On 2024/12/16 23:30, Ming Lei wrote: > On Tue, Dec 17, 2024 at 08:19:28AM +0100, Christoph Hellwig wrote: >> On Tue, Dec 17, 2024 at 03:05:48PM +0800, Ming Lei wrote: >>> On Tue, Dec 17, 2024 at 05:40:56AM +0100, Christoph Hellwig wrote: >>>> On Tue, Dec 17, 2024 at 09:52:51AM +0800, Ming Lei wrote: >>>>> The local copy can be updated in any way with any data, so does another >>>>> concurrent update on q->limits really matter? >>>> >>>> Yes, because that means one of the updates get lost even if it is >>>> for entirely separate fields. >>> >>> Right, but the limits are still valid anytime. >>> >>> Any suggestion for fixing this deadlock? >> >> What is "this deadlock"? > > The commit log provides two reports: > > - lockdep warning > > https://lore.kernel.org/linux-block/Z1A8fai9_fQFhs1s@xxxxxxxxxxxxxxxxxxxx/ > > - real deadlock report > > https://lore.kernel.org/linux-scsi/ZxG38G9BuFdBpBHZ@fedora/ > > It is actually one simple ABBA lock: > > 1) queue_attr_store() > > blk_mq_freeze_queue(q); //queue freeze lock > res = entry->store(disk, page, length); > queue_limits_start_update //->limits_lock > ... > queue_limits_commit_update > blk_mq_unfreeze_queue(q); The locking + freeze pattern should be: lim = queue_limits_start_update(q); ... blk_mq_freeze_queue(q); ret = queue_limits_commit_update(q, &lim); blk_mq_unfreeze_queue(q); This pattern is used in most places and anything that does not use it is likely susceptible to a similar ABBA deadlock. We should probably look into trying to integrate the freeze/unfreeze calls directly into queue_limits_commit_update(). Fixing queue_attr_store() to use this pattern seems simpler than trying to fix sd_revalidate_disk(). > > 2) sd_revalidate_disk() > > queue_limits_start_update //->limits_lock > sd_read_capacity() > scsi_execute_cmd > scsi_alloc_request > blk_queue_enter //queue freeze lock > > > Thanks, > Ming > > -- Damien Le Moal Western Digital Research
