lockdep: fix fs_reclaim warning
authorTetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Thu, 22 Mar 2018 23:17:10 +0000 (16:17 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Fri, 23 Mar 2018 00:07:01 +0000 (17:07 -0700)
Dave Jones reported fs_reclaim lockdep warnings.

  ============================================
  WARNING: possible recursive locking detected
  4.15.0-rc9-backup-debug+ #1 Not tainted
  --------------------------------------------
  sshd/24800 is trying to acquire lock:
   (fs_reclaim){+.+.}, at: [<0000000084f438c2>] fs_reclaim_acquire.part.102+0x5/0x30

  but task is already holding lock:
   (fs_reclaim){+.+.}, at: [<0000000084f438c2>] fs_reclaim_acquire.part.102+0x5/0x30

  other info that might help us debug this:
   Possible unsafe locking scenario:

         CPU0
         ----
    lock(fs_reclaim);
    lock(fs_reclaim);

   *** DEADLOCK ***

   May be due to missing lock nesting notation

  2 locks held by sshd/24800:
   #0:  (sk_lock-AF_INET6){+.+.}, at: [<000000001a069652>] tcp_sendmsg+0x19/0x40
   #1:  (fs_reclaim){+.+.}, at: [<0000000084f438c2>] fs_reclaim_acquire.part.102+0x5/0x30

  stack backtrace:
  CPU: 3 PID: 24800 Comm: sshd Not tainted 4.15.0-rc9-backup-debug+ #1
  Call Trace:
   dump_stack+0xbc/0x13f
   __lock_acquire+0xa09/0x2040
   lock_acquire+0x12e/0x350
   fs_reclaim_acquire.part.102+0x29/0x30
   kmem_cache_alloc+0x3d/0x2c0
   alloc_extent_state+0xa7/0x410
   __clear_extent_bit+0x3ea/0x570
   try_release_extent_mapping+0x21a/0x260
   __btrfs_releasepage+0xb0/0x1c0
   btrfs_releasepage+0x161/0x170
   try_to_release_page+0x162/0x1c0
   shrink_page_list+0x1d5a/0x2fb0
   shrink_inactive_list+0x451/0x940
   shrink_node_memcg.constprop.88+0x4c9/0x5e0
   shrink_node+0x12d/0x260
   try_to_free_pages+0x418/0xaf0
   __alloc_pages_slowpath+0x976/0x1790
   __alloc_pages_nodemask+0x52c/0x5c0
   new_slab+0x374/0x3f0
   ___slab_alloc.constprop.81+0x47e/0x5a0
   __slab_alloc.constprop.80+0x32/0x60
   __kmalloc_track_caller+0x267/0x310
   __kmalloc_reserve.isra.40+0x29/0x80
   __alloc_skb+0xee/0x390
   sk_stream_alloc_skb+0xb8/0x340
   tcp_sendmsg_locked+0x8e6/0x1d30
   tcp_sendmsg+0x27/0x40
   inet_sendmsg+0xd0/0x310
   sock_write_iter+0x17a/0x240
   __vfs_write+0x2ab/0x380
   vfs_write+0xfb/0x260
   SyS_write+0xb6/0x140
   do_syscall_64+0x1e5/0xc05
   entry_SYSCALL64_slow_path+0x25/0x25

This warning is caused by commit d92a8cfcb37e ("locking/lockdep:
Rework FS_RECLAIM annotation") which replaced the use of
lockdep_{set,clear}_current_reclaim_state() in __perform_reclaim()
and lockdep_trace_alloc() in slab_pre_alloc_hook() with
fs_reclaim_acquire()/ fs_reclaim_release().

Since __kmalloc_reserve() from __alloc_skb() adds __GFP_NOMEMALLOC |
__GFP_NOWARN to gfp_mask, and all reclaim path simply propagates
__GFP_NOMEMALLOC, fs_reclaim_acquire() in slab_pre_alloc_hook() is
trying to grab the 'fake' lock again when __perform_reclaim() already
grabbed the 'fake' lock.

The

  /* this guy won't enter reclaim */
  if ((current->flags & PF_MEMALLOC) && !(gfp_mask & __GFP_NOMEMALLOC))
          return false;

test which causes slab_pre_alloc_hook() to try to grab the 'fake' lock
was added by commit cf40bd16fdad ("lockdep: annotate reclaim context
(__GFP_NOFS)").  But that test is outdated because PF_MEMALLOC thread
won't enter reclaim regardless of __GFP_NOMEMALLOC after commit
341ce06f69ab ("page allocator: calculate the alloc_flags for allocation
only once") added the PF_MEMALLOC safeguard (

  /* Avoid recursion of direct reclaim */
  if (p->flags & PF_MEMALLOC)
          goto nopage;

in __alloc_pages_slowpath()).

Thus, let's fix outdated test by removing __GFP_NOMEMALLOC test and
allow __need_fs_reclaim() to return false.

Link: http://lkml.kernel.org/r/201802280650.FJC73911.FOSOMLJVFFQtHO@I-love.SAKURA.ne.jp
Fixes: d92a8cfcb37ecd13 ("locking/lockdep: Rework FS_RECLAIM annotation")
Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Reported-by: Dave Jones <davej@codemonkey.org.uk>
Tested-by: Dave Jones <davej@codemonkey.org.uk>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Nick Piggin <npiggin@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Nikolay Borisov <nborisov@suse.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: <stable@vger.kernel.org> [4.14+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/page_alloc.c

index 635d7dd..010dee0 100644 (file)
@@ -3596,7 +3596,7 @@ static bool __need_fs_reclaim(gfp_t gfp_mask)
                return false;
 
        /* this guy won't enter reclaim */
-       if ((current->flags & PF_MEMALLOC) && !(gfp_mask & __GFP_NOMEMALLOC))
+       if (current->flags & PF_MEMALLOC)
                return false;
 
        /* We're only interested __GFP_FS allocations for now */