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

[Max Gurtovoy]

On 12/20/2021 8:58 PM, Jens Axboe wrote:
> On 12/20/21 11:48 AM, Max Gurtovoy wrote:
>> On 12/20/2021 6:34 PM, Jens Axboe wrote:
>>> On 12/20/21 8:29 AM, Max Gurtovoy wrote:
>>>> On 12/20/2021 4:19 PM, Jens Axboe wrote:
>>>>> On 12/20/21 3:11 AM, Max Gurtovoy wrote:
>>>>>> On 12/19/2021 4:48 PM, Jens Axboe wrote:
>>>>>>> On 12/19/21 5:14 AM, Max Gurtovoy wrote:
>>>>>>>> On 12/16/2021 7:16 PM, Jens Axboe wrote:
>>>>>>>>> On 12/16/21 9:57 AM, Max Gurtovoy wrote:
>>>>>>>>>> On 12/16/2021 6:36 PM, Jens Axboe wrote:
>>>>>>>>>>> On 12/16/21 9:34 AM, Max Gurtovoy wrote:
>>>>>>>>>>>> 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 ?
>>>>>>>>>>> --iodepth just controls the overall depth, the batch submit count
>>>>>>>>>>> dictates what happens further down. If you run queue depth 28 and submit
>>>>>>>>>>> one at the time, then you'll get one at the time further down too. Hence
>>>>>>>>>>> the batching is directly driven by what the application is already
>>>>>>>>>>> doing.
>>>>>>>>>> I see. Thanks for the explanation.
>>>>>>>>>>
>>>>>>>>>> So it works only for io_uring based applications ?
>>>>>>>>> It's only enabled for io_uring right now, but it's generically available
>>>>>>>>> for anyone that wants to use it... Would be trivial to do for aio, and
>>>>>>>>> other spots that currently use blk_start_plug() and has an idea of how
>>>>>>>>> many IOs will be submitted
>>>>>>>> Can you please share an example application (or is it fio patches) that
>>>>>>>> can submit batches ? The same that was used to test this patchset is
>>>>>>>> fine too.
>>>>>>>>
>>>>>>>> I would like to test it with our NVMe SNAP controllers and also to
>>>>>>>> develop NVMe/RDMA queue_rqs code and test the perf with it.
>>>>>>> You should just be able to use iodepth_batch with fio. For my peak
>>>>>>> testing, I use t/io_uring from the fio repo. By default, it'll run QD of
>>>>>>> and do batches of 32 for complete and submit. You can just run:
>>>>>>>
>>>>>>> t/io_uring <dev or file>
>>>>>>>
>>>>>>> maybe adding -p0 for IRQ driven rather than polled IO.
>>>>>> I used your block/for-next branch and implemented queue_rqs in NVMe/RDMA
>>>>>> but it was never called using the t/io_uring test nor fio with
>>>>>> iodepth_batch=32 flag with io_uring engine.
>>>>>>
>>>>>> Any idea what might be the issue ?
>>>>>>
>>>>>> I installed fio from sources..
>>>>> The two main restrictions right now are a scheduler and shared tags, are
>>>>> you using any of those?
>>>> No.
>>>>
>>>> But maybe I'm missing the .commit_rqs callback. is it mandatory for this
>>>> feature ?
>>> I've only tested with nvme pci which does have it, but I don't think so.
>>> Unless there's some check somewhere that makes it necessary. Can you
>>> share the patch you're currently using on top?
>> The attached POC patches apply cleanly on block/for-next branch
> Looks reasonable to me from a quick glance. Not sure why you're not
> seeing it hit, maybe try and instrument
> block/blk-mq.c:blk_mq_flush_plug_list() and find out why it isn't being
> called? As mentioned, no elevator or shared tags, should work for
> anything else basically.

Yes. I saw that the blk layer converted the original non-shared tagset 
of NVMe/RDMA to a shared one because of the nvmf connect request queue 
that is using the same tagset (uses only the reserved tag).

So I guess this is the reason that the I couldn't reach the new code of 
queue_rqs.

The question is how we can overcome this ?

Should we create new tagset for the NVMf fabrics connect_q ? or maybe 
not mark the tagset as shared for reserved ids ?

Christoph, any suggestion here ?

>