On 4/6/20 4:51 AM, Damien Le Moal wrote:
> Commit b27aef6abfba ("zbd: use zone_lock to lock a zone") to fix
> potential deadlocks with zonemode=zbd zone locking was incomplete.
> The execution of the zone lock stress test t/zbd test case 48 still
> sometimes lead to deadlocks (a large number of repeated execution is
> sometimes needed).
>
> The remaining deadlock pattern identified with the repeated execution
> of this test is due to the concurrent execution of jobs doing random
> async writes to zones. In such case, any of the job may trigger an all
> zone reset through the path get_next_rand_block() -> fio_file_reset()
> while async writes are still inflight. The fix for this is to use the
> zone_lock() function instead of directly calling pthread_mutex_lock()i
> to ensure that no async IO is inflight for a zone that is part of a
> reset range.
Applied, thanks.
--
Jens Axboe
