Btrfs: fix deadlock between fiemap and transaction commits
authorFilipe Manana <fdmanana@suse.com>
Mon, 29 Jul 2019 08:37:10 +0000 (09:37 +0100)
committerDavid Sterba <dsterba@suse.com>
Tue, 30 Jul 2019 16:25:12 +0000 (18:25 +0200)
commita6d155d2e363f26290ffd50591169cb96c2a609e
treee7c2163140c230c8f0b3c1df5714bd086ed8781e
parentcb2d3daddbfb6318d170e79aac1f7d5e4d49f0d7
Btrfs: fix deadlock between fiemap and transaction commits

The fiemap handler locks a file range that can have unflushed delalloc,
and after locking the range, it tries to attach to a running transaction.
If the running transaction started its commit, that is, it is in state
TRANS_STATE_COMMIT_START, and either the filesystem was mounted with the
flushoncommit option or the transaction is creating a snapshot for the
subvolume that contains the file that fiemap is operating on, we end up
deadlocking. This happens because fiemap is blocked on the transaction,
waiting for it to complete, and the transaction is waiting for the flushed
dealloc to complete, which requires locking the file range that the fiemap
task already locked. The following stack traces serve as an example of
when this deadlock happens:

  (...)
  [404571.515510] Workqueue: btrfs-endio-write btrfs_endio_write_helper [btrfs]
  [404571.515956] Call Trace:
  [404571.516360]  ? __schedule+0x3ae/0x7b0
  [404571.516730]  schedule+0x3a/0xb0
  [404571.517104]  lock_extent_bits+0x1ec/0x2a0 [btrfs]
  [404571.517465]  ? remove_wait_queue+0x60/0x60
  [404571.517832]  btrfs_finish_ordered_io+0x292/0x800 [btrfs]
  [404571.518202]  normal_work_helper+0xea/0x530 [btrfs]
  [404571.518566]  process_one_work+0x21e/0x5c0
  [404571.518990]  worker_thread+0x4f/0x3b0
  [404571.519413]  ? process_one_work+0x5c0/0x5c0
  [404571.519829]  kthread+0x103/0x140
  [404571.520191]  ? kthread_create_worker_on_cpu+0x70/0x70
  [404571.520565]  ret_from_fork+0x3a/0x50
  [404571.520915] kworker/u8:6    D    0 31651      2 0x80004000
  [404571.521290] Workqueue: btrfs-flush_delalloc btrfs_flush_delalloc_helper [btrfs]
  (...)
  [404571.537000] fsstress        D    0 13117  13115 0x00004000
  [404571.537263] Call Trace:
  [404571.537524]  ? __schedule+0x3ae/0x7b0
  [404571.537788]  schedule+0x3a/0xb0
  [404571.538066]  wait_current_trans+0xc8/0x100 [btrfs]
  [404571.538349]  ? remove_wait_queue+0x60/0x60
  [404571.538680]  start_transaction+0x33c/0x500 [btrfs]
  [404571.539076]  btrfs_check_shared+0xa3/0x1f0 [btrfs]
  [404571.539513]  ? extent_fiemap+0x2ce/0x650 [btrfs]
  [404571.539866]  extent_fiemap+0x2ce/0x650 [btrfs]
  [404571.540170]  do_vfs_ioctl+0x526/0x6f0
  [404571.540436]  ksys_ioctl+0x70/0x80
  [404571.540734]  __x64_sys_ioctl+0x16/0x20
  [404571.540997]  do_syscall_64+0x60/0x1d0
  [404571.541279]  entry_SYSCALL_64_after_hwframe+0x49/0xbe
  (...)
  [404571.543729] btrfs           D    0 14210  14208 0x00004000
  [404571.544023] Call Trace:
  [404571.544275]  ? __schedule+0x3ae/0x7b0
  [404571.544526]  ? wait_for_completion+0x112/0x1a0
  [404571.544795]  schedule+0x3a/0xb0
  [404571.545064]  schedule_timeout+0x1ff/0x390
  [404571.545351]  ? lock_acquire+0xa6/0x190
  [404571.545638]  ? wait_for_completion+0x49/0x1a0
  [404571.545890]  ? wait_for_completion+0x112/0x1a0
  [404571.546228]  wait_for_completion+0x131/0x1a0
  [404571.546503]  ? wake_up_q+0x70/0x70
  [404571.546775]  btrfs_wait_ordered_extents+0x27c/0x400 [btrfs]
  [404571.547159]  btrfs_commit_transaction+0x3b0/0xae0 [btrfs]
  [404571.547449]  ? btrfs_mksubvol+0x4a4/0x640 [btrfs]
  [404571.547703]  ? remove_wait_queue+0x60/0x60
  [404571.547969]  btrfs_mksubvol+0x605/0x640 [btrfs]
  [404571.548226]  ? __sb_start_write+0xd4/0x1c0
  [404571.548512]  ? mnt_want_write_file+0x24/0x50
  [404571.548789]  btrfs_ioctl_snap_create_transid+0x169/0x1a0 [btrfs]
  [404571.549048]  btrfs_ioctl_snap_create_v2+0x11d/0x170 [btrfs]
  [404571.549307]  btrfs_ioctl+0x133f/0x3150 [btrfs]
  [404571.549549]  ? mem_cgroup_charge_statistics+0x4c/0xd0
  [404571.549792]  ? mem_cgroup_commit_charge+0x84/0x4b0
  [404571.550064]  ? __handle_mm_fault+0xe3e/0x11f0
  [404571.550306]  ? do_raw_spin_unlock+0x49/0xc0
  [404571.550608]  ? _raw_spin_unlock+0x24/0x30
  [404571.550976]  ? __handle_mm_fault+0xedf/0x11f0
  [404571.551319]  ? do_vfs_ioctl+0xa2/0x6f0
  [404571.551659]  ? btrfs_ioctl_get_supported_features+0x30/0x30 [btrfs]
  [404571.552087]  do_vfs_ioctl+0xa2/0x6f0
  [404571.552355]  ksys_ioctl+0x70/0x80
  [404571.552621]  __x64_sys_ioctl+0x16/0x20
  [404571.552864]  do_syscall_64+0x60/0x1d0
  [404571.553104]  entry_SYSCALL_64_after_hwframe+0x49/0xbe
  (...)

If we were joining the transaction instead of attaching to it, we would
not risk a deadlock because a join only blocks if the transaction is in a
state greater then or equals to TRANS_STATE_COMMIT_DOING, and the delalloc
flush performed by a transaction is done before it reaches that state,
when it is in the state TRANS_STATE_COMMIT_START. However a transaction
join is intended for use cases where we do modify the filesystem, and
fiemap only needs to peek at delayed references from the current
transaction in order to determine if extents are shared, and, besides
that, when there is no current transaction or when it blocks to wait for
a current committing transaction to complete, it creates a new transaction
without reserving any space. Such unnecessary transactions, besides doing
unnecessary IO, can cause transaction aborts (-ENOSPC) and unnecessary
rotation of the precious backup roots.

So fix this by adding a new transaction join variant, named join_nostart,
which behaves like the regular join, but it does not create a transaction
when none currently exists or after waiting for a committing transaction
to complete.

Fixes: 03628cdbc64db6 ("Btrfs: do not start a transaction during fiemap")
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
fs/btrfs/backref.c
fs/btrfs/transaction.c
fs/btrfs/transaction.h