Re: [PATCH] btrfs: fix deadlock between quota disable and qgroup rescan worker

[Nikolay Borisov <nborisov@suse.com>]

On 13.01.22 г. 12:40, Shin'ichiro Kawasaki wrote:
> Quota disable ioctl starts a transaction before waiting for the qgroup
> rescan worker completes. However, this wait can be infinite and results
> in deadlock because of circular dependency among the quota disable
> ioctl, the qgroup rescan worker and the other task with transaction such
> as block group relocation task.
> 
> The deadlock happens with the steps following:
> 
> 1) Task A calls ioctl to disable quota. It starts a transaction and
>    waits for qgroup rescan worker completes.
> 2) Task B such as block group relocation task starts a transaction and
>    joins to the transaction that task A started. Then task B commits to
>    the transaction. In this commit, task B waits for a commit by task A.
> 3) Task C as the qgroup rescan worker starts its job and starts a
>    transaction. In this transaction start, task C waits for completion
>    of the transaction that task A started and task B committed.
> 
> This deadlock was found with fstests test case block/115 and a zoned
> null_blk device. The test case enables and disables quota, and the
> block group reclaim was triggered during the quota disable by chance.
> The deadlock was also observed by running quota enable and disable in
> parallel with 'btrfs balance' command on regular null_blk devices.
> 


In btrfs_qgroup_rescan_worker can't we make the check for
BTRFS_FS_QUOTA_ENABLED be done _before_ starting the transaction by
refactoring the body of the loop? Even better will be to move the
FS_QUOTA_ENABLED check in rescan_should_stop. That way transaction
commit/quota disable will be delayed until the end of the current
qgroup_rescan_leaf invocation, no?