btrfs: fix hang during unmount when block group reclaim task is running

commit 31e70e527806c546a72262f2fc3d982ee23c42d3 upstream.

When we start an unmount, at close_ctree(), if we have the reclaim task
running and in the middle of a data block group relocation, we can trigger
a deadlock when stopping an async reclaim task, producing a trace like the
following:

[629724.498185] task:kworker/u16:7   state:D stack:    0 pid:681170 ppid:     2 flags:0x00004000
[629724.499760] Workqueue: events_unbound btrfs_async_reclaim_metadata_space [btrfs]
[629724.501267] Call Trace:
[629724.501759]  <TASK>
[629724.502174]  __schedule+0x3cb/0xed0
[629724.502842]  schedule+0x4e/0xb0
[629724.503447]  btrfs_wait_on_delayed_iputs+0x7c/0xc0 [btrfs]
[629724.504534]  ? prepare_to_wait_exclusive+0xc0/0xc0
[629724.505442]  flush_space+0x423/0x630 [btrfs]
[629724.506296]  ? rcu_read_unlock_trace_special+0x20/0x50
[629724.507259]  ? lock_release+0x220/0x4a0
[629724.507932]  ? btrfs_get_alloc_profile+0xb3/0x290 [btrfs]
[629724.508940]  ? do_raw_spin_unlock+0x4b/0xa0
[629724.509688]  btrfs_async_reclaim_metadata_space+0x139/0x320 [btrfs]
[629724.510922]  process_one_work+0x252/0x5a0
[629724.511694]  ? process_one_work+0x5a0/0x5a0
[629724.512508]  worker_thread+0x52/0x3b0
[629724.513220]  ? process_one_work+0x5a0/0x5a0
[629724.514021]  kthread+0xf2/0x120
[629724.514627]  ? kthread_complete_and_exit+0x20/0x20
[629724.515526]  ret_from_fork+0x22/0x30
[629724.516236]  </TASK>
[629724.516694] task:umount          state:D stack:    0 pid:719055 ppid:695412 flags:0x00004000
[629724.518269] Call Trace:
[629724.518746]  <TASK>
[629724.519160]  __schedule+0x3cb/0xed0
[629724.519835]  schedule+0x4e/0xb0
[629724.520467]  schedule_timeout+0xed/0x130
[629724.521221]  ? lock_release+0x220/0x4a0
[629724.521946]  ? lock_acquired+0x19c/0x420
[629724.522662]  ? trace_hardirqs_on+0x1b/0xe0
[629724.523411]  __wait_for_common+0xaf/0x1f0
[629724.524189]  ? usleep_range_state+0xb0/0xb0
[629724.524997]  __flush_work+0x26d/0x530
[629724.525698]  ? flush_workqueue_prep_pwqs+0x140/0x140
[629724.526580]  ? lock_acquire+0x1a0/0x310
[629724.527324]  __cancel_work_timer+0x137/0x1c0
[629724.528190]  close_ctree+0xfd/0x531 [btrfs]
[629724.529000]  ? evict_inodes+0x166/0x1c0
[629724.529510]  generic_shutdown_super+0x74/0x120
[629724.530103]  kill_anon_super+0x14/0x30
[629724.530611]  btrfs_kill_super+0x12/0x20 [btrfs]
[629724.531246]  deactivate_locked_super+0x31/0xa0
[629724.531817]  cleanup_mnt+0x147/0x1c0
[629724.532319]  task_work_run+0x5c/0xa0
[629724.532984]  exit_to_user_mode_prepare+0x1a6/0x1b0
[629724.533598]  syscall_exit_to_user_mode+0x16/0x40
[629724.534200]  do_syscall_64+0x48/0x90
[629724.534667]  entry_SYSCALL_64_after_hwframe+0x44/0xae
[629724.535318] RIP: 0033:0x7fa2b90437a7
[629724.535804] RSP: 002b:00007ffe0b7e4458 EFLAGS: 00000246 ORIG_RAX: 00000000000000a6
[629724.536912] RAX: 0000000000000000 RBX: 00007fa2b9182264 RCX: 00007fa2b90437a7
[629724.538156] RDX: 0000000000000000 RSI: 0000000000000000 RDI: 0000555d6cf20dd0
[629724.539053] RBP: 0000555d6cf20ba0 R08: 0000000000000000 R09: 00007ffe0b7e3200
[629724.539956] R10: 0000000000000000 R11: 0000000000000246 R12: 0000000000000000
[629724.540883] R13: 0000555d6cf20dd0 R14: 0000555d6cf20cb0 R15: 0000000000000000
[629724.541796]  </TASK>

This happens because:

1) Before entering close_ctree() we have the async block group reclaim
   task running and relocating a data block group;

2) There's an async metadata (or data) space reclaim task running;

3) We enter close_ctree() and park the cleaner kthread;

4) The async space reclaim task is at flush_space() and runs all the
   existing delayed iputs;

5) Before the async space reclaim task calls
   btrfs_wait_on_delayed_iputs(), the block group reclaim task which is
   doing the data block group relocation, creates a delayed iput at
   replace_file_extents() (called when COWing leaves that have file extent
   items pointing to relocated data extents, during the merging phase
   of relocation roots);

6) The async reclaim space reclaim task blocks at
   btrfs_wait_on_delayed_iputs(), since we have a new delayed iput;

7) The task at close_ctree() then calls cancel_work_sync() to stop the
   async space reclaim task, but it blocks since that task is waiting for
   the delayed iput to be run;

8) The delayed iput is never run because the cleaner kthread is parked,
   and no one else runs delayed iputs, resulting in a hang.

So fix this by stopping the async block group reclaim task before we
park the cleaner kthread.

Fixes: 18bb8bbf13c183 ("btrfs: zoned: automatically reclaim zones")
CC: stable@vger.kernel.org # 5.15+
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>

diff --git a/fs/btrfs/disk-io.c b/fs/btrfs/disk-io.c
index 9ab2792055271a36b4dd522aefdc8a6613fa5875..233d894f6febaf872858b6a59899dddccb0fbafb 100644 (file)
--- a/fs/btrfs/disk-io.c
+++ b/fs/btrfs/disk-io.c
@@ -4360,6 +4360,17 @@ void __cold close_ctree(struct btrfs_fs_info *fs_info)
        int ret;
 
        set_bit(BTRFS_FS_CLOSING_START, &fs_info->flags);
+
+       /*
+        * We may have the reclaim task running and relocating a data block group,
+        * in which case it may create delayed iputs. So stop it before we park
+        * the cleaner kthread otherwise we can get new delayed iputs after
+        * parking the cleaner, and that can make the async reclaim task to hang
+        * if it's waiting for delayed iputs to complete, since the cleaner is
+        * parked and can not run delayed iputs - this will make us hang when
+        * trying to stop the async reclaim task.
+        */
+       cancel_work_sync(&fs_info->reclaim_bgs_work);
        /*
         * We don't want the cleaner to start new transactions, add more delayed
         * iputs, etc. while we're closing. We can't use kthread_stop() yet
@@ -4400,8 +4411,6 @@ void __cold close_ctree(struct btrfs_fs_info *fs_info)
        cancel_work_sync(&fs_info->async_data_reclaim_work);
        cancel_work_sync(&fs_info->preempt_reclaim_work);
 
-       cancel_work_sync(&fs_info->reclaim_bgs_work);
-
        /* Cancel or finish ongoing discard work */
        btrfs_discard_cleanup(fs_info);