On Fri, 2018-07-27 at 10:46 -0600, Keith Busch wrote:

On Fri, Jul 27, 2018 at 09:20:42AM -0700, Bart Van Assche wrote:

quoted

+	ret = req->q->mq_ops->timeout(req, reser=

ved);

quoted

+	/*
+	 * BLK_EH_DONT_RESET_TIMER means that th=

e block driver either

quoted

+	 * completed the request or still owns the request and w=

ill

quoted

+	 * continue processing the timeout asynchronously. In th=

e

quoted

+	 * latter case, if blk_mq_complete_request()=

 was called while

quoted

+	 * the timeout handler was in progress, ignore that call=

.

quoted

+	 */
+	if (ret == BLK_EH_DONT_RESET_TIMER)
+		return;

=20
This is how completions get lost.

The new approach for handling completions that occur while the .timeout()
callback in progress is as follows:
* blk_mq_complete_request() executes the following code:
       if (blk_mq_change_rq_state(rq, MQ_RQ_TIMED�
F8-OUT, MQ_RQ_COMPLETE))
               return;
* blk_mq_rq_timed_out() executes the following code:
       if (blk_mq_rq_state(req) == MQ_RQ_COMPLETE=
) {
               __blk_mq_complete_request(req);
               return;
       }

As one can see __blk_mq_complete_request() gets called if=
 this race occurs.

Bart.