Btrfs: fix deadlock when writing out free space caches
authorFilipe Manana <fdmanana@suse.com>
Fri, 12 Oct 2018 09:03:55 +0000 (10:03 +0100)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Tue, 13 Nov 2018 19:08:58 +0000 (11:08 -0800)
commitea9c846f54dbf03da159a1b7566aa95e9bf1674b
tree6b0451cb0c3becb73f518a08cc2f81812ffd88de
parente17af96eebd39a53d467c3beb6f946f64f5c3637
Btrfs: fix deadlock when writing out free space caches

commit 5ce555578e0919237fa4bda92b4670e2dd176f85 upstream.

When writing out a block group free space cache we can end deadlocking
with ourselves on an extent buffer lock resulting in a warning like the
following:

  [245043.379979] WARNING: CPU: 4 PID: 2608 at fs/btrfs/locking.c:251 btrfs_tree_lock+0x1be/0x1d0 [btrfs]
  [245043.392792] CPU: 4 PID: 2608 Comm: btrfs-transacti Tainted: G
    W I      4.16.8 #1
  [245043.395489] RIP: 0010:btrfs_tree_lock+0x1be/0x1d0 [btrfs]
  [245043.396791] RSP: 0018:ffffc9000424b840 EFLAGS: 00010246
  [245043.398093] RAX: 0000000000000a30 RBX: ffff8807e20a3d20 RCX: 0000000000000001
  [245043.399414] RDX: 0000000000000001 RSI: 0000000000000002 RDI: ffff8807e20a3d20
  [245043.400732] RBP: 0000000000000001 R08: ffff88041f39a700 R09: ffff880000000000
  [245043.402021] R10: 0000000000000040 R11: ffff8807e20a3d20 R12: ffff8807cb220630
  [245043.403296] R13: 0000000000000001 R14: ffff8807cb220628 R15: ffff88041fbdf000
  [245043.404780] FS:  0000000000000000(0000) GS:ffff88082fc80000(0000) knlGS:0000000000000000
  [245043.406050] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
  [245043.407321] CR2: 00007fffdbdb9f10 CR3: 0000000001c09005 CR4: 00000000000206e0
  [245043.408670] Call Trace:
  [245043.409977]  btrfs_search_slot+0x761/0xa60 [btrfs]
  [245043.411278]  btrfs_insert_empty_items+0x62/0xb0 [btrfs]
  [245043.412572]  btrfs_insert_item+0x5b/0xc0 [btrfs]
  [245043.413922]  btrfs_create_pending_block_groups+0xfb/0x1e0 [btrfs]
  [245043.415216]  do_chunk_alloc+0x1e5/0x2a0 [btrfs]
  [245043.416487]  find_free_extent+0xcd0/0xf60 [btrfs]
  [245043.417813]  btrfs_reserve_extent+0x96/0x1e0 [btrfs]
  [245043.419105]  btrfs_alloc_tree_block+0xfb/0x4a0 [btrfs]
  [245043.420378]  __btrfs_cow_block+0x127/0x550 [btrfs]
  [245043.421652]  btrfs_cow_block+0xee/0x190 [btrfs]
  [245043.422979]  btrfs_search_slot+0x227/0xa60 [btrfs]
  [245043.424279]  ? btrfs_update_inode_item+0x59/0x100 [btrfs]
  [245043.425538]  ? iput+0x72/0x1e0
  [245043.426798]  write_one_cache_group.isra.49+0x20/0x90 [btrfs]
  [245043.428131]  btrfs_start_dirty_block_groups+0x102/0x420 [btrfs]
  [245043.429419]  btrfs_commit_transaction+0x11b/0x880 [btrfs]
  [245043.430712]  ? start_transaction+0x8e/0x410 [btrfs]
  [245043.432006]  transaction_kthread+0x184/0x1a0 [btrfs]
  [245043.433341]  kthread+0xf0/0x130
  [245043.434628]  ? btrfs_cleanup_transaction+0x4e0/0x4e0 [btrfs]
  [245043.435928]  ? kthread_create_worker_on_cpu+0x40/0x40
  [245043.437236]  ret_from_fork+0x1f/0x30
  [245043.441054] ---[ end trace 15abaa2aaf36827f ]---

This is because at write_one_cache_group() when we are COWing a leaf from
the extent tree we end up allocating a new block group (chunk) and,
because we have hit a threshold on the number of bytes reserved for system
chunks, we attempt to finalize the creation of new block groups from the
current transaction, by calling btrfs_create_pending_block_groups().
However here we also need to modify the extent tree in order to insert
a block group item, and if the location for this new block group item
happens to be in the same leaf that we were COWing earlier, we deadlock
since btrfs_search_slot() tries to write lock the extent buffer that we
locked before at write_one_cache_group().

We have already hit similar cases in the past and commit d9a0540a79f8
("Btrfs: fix deadlock when finalizing block group creation") fixed some
of those cases by delaying the creation of pending block groups at the
known specific spots that could lead to a deadlock. This change reworks
that commit to be more generic so that we don't have to add similar logic
to every possible path that can lead to a deadlock. This is done by
making __btrfs_cow_block() disallowing the creation of new block groups
(setting the transaction's can_flush_pending_bgs to false) before it
attempts to allocate a new extent buffer for either the extent, chunk or
device trees, since those are the trees that pending block creation
modifies. Once the new extent buffer is allocated, it allows creation of
pending block groups to happen again.

This change depends on a recent patch from Josef which is not yet in
Linus' tree, named "btrfs: make sure we create all new block groups" in
order to avoid occasional warnings at btrfs_trans_release_chunk_metadata().

Fixes: d9a0540a79f8 ("Btrfs: fix deadlock when finalizing block group creation")
CC: stable@vger.kernel.org # 4.4+
Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=199753
Link: https://lore.kernel.org/linux-btrfs/CAJtFHUTHna09ST-_EEiyWmDH6gAqS6wa=zMNMBsifj8ABu99cw@mail.gmail.com/
Reported-by: E V <eliventer@gmail.com>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
fs/btrfs/ctree.c
fs/btrfs/extent-tree.c