Re: [PATCH V4 0/5] blk-mq: improvement on handling IO during CPU hotplug

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

On Fri, Oct 25, 2019 at 05:33:35PM +0100, John Garry wrote:
> > > There might be two reasons:
> > > 
> > > 1) You are still testing a multiple reply-queue device?
> > 
> > As before, I am testing by exposing mutliple queues to the SCSI
> > midlayer. I had to make this change locally, as on mainline we still
> > only expose a single queue and use the internal reply queue when
> > enabling managed interrupts.
> > 
> > As I
> > > mentioned last times, it is hard to map reply-queue into blk-mq
> > > hctx correctly.
> > 
> > Here's my branch, if you want to check:
> > 
> > https://github.com/hisilicon/kernel-dev/commits/private-topic-sas-5.4-mq-v4
> > 
> > It's a bit messy (sorry), but you can see that the reply-queue in the
> > LLDD is removed in commit 087b95af374.
> > 
> > I am now thinking of actually making this change to the LLDD in mainline
> > to avoid any doubt in future.
> > 
> > > 
> > > 2) concurrent dispatch to device, which can be observed by the
> > > following patch.
> > > 
> > > diff --git a/block/blk-mq.c b/block/blk-mq.c
> > > index 06081966549f..3590f6f947eb 100644
> > > --- a/block/blk-mq.c
> > > +++ b/block/blk-mq.c
> > > @@ -679,6 +679,8 @@ void blk_mq_start_request(struct request *rq)
> > >  {
> > >         struct request_queue *q = rq->q;
> > > 
> > > +       WARN_ON_ONCE(test_bit(BLK_MQ_S_INTERNAL_STOPPED,
> > > &rq->mq_hctx->state));
> > > +
> > >         trace_block_rq_issue(q, rq);
> > > 
> > >         if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
> > > 
> > > However, I think it is hard to be 2#, since the current CPU is the last
> > > CPU in hctx->cpu_mask.
> > > 
> > 
> > I'll try it.
> > 
> 
> Hi Ming,
> 
> I am looking at this issue again.
> 
> I am using https://lore.kernel.org/linux-scsi/1571926881-75524-1-git-send-email-john.garry@huawei.com/T/#t
> with expose_mq_experimental set. I guess you're going to say that this
> series is wrong, but I think it's ok for this purpose.
> 
> Forgetting that for a moment, maybe I have found an issue.
> 
> For the SCSI commands which timeout, I notice that
> scsi_set_blocked(reason=SCSI_MLQUEUE_EH_RETRY) was called 30 seconds
> earlier.
> 
>  scsi_set_blocked+0x20/0xb8
>  __scsi_queue_insert+0x40/0x90
>  scsi_softirq_done+0x164/0x1c8
>  __blk_mq_complete_request_remote+0x18/0x20
>  flush_smp_call_function_queue+0xa8/0x150
>  generic_smp_call_function_single_interrupt+0x10/0x18
>  handle_IPI+0xec/0x1a8
>  arch_cpu_idle+0x10/0x18
>  do_idle+0x1d0/0x2b0
>  cpu_startup_entry+0x24/0x40
>  secondary_start_kernel+0x1b4/0x208

Could you investigate a bit the reason why timeout is triggered?

Especially we suppose to drain all in-flight requests before the
last CPU of this hctx becomes offline, and it shouldn't be caused by
the hctx becoming dead, so still need you to confirm that all
in-flight requests are really drained in your test. Or is it still
possible to dispatch to LDD after BLK_MQ_S_INTERNAL_STOPPED is set?

In theory, it shouldn't be possible, given we drain in-flight request
on the last CPU of this hctx.

Or blk_mq_hctx_next_cpu() may still run WORK_CPU_UNBOUND schedule after
all CPUs are offline, could you add debug message in that branch?

> 
> I also notice that the __scsi_queue_insert() call, above, seems to retry to
> requeue the request on a dead rq in calling
> __scsi_queue_insert()->blk_mq_requeue_requet()->__blk_mq_requeue_request(),
> ***:
> 
> [ 1185.235243] psci: CPU1 killed.
> [ 1185.238610] blk_mq_hctx_notify_dead cpu1 dead
> request_queue=0xffff0023ace24f60 (id=19)
> [ 1185.246530] blk_mq_hctx_notify_dead cpu1 dead
> request_queue=0xffff0023ace23f80 (id=17)
> [ 1185.254443] blk_mq_hctx_notify_dead cpu1 dead
> request_queue=0xffff0023ace22fa0 (id=15)
> [ 1185.262356] blk_mq_hctx_notify_dead cpu1 dead
> request_queue=0xffff0023ace21fc0 (id=13)***
> [ 1185.270271] blk_mq_hctx_notify_dead cpu1 dead
> request_queue=0xffff0023ace20fe0 (id=11)
> [ 1185.939451] scsi_softirq_done NEEDS_RETRY rq=0xffff0023b7416000
> [ 1185.945359] scsi_set_blocked reason=0x1057
> [ 1185.949444] __blk_mq_requeue_request request_queue=0xffff0023ace21fc0
> id=13 rq=0xffff0023b7416000***
> 
> [...]
> 
> [ 1214.903455] scsi_timeout req=0xffff0023add29000 reserved=0
> [ 1214.908946] scsi_timeout req=0xffff0023add29300 reserved=0
> [ 1214.914424] scsi_timeout req=0xffff0023add29600 reserved=0
> [ 1214.919909] scsi_timeout req=0xffff0023add29900 reserved=0
> 
> I guess that we're retrying as the SCSI failed in the LLDD for some reason.
> 
> So could this be the problem - we're attempting to requeue on a dead request
> queue?

If there are any in-flight requests originated from hctx which is going
to become dead, they should have been drained before CPU becomes offline.

Thanks, 
Ming