Re: commit 01e99aeca397 causes longer runtime of block/004

From: Shinichiro Kawasaki <shinichiro.kawasaki@wdc.com>
To: Ming Lei <ming.lei@redhat.com>
Cc: "linux-block@vger.kernel.org" <linux-block@vger.kernel.org>,
	Jens Axboe <axboe@kernel.dk>,
	Damien Le Moal <Damien.LeMoal@wdc.com>
Subject: Re: commit 01e99aeca397 causes longer runtime of block/004
Date: Thu, 5 Mar 2020 01:19:02 +0000	[thread overview]
Message-ID: <20200305011900.2rtgtmclovmr2fbw@shindev.dhcp.fujisawa.hgst.com> (raw)
In-Reply-To: <20200304095344.GA10390@ming.t460p>

On Mar 04, 2020 / 17:53, Ming Lei wrote:
> On Wed, Mar 04, 2020 at 06:11:37AM +0000, Shinichiro Kawasaki wrote:
> > On Mar 04, 2020 / 11:46, Ming Lei wrote:
> > > On Wed, Mar 04, 2020 at 02:38:43AM +0000, Shinichiro Kawasaki wrote:
> > > > I noticed that blktests block/004 takes longer runtime with 5.6-rc4 than
> > > > 5.6-rc3, and found that the commit 01e99aeca397 ("blk-mq: insert passthrough
> > > > request into hctx->dispatch directly") triggers it.
> > > > 
> > > > The longer runtime was observed with dm-linear device which maps SATA SMR HDD
> > > > connected via AHCI. It was not observed with dm-linear on SAS/SATA SMR HDDs
> > > > connected via SAS-HBA. Not observed with dm-linear on non-SMR HDDs either.
> > > > 
> > > > Before the commit, block/004 took around 130 seconds. After the commit, it takes
> > > > around 300 seconds. I need to dig in further details to understand why the
> > > > commit makes the test case longer.
> > > > 
> > > > The test case block/004 does "flush intensive workload". Is this longer runtime
> > > > expected?
> > > 
> > > The following patch might address this issue:
> > > 
> > > https://lore.kernel.org/linux-block/20200207190416.99928-1-sqazi@google.com/#t
> > > 
> > > Please test and provide us the result.
> > > 
> > > thanks,
> > > Ming
> > >
> > 
> > Hi Ming,
> > 
> > I applied the patch to 5.6-rc4 but I observed the longer runtime of block/004.
> > Still it takes around 300 seconds.
> 
> Hello Shinichiro,
> 
> block/004 only sends 1564 sync randwrite, and seems 130s has been slow
> enough.
> 
> There are two related effect in that commit for your issue:
> 
> 1) 'at_head' is applied in blk_mq_sched_insert_request() for flush
> request
> 
> 2) all IO is added back to tail of hctx->dispatch after .queue_rq()
> returns STS_RESOURCE
> 
> Seems it is more related with 2) given you can't reproduce the issue on 
> SAS.
> 
> So please test the following two patches, and see which one makes a
> difference for you.
> 
> BTW, both two looks not reasonable, just for narrowing down the issue.
> 
> 1) patch 1
> 
> diff --git a/block/blk-mq-sched.c b/block/blk-mq-sched.c
> index 856356b1619e..86137c75283c 100644
> --- a/block/blk-mq-sched.c
> +++ b/block/blk-mq-sched.c
> @@ -398,7 +398,7 @@ void blk_mq_sched_insert_request(struct request *rq, bool at_head,
>  	WARN_ON(e && (rq->tag != -1));
>  
>  	if (blk_mq_sched_bypass_insert(hctx, !!e, rq)) {
> -		blk_mq_request_bypass_insert(rq, at_head, false);
> +		blk_mq_request_bypass_insert(rq, true, false);
>  		goto run;
>  	}

Ming, thank you for the trial patches.
This "patch 1" reduced the runtime, as short as rc3.

> 
> 
> 2) patch 2
> diff --git a/block/blk-mq.c b/block/blk-mq.c
> index d92088dec6c3..447d5cb39832 100644
> --- a/block/blk-mq.c
> +++ b/block/blk-mq.c
> @@ -1286,7 +1286,7 @@ 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_tail_init(list, &hctx->dispatch);
> +		list_splice_init(list, &hctx->dispatch);
>  		spin_unlock(&hctx->lock);
>  
>  		/*

This patch 2 didn't reduce the runtime.

Wish this report helps.

-- 
Best Regards,
Shin'ichiro Kawasaki