Re: [PATCH] blk-mq: run queue after issuing the last request of the plug list

[Ming Lei <ming.lei@redhat.com>]

On Tue, Jul 26, 2022 at 11:14:23AM +0800, Yu Kuai wrote:
> Hi, Ming
> 
> 在 2022/07/26 11:02, Ming Lei 写道:
> > On Tue, Jul 26, 2022 at 10:52:56AM +0800, Yu Kuai wrote:
> > > Hi, Ming
> > > 在 2022/07/26 10:32, Ming Lei 写道:
> > > > On Tue, Jul 26, 2022 at 10:08:13AM +0800, Yu Kuai wrote:
> > > > > 在 2022/07/26 9:46, Ming Lei 写道:
> > > > > > On Tue, Jul 26, 2022 at 09:08:19AM +0800, Yu Kuai wrote:
> > > > > > > Hi, Ming!
> > > > > > > 
> > > > > > > 在 2022/07/25 23:43, Ming Lei 写道:
> > > > > > > > On Sat, Jul 23, 2022 at 10:50:03AM +0800, Yu Kuai wrote:
> > > > > > > > > Hi, Ming!
> > > > > > > > > 
> > > > > > > > > 在 2022/07/19 17:26, Ming Lei 写道:
> > > > > > > > > > On Mon, Jul 18, 2022 at 08:35:28PM +0800, Yufen Yu wrote:
> > > > > > > > > > > We do test on a virtio scsi device (/dev/sda) and the default mq
> > > > > > > > > > > scheduler is 'none'. We found a IO hung as following:
> > > > > > > > > > > 
> > > > > > > > > > > blk_finish_plug
> > > > > > > > > > >        blk_mq_plug_issue_direct
> > > > > > > > > > >            scsi_mq_get_budget
> > > > > > > > > > >            //get budget_token fail and sdev->restarts=1
> > > > > > > > > > > 
> > > > > > > > > > > 			     	 scsi_end_request
> > > > > > > > > > > 				   scsi_run_queue_async
> > > > > > > > > > >                                         //sdev->restart=0 and run queue
> > > > > > > > > > > 
> > > > > > > > > > >           blk_mq_request_bypass_insert
> > > > > > > > > > >              //add request to hctx->dispatch list
> > > > > > > > > > 
> > > > > > > > > > Here the issue shouldn't be related with scsi's get budget or
> > > > > > > > > > scsi_run_queue_async.
> > > > > > > > > > 
> > > > > > > > > > If blk-mq adds request into ->dispatch_list, it is blk-mq core's
> > > > > > > > > > responsibility to re-run queue for moving on. Can you investigate a
> > > > > > > > > > bit more why blk-mq doesn't run queue after adding request to
> > > > > > > > > > hctx dispatch list?
> > > > > > > > > 
> > > > > > > > > I think Yufen is probably thinking about the following Concurrent
> > > > > > > > > scenario:
> > > > > > > > > 
> > > > > > > > > blk_mq_flush_plug_list
> > > > > > > > > # assume there are three rq
> > > > > > > > >      blk_mq_plug_issue_direct
> > > > > > > > >       blk_mq_request_issue_directly
> > > > > > > > >       # dispatch rq1, succeed
> > > > > > > > >       blk_mq_request_issue_directly
> > > > > > > > >       # dispatch rq2
> > > > > > > > >        __blk_mq_try_issue_directly
> > > > > > > > >         blk_mq_get_dispatch_budget
> > > > > > > > >          scsi_mq_get_budget
> > > > > > > > >           atomic_inc(&sdev->restarts);
> > > > > > > > >           # rq2 failed to get budget
> > > > > > > > >           # restarts is 1 now
> > > > > > > > >                                             scsi_end_request
> > > > > > > > >                                             # rq1 is completed
> > > > > > > > >                                             ┊scsi_run_queue_async
> > > > > > > > >                                             ┊ atomic_cmpxchg(&sdev->restarts,
> > > > > > > > > old, 0) == old
> > > > > > > > >                                             ┊ # set restarts to 0
> > > > > > > > >                                             ┊ blk_mq_run_hw_queues
> > > > > > > > >                                             ┊ # hctx->dispatch list is empty
> > > > > > > > >       blk_mq_request_bypass_insert
> > > > > > > > >       # insert rq2 to hctx->dispatch list
> > > > > > > > 
> > > > > > > > After rq2 is added to ->dispatch_list in blk_mq_try_issue_list_directly(),
> > > > > > > > no matter if list_empty(list) is empty or not, queue will be run either from
> > > > > > > > blk_mq_request_bypass_insert() or blk_mq_sched_insert_requests().
> > > > > > > 
> > > > > > > 1) while inserting rq2 to dispatch list, blk_mq_request_bypass_insert()
> > > > > > > is called from blk_mq_try_issue_list_directly(), list_empty() won't
> > > > > > > pass, thus thus blk_mq_request_bypass_insert() won't run queue.
> > > > > > 
> > > > > > Yeah, but in blk_mq_try_issue_list_directly() after rq2 is inserted to dispatch
> > > > > > list, the loop is broken and blk_mq_try_issue_list_directly() returns to
> > > > > > blk_mq_sched_insert_requests() in which list_empty() is false, so
> > > > > > blk_mq_insert_requests() and blk_mq_run_hw_queue() are called, queue
> > > > > > is still run.
> > > > > > 
> > > > > > Also not sure why you make rq3 involved, since the list is local list on
> > > > > > stack, and it can be operated concurrently.
> > > > > 
> > > > > I make rq3 involved because there are some conditions that
> > > > > blk_mq_insert_requests() and blk_mq_run_hw_queue() won't be called from
> > > > > blk_mq_sched_insert_requests():
> > > > 
> > > > The two won't be called if list_empty() is true, and will be called if
> > > > !list_empty().
> > > > 
> > > > That is why I mentioned run queue has been done after rq2 is added to
> > > > ->dispatch_list.
> > > 
> > > I don't follow here, it's right after rq2 is inserted to dispatch list,
> > > list is not empty, and blk_mq_sched_insert_requests() will be called.
> > > However, do you think that it's impossible that
> > > blk_mq_sched_insert_requests() can dispatch rq in the list and list
> > > will become empty?
> > 
> > Please take a look at blk_mq_sched_insert_requests().
> > 
> > When codes runs into blk_mq_sched_insert_requests(), the following
> > blk_mq_run_hw_queue() will be run always, how does list empty or not
> > make a difference there?
> 
> This is strange, always blk_mq_run_hw_queue() is exactly what Yufen
> tries to do in this patch, are we look at different code?

No.

> 
> I'm copying blk_mq_sched_insert_requests() here, the code is from
> latest linux-next:
> 
> 461 void blk_mq_sched_insert_requests(struct blk_mq_hw_ctx *hctx,
> 462                                 ┊ struct blk_mq_ctx *ctx,
> 463                                 ┊ struct list_head *list, bool
> run_queue_async)
> 464 {
> 465         struct elevator_queue *e;
> 466         struct request_queue *q = hctx->queue;
> 467
> 468         /*
> 469         ┊* blk_mq_sched_insert_requests() is called from flush plug
> 470         ┊* context only, and hold one usage counter to prevent queue
> 471         ┊* from being released.
> 472         ┊*/
> 473         percpu_ref_get(&q->q_usage_counter);
> 474
> 475         e = hctx->queue->elevator;
> 476         if (e) {
> 477                 e->type->ops.insert_requests(hctx, list, false);
> 478         } else {
> 479                 /*
> 480                 ┊* try to issue requests directly if the hw queue isn't
> 481                 ┊* busy in case of 'none' scheduler, and this way may
> save
> 482                 ┊* us one extra enqueue & dequeue to sw queue.
> 483                 ┊*/
> 484                 if (!hctx->dispatch_busy && !run_queue_async) {
> 485                         blk_mq_run_dispatch_ops(hctx->queue,
> 486                                 blk_mq_try_issue_list_directly(hctx,
> list));
> 487                         if (list_empty(list))
> 488                                 goto out;
> 489                 }
> 490                 blk_mq_insert_requests(hctx, ctx, list);
> 491         }
> 492
> 493         blk_mq_run_hw_queue(hctx, run_queue_async);
> 494  out:
> 495         percpu_ref_put(&q->q_usage_counter);
> 496 }
> 
> Here in line 487, if list_empty() is true, out label will skip
> run_queue().

If list_empty() is true, run queue is guaranteed to run
in blk_mq_try_issue_list_directly() in case that BLK_STS_*RESOURCE
is returned from blk_mq_request_issue_directly().

		ret = blk_mq_request_issue_directly(rq, list_empty(list));
		if (ret != BLK_STS_OK) {
			if (ret == BLK_STS_RESOURCE ||
					ret == BLK_STS_DEV_RESOURCE) {
				blk_mq_request_bypass_insert(rq, false,
							list_empty(list));	//run queue
				break;
			}
			blk_mq_end_request(rq, ret);
			errors++;
		} else
			queued++;

So why do you try to add one extra run queue?


Thanks,
Ming