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

   1 From a264df74df38855096393447f1b8f386069a94b9 Mon Sep 17 00:00:00 2001
   2 From: Roman Gushchin <guro@fb.com>
   3 Date: Wed, 4 Dec 2019 16:49:46 -0800
   4 Subject: mm: memcg/slab: wait for !root kmem_cache refcnt killing on root kmem_cache destruction
   5
   6 From: Roman Gushchin <guro@fb.com>
   7
   8 commit a264df74df38855096393447f1b8f386069a94b9 upstream.
   9
  10 Christian reported a warning like the following obtained during running
  11 some KVM-related tests on s390:
  12
  13     WARNING: CPU: 8 PID: 208 at lib/percpu-refcount.c:108 percpu_ref_exit+0x50/0x58
  14     Modules linked in: kvm(-) xt_CHECKSUM xt_MASQUERADE bonding xt_tcpudp ip6t_rpfilter ip6t_REJECT nf_reject_ipv6 ipt_REJECT nf_reject_ipv4 xt_conntrack ip6table_na>
  15     CPU: 8 PID: 208 Comm: kworker/8:1 Not tainted 5.2.0+ #66
  16     Hardware name: IBM 2964 NC9 712 (LPAR)
  17     Workqueue: events sysfs_slab_remove_workfn
  18     Krnl PSW : 0704e00180000000 0000001529746850 (percpu_ref_exit+0x50/0x58)
  19                R:0 T:1 IO:1 EX:1 Key:0 M:1 W:0 P:0 AS:3 CC:2 PM:0 RI:0 EA:3
  20     Krnl GPRS: 00000000ffff8808 0000001529746740 000003f4e30e8e18 0036008100000000
  21                0000001f00000000 0035008100000000 0000001fb3573ab8 0000000000000000
  22                0000001fbdb6de00 0000000000000000 0000001529f01328 0000001fb3573b00
  23                0000001fbb27e000 0000001fbdb69300 000003e009263d00 000003e009263cd0
  24     Krnl Code: 0000001529746842: f0a0000407fe        srp        4(11,%r0),2046,0
  25                0000001529746848: 47000700            bc         0,1792
  26               #000000152974684c: a7f40001            brc        15,152974684e
  27               >0000001529746850: a7f4fff2            brc        15,1529746834
  28                0000001529746854: 0707                bcr        0,%r7
  29                0000001529746856: 0707                bcr        0,%r7
  30                0000001529746858: eb8ff0580024        stmg       %r8,%r15,88(%r15)
  31                000000152974685e: a738ffff            lhi        %r3,-1
  32     Call Trace:
  33     ([<000003e009263d00>] 0x3e009263d00)
  34      [<00000015293252ea>] slab_kmem_cache_release+0x3a/0x70
  35      [<0000001529b04882>] kobject_put+0xaa/0xe8
  36      [<000000152918cf28>] process_one_work+0x1e8/0x428
  37      [<000000152918d1b0>] worker_thread+0x48/0x460
  38      [<00000015291942c6>] kthread+0x126/0x160
  39      [<0000001529b22344>] ret_from_fork+0x28/0x30
  40      [<0000001529b2234c>] kernel_thread_starter+0x0/0x10
  41     Last Breaking-Event-Address:
  42      [<000000152974684c>] percpu_ref_exit+0x4c/0x58
  43     ---[ end trace b035e7da5788eb09 ]---
  44
  45 The problem occurs because kmem_cache_destroy() is called immediately
  46 after deleting of a memcg, so it races with the memcg kmem_cache
  47 deactivation.
  48
  49 flush_memcg_workqueue() at the beginning of kmem_cache_destroy() is
  50 supposed to guarantee that all deactivation processes are finished, but
  51 failed to do so.  It waits for an rcu grace period, after which all
  52 children kmem_caches should be deactivated.  During the deactivation
  53 percpu_ref_kill() is called for non root kmem_cache refcounters, but it
  54 requires yet another rcu grace period to finish the transition to the
  55 atomic (dead) state.
  56
  57 So in a rare case when not all children kmem_caches are destroyed at the
  58 moment when the root kmem_cache is about to be gone, we need to wait
  59 another rcu grace period before destroying the root kmem_cache.
  60
  61 This issue can be triggered only with dynamically created kmem_caches
  62 which are used with memcg accounting.  In this case per-memcg child
  63 kmem_caches are created.  They are deactivated from the cgroup removing
  64 path.  If the destruction of the root kmem_cache is racing with the
  65 removal of the cgroup (both are quite complicated multi-stage
  66 processes), the described issue can occur.  The only known way to
  67 trigger it in the real life, is to unload some kernel module which
  68 creates a dedicated kmem_cache, used from different memory cgroups with
  69 GFP_ACCOUNT flag.  If the unloading happens immediately after calling
  70 rmdir on the corresponding cgroup, there is some chance to trigger the
  71 issue.
  72
  73 Link: http://lkml.kernel.org/r/20191129025011.3076017-1-guro@fb.com
  74 Fixes: f0a3a24b532d ("mm: memcg/slab: rework non-root kmem_cache lifecycle management")
  75 Signed-off-by: Roman Gushchin <guro@fb.com>
  76 Reported-by: Christian Borntraeger <borntraeger@de.ibm.com>
  77 Tested-by: Christian Borntraeger <borntraeger@de.ibm.com>
  78 Reviewed-by: Shakeel Butt <shakeelb@google.com>
  79 Acked-by: Michal Hocko <mhocko@suse.com>
  80 Cc: Johannes Weiner <hannes@cmpxchg.org>
  81 Cc: <stable@vger.kernel.org>
  82 Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
  83 Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
  84 Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
  85
  86 ---
  87  mm/slab_common.c |   12 ++++++++++++
  88  1 file changed, 12 insertions(+)
  89
  90 --- a/mm/slab_common.c
  91 +++ b/mm/slab_common.c
  92 @@ -904,6 +904,18 @@ static void flush_memcg_workqueue(struct
  93          * previous workitems on workqueue are processed.
  94          */
  95         flush_workqueue(memcg_kmem_cache_wq);
  96 +
  97 +       /*
  98 +        * If we're racing with children kmem_cache deactivation, it might
  99 +        * take another rcu grace period to complete their destruction.
 100 +        * At this moment the corresponding percpu_ref_kill() call should be
 101 +        * done, but it might take another rcu grace period to complete
 102 +        * switching to the atomic mode.
 103 +        * Please, note that we check without grabbing the slab_mutex. It's safe
 104 +        * because at this moment the children list can't grow.
 105 +        */
 106 +       if (!list_empty(&s->memcg_params.children))
 107 +               rcu_barrier();
 108  }
 109  #else
 110  static inline int shutdown_memcg_caches(struct kmem_cache *s)