[PATCH 1/1] block: Use RCU in blk_mq_[un]quiesce_tagset() instead of set->tag_list_lock

[Mohamed Khalfella]

On Thu 2025-12-04 08:22:23 -1000, Bart Van Assche wrote:
> On 12/4/25 8:11 AM, Mohamed Khalfella wrote:
> > @@ -4302,6 +4302,8 @@ static void blk_mq_del_queue_tag_set(struct request_queue *q)
> >   		blk_mq_update_tag_set_shared(set, false);
> >   	}
> >   	mutex_unlock(&set->tag_list_lock);
> > +
> > +	synchronize_rcu();
> >   	INIT_LIST_HEAD(&q->tag_set_list);
> >   }
> Yikes. This change slows down all blk_mq_del_queue_tag_set() callers.

synchronize_rcu() is necessary before re-initializing q->tag_set_list
because of list_for_each_entry_rcu() in blk_mq_[un]quiesce_tagset().

Is blk_mq_del_queue_tag_set() performance sensitive such that it can not
take synchronize_rcu()? It is not in IO codepath, right?

> Please fix the reported deadlock by modifying the NVMe code instead of
> slowing down the block layer.

I can not think of an easy way to do that. Suggestions are welcomed.

> 
> Thanks,
> 
> Bart.