Re: [PATCH 4/4] nvme: add support for mq_ops->queue_rqs()

[Max Gurtovoy <mgurtovoy@nvidia.com>]

On 12/16/2021 6:25 PM, Jens Axboe wrote:
> On 12/16/21 9:19 AM, Max Gurtovoy wrote:
>> On 12/16/2021 6:05 PM, Jens Axboe wrote:
>>> On 12/16/21 9:00 AM, Max Gurtovoy wrote:
>>>> On 12/16/2021 5:48 PM, Jens Axboe wrote:
>>>>> On 12/16/21 6:06 AM, Max Gurtovoy wrote:
>>>>>> On 12/16/2021 11:08 AM, Christoph Hellwig wrote:
>>>>>>> On Wed, Dec 15, 2021 at 09:24:21AM -0700, Jens Axboe wrote:
>>>>>>>> +	spin_lock(&nvmeq->sq_lock);
>>>>>>>> +	while (!rq_list_empty(*rqlist)) {
>>>>>>>> +		struct request *req = rq_list_pop(rqlist);
>>>>>>>> +		struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
>>>>>>>> +
>>>>>>>> +		memcpy(nvmeq->sq_cmds + (nvmeq->sq_tail << nvmeq->sqes),
>>>>>>>> +				absolute_pointer(&iod->cmd), sizeof(iod->cmd));
>>>>>>>> +		if (++nvmeq->sq_tail == nvmeq->q_depth)
>>>>>>>> +			nvmeq->sq_tail = 0;
>>>>>>> So this doesn't even use the new helper added in patch 2?  I think this
>>>>>>> should call nvme_sq_copy_cmd().
>>>>>> I also noticed that.
>>>>>>
>>>>>> So need to decide if to open code it or use the helper function.
>>>>>>
>>>>>> Inline helper sounds reasonable if you have 3 places that will use it.
>>>>> Yes agree, that's been my stance too :-)
>>>>>
>>>>>>> The rest looks identical to the incremental patch I posted, so I guess
>>>>>>> the performance degration measured on the first try was a measurement
>>>>>>> error?
>>>>>> giving 1 dbr for a batch of N commands sounds good idea. Also for RDMA host.
>>>>>>
>>>>>> But how do you moderate it ? what is the batch_sz <--> time_to_wait
>>>>>> algorithm ?
>>>>> The batching is naturally limited at BLK_MAX_REQUEST_COUNT, which is 32
>>>>> in total. I do agree that if we ever made it much larger, then we might
>>>>> want to cap it differently. But 32 seems like a pretty reasonable number
>>>>> to get enough gain from the batching done in various areas, while still
>>>>> not making it so large that we have a potential latency issue. That
>>>>> batch count is already used consistently for other items too (like tag
>>>>> allocation), so it's not specific to just this one case.
>>>> I'm saying that the you can wait to the batch_max_count too long and it
>>>> won't be efficient from latency POV.
>>>>
>>>> So it's better to limit the block layar to wait for the first to come: x
>>>> usecs or batch_max_count before issue queue_rqs.
>>> There's no waiting specifically for this, it's just based on the plug.
>>> We just won't do more than 32 in that plug. This is really just an
>>> artifact of the plugging, and if that should be limited based on "max of
>>> 32 or xx time", then that should be done there.
>>>
>>> But in general I think it's saner and enough to just limit the total
>>> size. If we spend more than xx usec building up the plug list, we're
>>> doing something horribly wrong. That really should not happen with 32
>>> requests, and we'll never eg wait on requests if we're out of tags. That
>>> will result in a plug flush to begin with.
>> I'm not aware of the plug. I hope to get to it soon.
>>
>> My concern is if the user application submitted only 28 requests and
>> then you'll wait forever ? or for very long time.
>>
>> I guess not, but I'm asking how do you know how to batch and when to
>> stop in case 32 commands won't arrive anytime soon.
> The plug is in the stack of the task, so that condition can never
> happen. If the application originally asks for 32 but then only submits
> 28, then once that last one is submitted the plug is flushed and
> requests are issued.

So if I'm running fio with --iodepth=28 what will plug do ? send batches 
of 28 ? or 1 by 1 ?

>>>> Also, This batch is per HW queue or SW queue or the entire request queue ?
>>> It's per submitter, so whatever the submitter ends up queueing IO
>>> against. In general it'll be per-queue.
>> struct request_queue ?
>>
>> I think the best is to batch per struct blk_mq_hw_ctx.
>>
>> I see that you check this in the nvme_pci driver but shouldn't it go to
>> the block layer ?
> That's not how plugging works. In general, unless your task bounces
> around, then it'll be a single queue and a single hw queue as well.
> Adding code to specifically check the mappings and flush at that point
> would be a net loss, rather than just deal with it if it happens for
> some cases.
>