b1f6b5472 upstream.
   9
  10 Commit 766a4c19d880 ("mm/memcontrol.c: keep local VM counters in sync
  11 with the hierarchical ones") effectively decreased the precision of
  12 per-memcg vmstats_local and per-memcg-per-node lruvec percpu counters.
  13
  14 That's good for displaying in memory.stat, but brings a serious
  15 regression into the reclaim process.
  16
  17 One issue I've discovered and debugged is the following:
  18 lruvec_lru_size() can return 0 instead of the actual number of pages in
  19 the lru list, preventing the kernel to reclaim last remaining pages.
  20 Result is yet another dying memory cgroups flooding.  The opposite is
  21 also happening: scanning an empty lru list is the waste of cpu time.
  22
  23 Also, inactive_list_is_low() can return incorrect values, preventing the
  24 active lru from being scanned and freed.  It can fail both because the
  25 size of active and inactive lists are inaccurate, and because the number
  26 of workingset refaults isn't precise.  In other words, the result is
  27 pretty random.
  28
  29 I'm not sure, if using the approximate number of slab pages in
  30 count_shadow_number() is acceptable, but issues described above are
  31 enough to partially revert the patch.
  32
  33 Let's keep per-memcg vmstat_local batched (they are only used for
  34 displaying stats to the userspace), but keep lruvec stats precise.  This
  35 change fixes the dead memcg flooding on my setup.
  36
  37 Link: http://lkml.kernel.org/r/20190817004726.2530670-1-guro@fb.com
  38 Fixes: 766a4c19d880 ("mm/memcontrol.c: keep local VM counters in sync with the hierarchical ones")
  39 Signed-off-by: Roman Gushchin <guro@fb.com>
  40 Acked-by: Yafang Shao <laoar.shao@gmail.com>
  41 Cc: Johannes Weiner <hannes@cmpxchg.org>
  42 Cc: Michal Hocko <mhocko@kernel.org>
  43 Cc: <stable@vger.kernel.org>
  44 Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
  45 Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
  46 Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
  47
  48 ---
  49  mm/memcontrol.c |    8 +++-----
  50  1 file changed, 3 insertions(+), 5 deletions(-)
  51
  52 --- a/mm/memcontrol.c
  53 +++ b/mm/memcontrol.c
  54 @@ -748,15 +748,13 @@ void __mod_lruvec_state(struct lruvec *l
  55         /* Update memcg */
  56         __mod_memcg_state(memcg, idx, val);
  57
  58 +       /* Update lruvec */
  59 +       __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
  60 +
  61         x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
  62         if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
  63                 struct mem_cgroup_per_node *pi;
  64
  65 -               /*
  66 -                * Batch local counters to keep them in sync with
  67 -                * the hierarchical ones.
  68 -                */
  69 -               __this_cpu_add(pn->lruvec_stat_local->count[idx], x);
  70                 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
  71                         atomic_long_add(x, &pi->lruvec_stat[idx]);
  72                 x = 0;