Re: [dm-devel] should blk-mq halt requeue processing while queue is frozen?

On 09/07/2016 06:41 PM, Mike Snitzer wrote:

On Fri, Sep 02 2016 at  6:42pm -0400,
Bart Van Assche [off-list ref] wrote:

quoted

+/**
+ * blk_mq_quiesce_queue - wait until all pending queue_rq calls have finished
+ *
+ * Prevent that new I/O requests are queued and wait until all pending
+ * queue_rq() calls have finished.
+ */
+void blk_mq_quiesce_queue(struct request_queue *q)
+{
+	spin_lock_irq(q->queue_lock);
+	WARN_ON_ONCE(blk_queue_quiescing(q));
+	queue_flag_set(QUEUE_FLAG_QUIESCING, q);
+	spin_unlock_irq(q->queue_lock);
+
+	atomic_inc_return(&q->mq_freeze_depth);
+	blk_mq_run_hw_queues(q, false);
+	synchronize_rcu();

Why the synchronize_rcu()?

Hello Mike,

Adding read_lock() + read_unlock() in __blk_mq_run_hw_queue() and synchronize_rcu() in blk_mq_quiesce_queue() is the lowest overhead mechanism I know of to make the latter function wait until the former has finished.

Also, you're effectively open-coding blk_mq_freeze_queue_start() minus
the q->q_usage_counter mgmt.  Why not add a flag to conditionally manage
q->q_usage_counter to blk_mq_freeze_queue_start()?

I will consider this.

But I'm concerned about blk_mq_{quiesce,resume}_queue vs
blk_mq_{freeze,unfreeze}_queue -- e.g. if "freeze" is nested after
"queue" (but before "resume") it would still need the q->q_usage_counter
management.  Your patch as-is would break the blk-mq freeze interface.

Agreed. blk_mq_{quiesce,resume}_queue() has to manipulate q_usage_counter in the same way as blk_mq_{freeze,unfreeze}_queue(). Once I am back in the office I will rework this patch and send it to Jens.

Bart.