On 10/13/2017 10:07 AM, Ming Lei wrote: > On Fri, Oct 13, 2017 at 08:44:23AM -0600, Jens Axboe wrote: >> On 10/12/2017 06:19 PM, Ming Lei wrote: >>> On Thu, Oct 12, 2017 at 12:46:24PM -0600, Jens Axboe wrote: >>>> On 10/12/2017 12:37 PM, Ming Lei wrote: >>>>> For SCSI devices, there is often per-request-queue depth, which need >>>>> to be respected before queuing one request. >>>>> >>>>> The current blk-mq always dequeues one request first, then calls .queue_rq() >>>>> to dispatch the request to lld. One obvious issue of this way is that I/O >>>>> merge may not be good, because when the per-request-queue depth can't be >>>>> respected, .queue_rq() has to return BLK_STS_RESOURCE, then this request >>>>> has to staty in hctx->dispatch list, and never got chance to participate >>>>> into I/O merge. >>>>> >>>>> This patch introduces .get_budget and .put_budget callback in blk_mq_ops, >>>>> then we can try to get reserved budget first before dequeuing request. >>>>> Once we can't get budget for queueing I/O, we don't need to dequeue request >>>>> at all, then I/O merge can get improved a lot. >>>> >>>> I can't help but think that it would be cleaner to just be able to >>>> reinsert the request into the scheduler properly, if we fail to >>>> dispatch it. Bart hinted at that earlier as well. >>> >>> Actually when I start to investigate the issue, the 1st thing I tried >>> is to reinsert, but that way is even worse on qla2xxx. >>> >>> Once request is dequeued, the IO merge chance is decreased a lot. >>> With none scheduler, it becomes not possible to merge because >>> we only try to merge over the last 8 requests. With mq-deadline, >>> when one request is reinserted, another request may be dequeued >>> at the same time. >> >> I don't care too much about 'none'. If perfect merging is crucial for >> getting to the performance level you want on the hardware you are using, >> you should not be using 'none'. 'none' will work perfectly fine for NVMe >> etc style devices, where we are not dependent on merging to the same >> extent that we are on other devices. >> >> mq-deadline reinsertion will be expensive, that's in the nature of that >> beast. It's basically the same as a normal request inserition. So for >> that, we'd have to be a bit careful not to run into this too much. Even >> with a dumb approach, it should only happen 1 out of N times, where N is >> the typical point at which the device will return STS_RESOURCE. The >> reinsertion vs dequeue should be serialized with your patch to do that, >> at least for the single queue mq-deadline setup. In fact, I think your >> approach suffers from that same basic race, in that the budget isn't a >> hard allocation, it's just a hint. It can change from the time you check >> it, and when you go and dispatch the IO, if you don't serialize that >> part. So really should be no different in that regard. > > In case of SCSI, the .get_buget is done as atomic counting, > and it is completely effective to avoid unnecessary dequeue, please take > a look at patch 6. Looks like you are right, I had initially misread that as just checking the busy count. But you are actually getting the count at that point, so it should be solid. >>> Not mention the cost of acquiring/releasing lock, that work >>> is just doing useless work and wasting CPU. >> >> Sure, my point is that if it doesn't happen too often, it doesn't really >> matter. It's not THAT expensive. > > Actually it is in hot path, for example, lpfc and qla2xx's queue depth is 3, > it is quite easy to trigger STS_RESOURCE. Ugh, that is low. OK, I think we should just roll with this and see how far we can go. I'll apply it for 4.15. -- Jens Axboe
