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

From: John Garry <john.garry@huawei.com>
To: Ming Lei <ming.lei@redhat.com>
Cc: Jens Axboe <axboe@kernel.dk>, <linux-block@vger.kernel.org>,
	"Bart Van Assche" <bvanassche@acm.org>,
	Hannes Reinecke <hare@suse.com>, "Christoph Hellwig" <hch@lst.de>,
	Thomas Gleixner <tglx@linutronix.de>,
	Keith Busch <keith.busch@intel.com>
Subject: Re: [PATCH V4 0/5] blk-mq: improvement on handling IO during CPU hotplug
Date: Tue, 22 Oct 2019 12:19:17 +0100	[thread overview]
Message-ID: <4a42a062-8bca-9e03-c158-1c149986d383@huawei.com> (raw)
In-Reply-To: <20191022001613.GA32193@ming.t460p>

On 22/10/2019 01:16, Ming Lei wrote:
> On Mon, Oct 21, 2019 at 03:02:56PM +0100, John Garry wrote:
>> On 21/10/2019 13:53, Ming Lei wrote:
>>> On Mon, Oct 21, 2019 at 12:49:53PM +0100, John Garry wrote:
>>>>>>>
>>>>>>
>>>>>> Yes, we share tags among all queues, but we generate the tag - known as IPTT
>>>>>> - in the LLDD now, as we can no longer use the request tag (as it is not
>>>>>> unique per all queues):
>>>>>>
>>>>>> https://github.com/hisilicon/kernel-dev/commit/087b95af374be6965583c1673032fb33bc8127e8#diff-f5d8fff19bc539a7387af5230d4e5771R188
>>>>>>
>>>>>> As I said, the branch is messy and I did have to fix 087b95af374.
>>>>>
>>>>> Firstly this way may waste lots of memory, especially the queue depth is
>>>>> big, such as, hisilicon V3's queue depth is 4096.
>>>>>
>>>>> Secondly, you have to deal with queue busy efficiently and correctly,
>>>>> for example, your real hw tags(IPTT) can be used up easily, and how
>>>>> will you handle these dispatched request?
>>>>
>>>> I have not seen scenario of exhausted IPTT. And IPTT count is same as SCSI
>>>> host.can_queue, so SCSI midlayer should ensure that this does not occur.
>>>
>>
>> Hi Ming,

Hi Ming,

>>
>>> That check isn't correct, and each hw queue should have allowed
>>> .can_queue in-flight requests.
>>
>> There always seems to be some confusion or disagreement on this topic.
>>
>> I work according to the comment in scsi_host.h:
>>
>> "Note: it is assumed that each hardware queue has a queue depth of
>>   can_queue. In other words, the total queue depth per host
>>   is nr_hw_queues * can_queue."
>>
>> So I set Scsi_host.can_queue = HISI_SAS_MAX_COMMANDS (=4096)
> 
> I believe all current drivers set .can_queue as single hw queue's depth.
> If you set .can_queue as HISI_SAS_MAX_COMMANDS which is HBA's queue
> depth, the hisilicon sas driver will HISI_SAS_MAX_COMMANDS * nr_hw_queues
> in-flight requests.

Yeah, but the SCSI host should still limit max IOs over all queues to 
.can_queue:

https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/Documentation/scsi/scsi_mid_low_api.txt#n1083

> 
>>
>>>
>>>>
>>>>>
>>>>> Finally, you have to evaluate the performance effect, this is highly
>>>>> related with how to deal with out-of-IPTT.
>>>>
>>>> Some figures from our previous testing:
>>>>
>>>> Managed interrupt without exposing multiple queues: 3M IOPs
>>>> Managed interrupt with exposing multiple queues: 2.6M IOPs
>>>
>>> Then you see the performance regression.
>>
>> Let's discuss this when I send the patches, so we don't get sidetracked on
>> this blk-mq improvement topic.
> 
> OK, what I meant is to use correct driver to test the patches, otherwise
> it might be hard to investigate.

Of course. I'm working on this now, and it looks like it will turn out 
complicated... you'll see.

BTW, I reran the test and never say the new WARN trigger (while SCSI 
timeouts did occur).

Thanks again,
John

> 
> 
> Thanks,
> Ming
> 
> .
> 