Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Mon, 28 Nov 2022 12:12:24 +0800
From:   Ming Lei <ming.lei@redhat.com>
To:     Yu Kuai <yukuai1@huaweicloud.com>
Cc:     jejb@linux.ibm.com, martin.petersen@oracle.com,
        linux-scsi@vger.kernel.org, linux-kernel@vger.kernel.org,
        yi.zhang@huawei.com, "yukuai (C)" <yukuai3@huawei.com>
Subject: Re: [PATCH RFC] scsi: core: remove unsed 'restarts' from scsi_device
Message-ID: <Y4Q1KPgBsDCZTyTW@T590>
References: <20221118113052.1324140-1-yukuai1@huaweicloud.com>
 <cefdae2e-67e3-b4b4-f569-31db960e991f@huaweicloud.com>
 <Y4QqtbXsuYmkOe88@T590>
 <7a747bc3-b902-6f0c-21ef-0ef470ec326e@huaweicloud.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
Content-Transfer-Encoding: 8bit
In-Reply-To: <7a747bc3-b902-6f0c-21ef-0ef470ec326e@huaweicloud.com>
Precedence: bulk

On Mon, Nov 28, 2022 at 11:35:18AM +0800, Yu Kuai wrote:
> 
> 
> 在 2022/11/28 11:27, Ming Lei 写道:
> > On Sat, Nov 26, 2022 at 04:54:46PM +0800, Yu Kuai wrote:
> > > Hi,
> > > 
> > > 在 2022/11/18 19:30, Yu Kuai 写道:
> > > > From: Yu Kuai <yukuai3@huawei.com>
> > > > 
> > > > During code review, I found that 'restarts' is not useful anymore after
> > > > the following commits:
> > > > 
> > > > 1) commit ab3cee3762e5 ("blk-mq: In blk_mq_dispatch_rq_list() "no budget"
> > > > is a reason to kick")
> > > > 2) commit d3b38596875d ("blk-mq: run queue no matter whether the request
> > > > is the last request")
> > > > 3) commit 673235f91531 ("scsi: core: Fix race between handling STS_RESOURCE
> > > > and completion")
> > > > 
> > > > Now that if get budget ever failed, block layer will make sure to
> > > > trigger new run queue for the hctx. Hence there is no need to run queue
> > > > from scsi layer in this case.
> > > > 
> > 
> > But scsi_run_queue_async() needs to run all hw queue because budget is
> > actually LUN/request queue wide.
> 
> Why the hw queue need to run if get budget never failed in this hw
> queue?

Because all hw queues share the queue wide budget, and once budget
is available, all hw queues are re-run, and the hw queue won't be
scheduled actually if there is nothing to run, see
blk_mq_run_hw_queue().

> 
> > 
> > > 
> > > Does anyone has suggestions about this patch?
> > > 
> > > More info why I tried to remove this:
> > > 
> > > while testing megaraid with 4 nvme with none elevator, the default
> > > queue_depth is 128, while I test it with fio 128 jobs and 1 iodepth,
> > > bw is about 4Gib/s, however, if I test with 128 jobs and 2 iodepth,
> > > bw is decreased to about 0.8Gib/s, and with this patch applied,
> > > bw can stay 4Gib/s in the later case.
> > 
> > What is .can_queue and nr_hw_queues in your setting?
> test cmd:
> fio -name=0 -ioengine=libaio -direct=1 -group_reporting=1 -randseed=2022
> -rwmixread=70 -refill_buffers -filename=/dev/sdg -numjobs=128 -size=1TB
> -runtime=60s -bs=4k -iodepth=2 -rw=randwrite
> 
> test environment:
> arm64 Kunpeng-920, 128 cpu
> megaraid with 4 NVMEs, 128 hctx and queue_depth is 128

From your setting, megaraid should sets ->host_tagset, that said there
is only 128 tags for all 4 NVMEs(128 hw queue shares the all 128 tags
too).

That looks one really bad setting.

BTW, why do you drive nvme via megaraid instead nvme driver?

> And by the way, after Jan's patch "blk-mq: Improve performance of non-mq
> IO schedulers with multiple HW queues", scsi_run_queue_async() can only
> garantee to run hw queue for the current cpu, not all the hw queues.

That isn't true, each hctx is still run in case of none & kyber scheduler.

thanks,
Ming