在 2021/12/02 22:48, Michal Koutný 写道:

Hello Kuai.

On Thu, Dec 02, 2021 at 09:04:40PM +0800, Yu Kuai [off-list ref] wrote:

quoted

For example, if user thread is throttled with low bps while it's
issuing large io, and the device is deleted. The user thread will
wait for a long time for io to return.

Do I understand correctly the "long time" here is
outstanding_IO_size/throttled_bandwidth? Or are you getting at some

Hi, Michal

Yes, this is exactly what I mean.

other cause/longer time?

quoted

+void blk_throtl_cancel_bios(struct request_queue *q)
+{
+	struct throtl_data *td = q->td;
+	struct bio_list bio_list_on_stack;
+	struct blkcg_gq *blkg;
+	struct cgroup_subsys_state *pos_css;
+	struct bio *bio;
+	int rw;
+
+	bio_list_init(&bio_list_on_stack);
+
+	/*
+	 * hold queue_lock to prevent concurrent with dispatching
+	 * throttled bios by timer.
+	 */
+	spin_lock_irq(&q->queue_lock);

You've replaced the rcu_read_lock() with the queue lock but...

quoted

+
+	/*
+	 * Drain each tg while doing post-order walk on the blkg tree, so
+	 * that all bios are propagated to td->service_queue.  It'd be
+	 * better to walk service_queue tree directly but blkg walk is
+	 * easier.
+	 */
+	blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg)
+		tg_drain_bios(&blkg_to_tg(blkg)->service_queue);

...you also need the rcu_read_lock() here since you may encounter a
(descendant) blkcg that's removed concurrently.

blkg_destroy() is protected by the queue_lock，so I think queue_lock can
protect such concurrent scenario.

Thanks,
Kuai

(I may miss some consequences of doing this under the queue_lock so if
the concurrent removal is ruled out, please make a comment about it.)


Regards,
Michal