> On Wed, Jul 22, 2020 at 11:00:45AM +0530, Kashyap Desai wrote: > > > On Tue, Jul 21, 2020 at 12:23:39PM +0530, Kashyap Desai wrote: > > > > > > > > > > > > > > Perf top (shared host tag. IOPS = 230K) > > > > > > > > > > > > > > 13.98% [kernel] [k] sbitmap_any_bit_set > > > > > > > 6.43% [kernel] [k] blk_mq_run_hw_queue > > > > > > > > > > > > blk_mq_run_hw_queue function take more CPU which is called from > " > > > > > > scsi_end_request" > > > > > > > > > > The problem could be that nr_hw_queues is increased a lot so > > > > > that sample > > > > on > > > > > blk_mq_run_hw_queue() can be observed now. > > > > > > > > Yes. That is correct. > > > > > > > > > > > > > > > It looks like " blk_mq_hctx_has_pending" handles only elevator > > > > > > (scheduler) case. If queue has ioscheduler=none, we can skip. > > > > > > I case of scheduler=none, IO will be pushed to hardware queue > > > > > > and it by pass > > > > > software queue. > > > > > > Based on above understanding, I added below patch and I can > > > > > > see performance scale back to expectation. > > > > > > > > > > > > Ming mentioned that - we cannot remove blk_mq_run_hw_queues() > > > from > > > > > > IO completion path otherwise we may see IO hang. So I have > > > > > > just modified completion path assuming it is only required for > > > > > > IO > > scheduler > > > case. > > > > > > https://www.spinics.net/lists/linux-block/msg55049.html > > > > > > > > > > > > Please review and let me know if this is good or we have to > > > > > > address with proper fix. > > > > > > > > > > > > diff --git a/block/blk-mq.c b/block/blk-mq.c index > > > > > > 1be7ac5a4040..b6a5b41b7fc2 100644 > > > > > > --- a/block/blk-mq.c > > > > > > +++ b/block/blk-mq.c > > > > > > @@ -1559,6 +1559,9 @@ void blk_mq_run_hw_queues(struct > > > > > request_queue > > > > > > *q, bool async) > > > > > > struct blk_mq_hw_ctx *hctx; > > > > > > int i; > > > > > > > > > > > > + if (!q->elevator) > > > > > > + return; > > > > > > + > > > > > > > > > > This way shouldn't be correct, blk_mq_run_hw_queues() is still > > > > > needed > > > > for > > > > > none because request may not be dispatched successfully by > > > > > direct > > issue. > > > > > > > > When block layer attempt posting request to h/w queue directly > > > > (for > > > > ioscheduler=none) and if it fails, it is calling > > > > blk_mq_request_bypass_insert(). > > > > blk_mq_request_bypass_insert function will start the h/w queue > > > > from submission context. Do we still have an issue if we skip > > > > running hw queue from completion ? > > > > > > The thing is that we can't guarantee that direct issue or adding > > > request > > into > > > hctx->dispatch is always done for MQ/none, for example, request > > > hctx->still > > > can be added to sw queue from blk_mq_flush_plug_list() when mq plug > > > is applied. > > > > I see even blk_mq_sched_insert_requests() from blk_mq_flush_plug_list > > make sure it run the h/w queue. If all the submission path which deals > > with s/w queue make sure they run h/w queue, can't we remove > > blk_mq_run_hw_queues() from scsi_end_request ? > > No, one purpose of blk_mq_run_hw_queues() is for rerun queue in case that > dispatch budget is running out of in submission path, and sdev->device_busy > is shared by all hw queues on this scsi device. > > I posted one patch for avoiding it in scsi_end_request() before, looks it never > lands upstream: > > https://lore.kernel.org/linux-block/20191118100640.3673-1- > ming.lei@xxxxxxxxxx/ Ming - I think above patch will fix the issue of performance on VD. I fix some hunk failure and ported to 5.8 kernel - I am testing this patch on my setup. If you post V4, I will use that. So far looks good. I have reduced device queue depth so that I hit budget busy code path frequently. Kashyap > > Thanks, > Ming
