Re: [PATCH V2] blk-mq: insert rq with DONTPREP to hctx dispatch list when requeue

[Ming Lei]

On Fri, Feb 15, 2019 at 10:34:39AM +0800, jianchao.wang wrote:
> Hi Ming
> 
> Thanks for your kindly response.
> 
> On 2/15/19 10:00 AM, Ming Lei wrote:
> > On Tue, Feb 12, 2019 at 09:56:25AM +0800, Jianchao Wang wrote:
> >> When requeue, if RQF_DONTPREP, rq has contained some driver
> >> specific data, so insert it to hctx dispatch list to avoid any
> >> merge. Take scsi as example, here is the trace event log (no
> >> io scheduler, because RQF_STARTED would prevent merging),
> >>
> >>    kworker/0:1H-339   [000] ...1  2037.209289: block_rq_insert: 8,0 R 4096 () 32768 + 8 [kworker/0:1H]
> >> scsi_inert_test-1987  [000] ....  2037.220465: block_bio_queue: 8,0 R 32776 + 8 [scsi_inert_test]
> >> scsi_inert_test-1987  [000] ...2  2037.220466: block_bio_backmerge: 8,0 R 32776 + 8 [scsi_inert_test]
> >>    kworker/0:1H-339   [000] ....  2047.220913: block_rq_issue: 8,0 R 8192 () 32768 + 16 [kworker/0:1H]
> >> scsi_inert_test-1996  [000] ..s1  2047.221007: block_rq_complete: 8,0 R () 32768 + 8 [0]
> >> scsi_inert_test-1996  [000] .Ns1  2047.221045: block_rq_requeue: 8,0 R () 32776 + 8 [0]
> >>    kworker/0:1H-339   [000] ...1  2047.221054: block_rq_insert: 8,0 R 4096 () 32776 + 8 [kworker/0:1H]
> >>    kworker/0:1H-339   [000] ...1  2047.221056: block_rq_issue: 8,0 R 4096 () 32776 + 8 [kworker/0:1H]
> >> scsi_inert_test-1986  [000] ..s1  2047.221119: block_rq_complete: 8,0 R () 32776 + 8 [0]
> >>
> >> (32768 + 8) was requeued by scsi_queue_insert and had RQF_DONTPREP.
> > 
> > scsi_mq_requeue_cmd() does uninit the request before requeuing, but
> > __scsi_queue_insert doesn't do that.
> 
> Yes.
> scsi layer use both of them.
> 
> > 
> > 
> >> Then it was merged with (32776 + 8) and issued. Due to RQF_DONTPREP,
> >> the sdb only contained the part of (32768 + 8), then only that part
> >> was completed. The lucky thing was that scsi_io_completion detected
> >> it and requeued the remaining part. So we didn't get corrupted data.
> >> However, the requeue of (32776 + 8) is not expected.
> >>
> >> Suggested-by: Jens Axboe <axboe@xxxxxxxxx>
> >> Signed-off-by: Jianchao Wang <jianchao.w.wang@xxxxxxxxxx>
> >> ---
> >> V2:
> >>  - refactor the code based on Jens' suggestion
> >>
> >>  block/blk-mq.c | 12 ++++++++++--
> >>  1 file changed, 10 insertions(+), 2 deletions(-)
> >>
> >> diff --git a/block/blk-mq.c b/block/blk-mq.c
> >> index 8f5b533..9437a5e 100644
> >> --- a/block/blk-mq.c
> >> +++ b/block/blk-mq.c
> >> @@ -737,12 +737,20 @@ static void blk_mq_requeue_work(struct work_struct *work)
> >>  	spin_unlock_irq(&q->requeue_lock);
> >>  
> >>  	list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
> >> -		if (!(rq->rq_flags & RQF_SOFTBARRIER))
> >> +		if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
> >>  			continue;
> >>  
> >>  		rq->rq_flags &= ~RQF_SOFTBARRIER;
> >>  		list_del_init(&rq->queuelist);
> >> -		blk_mq_sched_insert_request(rq, true, false, false);
> >> +		/*
> >> +		 * If RQF_DONTPREP, rq has contained some driver specific
> >> +		 * data, so insert it to hctx dispatch list to avoid any
> >> +		 * merge.
> >> +		 */
> >> +		if (rq->rq_flags & RQF_DONTPREP)
> >> +			blk_mq_request_bypass_insert(rq, false);
> >> +		else
> >> +			blk_mq_sched_insert_request(rq, true, false, false);
> >>  	}
> > 
> > Suppose it is one WRITE request to zone device, this way might break
> > the order.
> 
> I'm not sure about this.
> Since the request is dispatched, it should hold and zone write lock.
> And also mq-deadline doesn't have a .requeue_request, zone write lock
> wouldn't be released during requeue.

You are right, looks I misunderstood the zone write lock, sorry for
the noise.

> 
> IMO, this requeue action is similar with what blk_mq_dispatch_rq_list does.
> The latter one also issues the request to underlying driver and requeue rqs
> on dispatch_list if get BLK_STS_SOURCE or BLK_STS_DEV_SOURCE.
> 
> And in addition, RQF_STARTED is set by io scheduler .dispatch_request and
> it could be stop merging as RQF_NOMERGE_FLAGS contains it. 

Yes, that is correct.

Then another question is:

Why don't always requeue request in this way so that it can be simplified
into one code path?

1) in block legacy code, blk_requeue_request() doesn't insert the
request into scheduler queue, and simply put the request into
q->queue_head.

2) blk_mq_requeue_request() is basically run from completion context for
handling very unusual cases(partial completion, error, timeout, ...),
and there shouldn't have benefit to schedule/merge requeued request.

3) RQF_DONTPREP is like a driver private flag, and read/write by driver
only before this patch.

Thanks,
Ming