On 2021/10/15 7:02, Tejun Heo wrote:

On Wed, Oct 13, 2021 at 08:44:56PM +0800, Yu Kuai wrote:

quoted

@@ -1401,6 +1406,7 @@ void blkcg_deactivate_policy(struct request_queue *q,
  	if (queue_is_mq(q))
  		blk_mq_freeze_queue(q);
  
+	mutex_lock(&q->blkg_lock);
  	spin_lock_irq(&q->queue_lock);

Given that deactivation drains q_usage_counter through
blk_mq_freeze_queue(), can't the blkg_conf_prep() just pin the usage count?

In fact, the caller of blkcg_deactivate_policy() aready freeze the
queue, either from blk_cleanup_queue() or from elevator_switch().

Grab q_usage_counter in blk_conf_prep() can prevent concurrent with
blkcg_deactivate_policy(), I'm afraid will it be too much since it
will also be synchoronized with all the other freeze/unfreeze queues?

Thanks,
Kuai