btrfs: honor path->skip_locking in backref code
commitd819d97ea025f8c32c12adef0ff55b2c7bf5c853
authorJosef Bacik <josef@toxicpanda.com>
Wed, 16 Jan 2019 16:00:57 +0000 (16 11:00 -0500)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Fri, 31 May 2019 13:47:12 +0000 (31 06:47 -0700)
tree550a43f65fd808624dae01721ad985761b6c4270
parent7c9290b56da477b54fab5dc48e1d21cfb8dc46f4
btrfs: honor path->skip_locking in backref code

commit 38e3eebff643db725633657d1d87a3be019d1018 upstream.

Qgroups will do the old roots lookup at delayed ref time, which could be
while walking down the extent root while running a delayed ref.  This
should be fine, except we specifically lock eb's in the backref walking
code irrespective of path->skip_locking, which deadlocks the system.
Fix up the backref code to honor path->skip_locking, nobody will be
modifying the commit_root when we're searching so it's completely safe
to do.

This happens since fb235dc06fac ("btrfs: qgroup: Move half of the qgroup
accounting time out of commit trans"), kernel may lockup with quota
enabled.

There is one backref trace triggered by snapshot dropping along with
write operation in the source subvolume.  The example can be reliably
reproduced:

  btrfs-cleaner   D    0  4062      2 0x80000000
  Call Trace:
   schedule+0x32/0x90
   btrfs_tree_read_lock+0x93/0x130 [btrfs]
   find_parent_nodes+0x29b/0x1170 [btrfs]
   btrfs_find_all_roots_safe+0xa8/0x120 [btrfs]
   btrfs_find_all_roots+0x57/0x70 [btrfs]
   btrfs_qgroup_trace_extent_post+0x37/0x70 [btrfs]
   btrfs_qgroup_trace_leaf_items+0x10b/0x140 [btrfs]
   btrfs_qgroup_trace_subtree+0xc8/0xe0 [btrfs]
   do_walk_down+0x541/0x5e3 [btrfs]
   walk_down_tree+0xab/0xe7 [btrfs]
   btrfs_drop_snapshot+0x356/0x71a [btrfs]
   btrfs_clean_one_deleted_snapshot+0xb8/0xf0 [btrfs]
   cleaner_kthread+0x12b/0x160 [btrfs]
   kthread+0x112/0x130
   ret_from_fork+0x27/0x50

When dropping snapshots with qgroup enabled, we will trigger backref
walk.

However such backref walk at that timing is pretty dangerous, as if one
of the parent nodes get WRITE locked by other thread, we could cause a
dead lock.

For example:

           FS 260     FS 261 (Dropped)
            node A        node B
           /      \      /      \
       node C      node D      node E
      /   \         /  \        /     \
  leaf F|leaf G|leaf H|leaf I|leaf J|leaf K

The lock sequence would be:

      Thread A (cleaner)             |       Thread B (other writer)
-----------------------------------------------------------------------
write_lock(B)                        |
write_lock(D)                        |
^^^ called by walk_down_tree()       |
                                     |       write_lock(A)
                                     |       write_lock(D) << Stall
read_lock(H) << for backref walk     |
read_lock(D) << lock owner is        |
                the same thread A    |
                so read lock is OK   |
read_lock(A) << Stall                |

So thread A hold write lock D, and needs read lock A to unlock.
While thread B holds write lock A, while needs lock D to unlock.

This will cause a deadlock.

This is not only limited to snapshot dropping case.  As the backref
walk, even only happens on commit trees, is breaking the normal top-down
locking order, makes it deadlock prone.

Fixes: fb235dc06fac ("btrfs: qgroup: Move half of the qgroup accounting time out of commit trans")
CC: stable@vger.kernel.org # 4.14+
Reported-and-tested-by: David Sterba <dsterba@suse.com>
Reported-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: Filipe Manana <fdmanana@suse.com>
[ rebase to latest branch and fix lock assert bug in btrfs/007 ]
[ backport to linux-4.19.y branch, solve minor conflicts ]
Signed-off-by: Qu Wenruo <wqu@suse.com>
[ copy logs and deadlock analysis from Qu's patch ]
Signed-off-by: David Sterba <dsterba@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
fs/btrfs/backref.c