Re: [PATCH] blk-mq: insert passthrough request into hctx->dispatch directly

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

On Wed, Feb 19, 2020 at 03:47:50PM -0800, dongli.zhang@oracle.com wrote:
> 
> 
> On 2/19/20 2:10 PM, Ming Lei wrote:
> > On Wed, Feb 19, 2020 at 08:36:15AM -0800, Christoph Hellwig wrote:
> >> On Sat, Feb 15, 2020 at 11:21:40AM +0800, Ming Lei wrote:
> >>> For some reason, device may be in one situation which can't handle
> >>> FS request, so STS_RESOURCE is always returned and the FS request
> >>> will be added to hctx->dispatch. However passthrough request may
> >>> be required at that time for fixing the problem. If passthrough
> >>> request is added to scheduler queue, there isn't any chance for
> >>> blk-mq to dispatch it given we prioritize requests in hctx->dispatch.
> >>> Then the FS IO request may never be completed, and IO hang is caused.
> >>>
> >>> So passthrough request has to be added to hctx->dispatch directly.
> >>>
> >>> Fix this issue by inserting passthrough request into hctx->dispatch
> >>> directly. Then it becomes consistent with original legacy IO request
> >>> path, in which passthrough request is always added to q->queue_head.
> >>
> >> Do you have a description of an actual problem this fixes?  Maybe even
> >> a reproducer for blktests?
> >>
> > 
> > It is reported by one RH customer in the following test case:
> > 
> > 	1) Start IO on Emulex FC host
> > 	2) Fail one controller, wait 5 minutes
> > 	3) Bring controller back online
> > 
> > When we trace the problem, it is found that FS request started in device_add_disk()
> > from scsi disk probe context stuck because scsi_queue_rq() always return
> > STS_BUSY via scsi_setup_fs_cmnd() -> alua_prep_fn().
> > 
> > The kernel ALUA state is TRANSITIONING at that time, so it is reasonable to see
> > BLK_TYPE_FS requests won't go anywhere because of the check in alua_prep_fn().
> > 
> > However, the passthrough request(TEST UNIT READY) is submitted from alua_rtpg_work
> > when the FS request can't be dispatched to LLD. And SCSI stack should
> > have been allowed to handle this passthrough rquest. But it can't reach SCSI stack
> > via .queue_rq() because blk-mq won't dispatch it until hctx->dispatch is
> > empty.
> > 
> > The legacy IO request code always added passthrough request into head of q->queue_head
> > directly instead of scheduler queue or sw queue, so no such issue.
> > 
> > So far not figured out one blktests test case, but the problem is real.
> > 
> > BTW, I just found we need the extra following change:
> > 
> > @@ -1301,7 +1301,7 @@ bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
> >                         q->mq_ops->commit_rqs(hctx);
> > 
> >                 spin_lock(&hctx->lock);
> > -               list_splice_init(list, &hctx->dispatch);
> > +               list_splice_tail_init(list, &hctx->dispatch);
> >                 spin_unlock(&hctx->lock);
> > 
> 
> Is it fine to add to tail as the requests on dispatch would be reordered?

Wrt. FS request:

Firstly we never guarantee that the request is dispatched in order.

Secondly and more importantly, request can be added into hctx->dispatch
in any order. One usual case is that request is added to hctx->dispatch
concurrently when .queue_rq() fails. On the other side, in case of not
concurrent adding to hctx->dispatch, after one request is added to
hctx->dispatch, we always dispatch request from hctx->dispatch first,
instead of dequeuing request from scheduler queue and adding them to
hctx->dispatch again after .queue_rq() fails.

> 
> A, B, C and D are on the list. Suppose A is failed and the new list would become
> B, C D, A?

Right, I don't see there is any issue in this way, do you see issues?



Thanks,
Ming