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;
Subject: Re: [PATCH RFC] scsi: core: remove unsed 'restarts' from scsi_device
To:     Ming Lei <ming.lei@redhat.com>, 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>
References: <20221118113052.1324140-1-yukuai1@huaweicloud.com>
 <cefdae2e-67e3-b4b4-f569-31db960e991f@huaweicloud.com>
 <Y4QqtbXsuYmkOe88@T590>
 <7a747bc3-b902-6f0c-21ef-0ef470ec326e@huaweicloud.com>
 <Y4Q1KPgBsDCZTyTW@T590>
From:   Yu Kuai <yukuai1@huaweicloud.com>
Message-ID: <1caa7d49-0322-3eb9-7e99-6be96f77c80a@huaweicloud.com>
Date:   Mon, 28 Nov 2022 14:08:48 +0800
User-Agent: Mozilla/5.0 (Windows NT 10.0; WOW64; rv:60.0) Gecko/20100101
 Thunderbird/60.8.0
MIME-Version: 1.0
In-Reply-To: <Y4Q1KPgBsDCZTyTW@T590>
Content-Type: text/plain; charset=utf-8; format=flowed
Content-Transfer-Encoding: 8bit
Precedence: bulk

Hi,

在 2022/11/28 12:12, Ming Lei 写道:
> 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().

I still don't get it why all hw queues should be re-run, is this just
for performance or fixing a bug? And I'm not sure this behavior is good
for performance.

> 
>>
>>>
>>>>
>>>> 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.

It's right that is bad, but the point here is to triggered get budget
failed frequently. If I set queue_depth to 2048, and I use 128 numjobs
with 32 iodpeth, performance is still bad.
> 
> 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.

Yes, but current default hctx shared elevator is deadline.
> 
> thanks,
> Ming
> 
> .
>