26.03.2020 18:56, Stefan Reiter wrote:
All code-paths leading to backup_clean (via job_clean) have the job's context already acquired. The job's context is guaranteed to be the same as the one used by backup_top via backup_job_create.
As we already discussed, this is not quite right. So, may be this patch should be the last one...
Since the previous logic effectively acquired the lock twice, this broke cleanup of backups for disks using IO threads, since the BDRV_POLL_WHILE in bdrv_backup_top_drop -> bdrv_do_drained_begin would only release the lock once, thus deadlocking with the IO thread. This is a partial revert of 0abf2581717a19. Signed-off-by: Stefan Reiter <s.rei...@proxmox.com> --- block/backup.c | 4 ---- 1 file changed, 4 deletions(-) diff --git a/block/backup.c b/block/backup.c index 7430ca5883..a7a7dcaf4c 100644 --- a/block/backup.c +++ b/block/backup.c @@ -126,11 +126,7 @@ static void backup_abort(Job *job) static void backup_clean(Job *job) { BackupBlockJob *s = container_of(job, BackupBlockJob, common.job); - AioContext *aio_context = bdrv_get_aio_context(s->backup_top); - - aio_context_acquire(aio_context); bdrv_backup_top_drop(s->backup_top); - aio_context_release(aio_context); }void backup_do_checkpoint(BlockJob *job, Error **errp)
-- Best regards, Vladimir
