]> Git Repo - linux.git/commit
btrfs: take the cleaner_mutex earlier in qgroup disable
authorJosef Bacik <[email protected]>
Fri, 19 Apr 2024 18:38:48 +0000 (14:38 -0400)
committerDavid Sterba <[email protected]>
Thu, 25 Apr 2024 14:23:09 +0000 (16:23 +0200)
commit0f2b8098d72a93890e69aa24ec549ef4bc34f4db
tree10587186d12f6d345e2329eba5a7d8edd2793c9e
parent9af503d91298c3f2945e73703f0e00995be08c30
btrfs: take the cleaner_mutex earlier in qgroup disable

One of my CI runs popped the following lockdep splat

======================================================
WARNING: possible circular locking dependency detected
6.9.0-rc4+ #1 Not tainted
------------------------------------------------------
btrfs/471533 is trying to acquire lock:
ffff92ba46980850 (&fs_info->cleaner_mutex){+.+.}-{3:3}, at: btrfs_quota_disable+0x54/0x4c0

but task is already holding lock:
ffff92ba46980bd0 (&fs_info->subvol_sem){++++}-{3:3}, at: btrfs_ioctl+0x1c8f/0x2600

which lock already depends on the new lock.

the existing dependency chain (in reverse order) is:

-> #2 (&fs_info->subvol_sem){++++}-{3:3}:
       down_read+0x42/0x170
       btrfs_rename+0x607/0xb00
       btrfs_rename2+0x2e/0x70
       vfs_rename+0xaf8/0xfc0
       do_renameat2+0x586/0x600
       __x64_sys_rename+0x43/0x50
       do_syscall_64+0x95/0x180
       entry_SYSCALL_64_after_hwframe+0x76/0x7e

-> #1 (&sb->s_type->i_mutex_key#16){++++}-{3:3}:
       down_write+0x3f/0xc0
       btrfs_inode_lock+0x40/0x70
       prealloc_file_extent_cluster+0x1b0/0x370
       relocate_file_extent_cluster+0xb2/0x720
       relocate_data_extent+0x107/0x160
       relocate_block_group+0x442/0x550
       btrfs_relocate_block_group+0x2cb/0x4b0
       btrfs_relocate_chunk+0x50/0x1b0
       btrfs_balance+0x92f/0x13d0
       btrfs_ioctl+0x1abf/0x2600
       __x64_sys_ioctl+0x97/0xd0
       do_syscall_64+0x95/0x180
       entry_SYSCALL_64_after_hwframe+0x76/0x7e

-> #0 (&fs_info->cleaner_mutex){+.+.}-{3:3}:
       __lock_acquire+0x13e7/0x2180
       lock_acquire+0xcb/0x2e0
       __mutex_lock+0xbe/0xc00
       btrfs_quota_disable+0x54/0x4c0
       btrfs_ioctl+0x206b/0x2600
       __x64_sys_ioctl+0x97/0xd0
       do_syscall_64+0x95/0x180
       entry_SYSCALL_64_after_hwframe+0x76/0x7e

other info that might help us debug this:

Chain exists of:
  &fs_info->cleaner_mutex --> &sb->s_type->i_mutex_key#16 --> &fs_info->subvol_sem

 Possible unsafe locking scenario:

       CPU0                    CPU1
       ----                    ----
  lock(&fs_info->subvol_sem);
                               lock(&sb->s_type->i_mutex_key#16);
                               lock(&fs_info->subvol_sem);
  lock(&fs_info->cleaner_mutex);

 *** DEADLOCK ***

2 locks held by btrfs/471533:
 #0: ffff92ba4319e420 (sb_writers#14){.+.+}-{0:0}, at: btrfs_ioctl+0x3b5/0x2600
 #1: ffff92ba46980bd0 (&fs_info->subvol_sem){++++}-{3:3}, at: btrfs_ioctl+0x1c8f/0x2600

stack backtrace:
CPU: 1 PID: 471533 Comm: btrfs Kdump: loaded Not tainted 6.9.0-rc4+ #1
Call Trace:
 <TASK>
 dump_stack_lvl+0x77/0xb0
 check_noncircular+0x148/0x160
 ? lock_acquire+0xcb/0x2e0
 __lock_acquire+0x13e7/0x2180
 lock_acquire+0xcb/0x2e0
 ? btrfs_quota_disable+0x54/0x4c0
 ? lock_is_held_type+0x9a/0x110
 __mutex_lock+0xbe/0xc00
 ? btrfs_quota_disable+0x54/0x4c0
 ? srso_return_thunk+0x5/0x5f
 ? lock_acquire+0xcb/0x2e0
 ? btrfs_quota_disable+0x54/0x4c0
 ? btrfs_quota_disable+0x54/0x4c0
 btrfs_quota_disable+0x54/0x4c0
 btrfs_ioctl+0x206b/0x2600
 ? srso_return_thunk+0x5/0x5f
 ? __do_sys_statfs+0x61/0x70
 __x64_sys_ioctl+0x97/0xd0
 do_syscall_64+0x95/0x180
 ? srso_return_thunk+0x5/0x5f
 ? reacquire_held_locks+0xd1/0x1f0
 ? do_user_addr_fault+0x307/0x8a0
 ? srso_return_thunk+0x5/0x5f
 ? lock_acquire+0xcb/0x2e0
 ? srso_return_thunk+0x5/0x5f
 ? srso_return_thunk+0x5/0x5f
 ? find_held_lock+0x2b/0x80
 ? srso_return_thunk+0x5/0x5f
 ? lock_release+0xca/0x2a0
 ? srso_return_thunk+0x5/0x5f
 ? do_user_addr_fault+0x35c/0x8a0
 ? srso_return_thunk+0x5/0x5f
 ? trace_hardirqs_off+0x4b/0xc0
 ? srso_return_thunk+0x5/0x5f
 ? lockdep_hardirqs_on_prepare+0xde/0x190
 ? srso_return_thunk+0x5/0x5f

This happens because when we call rename we already have the inode mutex
held, and then we acquire the subvol_sem if we are a subvolume.  This
makes the dependency

inode lock -> subvol sem

When we're running data relocation we will preallocate space for the
data relocation inode, and we always run the relocation under the
->cleaner_mutex.  This now creates the dependency of

cleaner_mutex -> inode lock (from the prealloc) -> subvol_sem

Qgroup delete is doing this in the opposite order, it is acquiring the
subvol_sem and then it is acquiring the cleaner_mutex, which results in
this lockdep splat.  This deadlock can't happen in reality, because we
won't ever rename the data reloc inode, nor is the data reloc inode a
subvolume.

However this is fairly easy to fix, simply take the cleaner mutex in the
case where we are disabling qgroups before we take the subvol_sem.  This
resolves the lockdep splat.

Reviewed-by: Filipe Manana <[email protected]>
Signed-off-by: Josef Bacik <[email protected]>
Reviewed-by: David Sterba <[email protected]>
Signed-off-by: David Sterba <[email protected]>
fs/btrfs/ioctl.c
fs/btrfs/qgroup.c
This page took 0.05649 seconds and 4 git commands to generate.