Re: [PATCH RFC v7 00/12] blk-mq/scsi: Provide hostwide shared tags for SCSI HBAs

[Ming Lei <ming.lei@redhat.com>]

On Mon, Jun 22, 2020 at 08:24:39AM +0200, Hannes Reinecke wrote:
> On 6/17/20 1:26 PM, Kashyap Desai wrote:
> > > 
> > > ->queued is increased only and not decreased just for debug purpose so
> > > ->far, so
> > > it can't be relied for this purpose.
> > 
> > Thanks. I overlooked that that it is only incremental counter.
> > 
> > > 
> > > One approach is to add one similar counter, and maintain it by
> > scheduler's
> > > insert/dispatch callback.
> > 
> > I tried below  and I see performance is on expected range.
> > 
> > diff --git a/block/blk-mq-sched.c b/block/blk-mq-sched.c
> > index fdcc2c1..ea201d0 100644
> > --- a/block/blk-mq-sched.c
> > +++ b/block/blk-mq-sched.c
> > @@ -485,6 +485,7 @@ void blk_mq_sched_insert_request(struct request *rq,
> > bool at_head,
> > 
> >                  list_add(&rq->queuelist, &list);
> >                  e->type->ops.insert_requests(hctx, &list, at_head);
> > +               atomic_inc(&hctx->elevator_queued);
> >          } else {
> >                  spin_lock(&ctx->lock);
> >                  __blk_mq_insert_request(hctx, rq, at_head);
> > @@ -511,8 +512,10 @@ void blk_mq_sched_insert_requests(struct
> > blk_mq_hw_ctx *hctx,
> >          percpu_ref_get(&q->q_usage_counter);
> > 
> >          e = hctx->queue->elevator;
> > -       if (e && e->type->ops.insert_requests)
> > +       if (e && e->type->ops.insert_requests) {
> >                  e->type->ops.insert_requests(hctx, list, false);
> > +               atomic_inc(&hctx->elevator_queued);
> > +       }
> >          else {
> >                  /*
> >                   * try to issue requests directly if the hw queue isn't
> > diff --git a/block/blk-mq-sched.h b/block/blk-mq-sched.h
> > index 126021f..946b47a 100644
> > --- a/block/blk-mq-sched.h
> > +++ b/block/blk-mq-sched.h
> > @@ -74,6 +74,13 @@ static inline bool blk_mq_sched_has_work(struct
> > blk_mq_hw_ctx *hctx)
> >   {
> >          struct elevator_queue *e = hctx->queue->elevator;
> > 
> > +       /* If current hctx has not queued any request, there is no need to
> > run.
> > +        * blk_mq_run_hw_queue() on hctx which has queued IO will handle
> > +        * running specific hctx.
> > +        */
> > +       if (!atomic_read(&hctx->elevator_queued))
> > +               return false;
> > +
> >          if (e && e->type->ops.has_work)
> >                  return e->type->ops.has_work(hctx);
> > 
> > diff --git a/block/blk-mq.c b/block/blk-mq.c
> > index f73a2f9..48f1824 100644
> > --- a/block/blk-mq.c
> > +++ b/block/blk-mq.c
> > @@ -517,8 +517,10 @@ void blk_mq_free_request(struct request *rq)
> >          struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
> > 
> >          if (rq->rq_flags & RQF_ELVPRIV) {
> > -               if (e && e->type->ops.finish_request)
> > +               if (e && e->type->ops.finish_request) {
> >                          e->type->ops.finish_request(rq);
> > +                       atomic_dec(&hctx->elevator_queued);
> > +               }
> >                  if (rq->elv.icq) {
> >                          put_io_context(rq->elv.icq->ioc);
> >                          rq->elv.icq = NULL;
> > @@ -2571,6 +2573,7 @@ blk_mq_alloc_hctx(struct request_queue *q, struct
> > blk_mq_tag_set *set,
> >                  goto free_hctx;
> > 
> >          atomic_set(&hctx->nr_active, 0);
> > +       atomic_set(&hctx->elevator_queued, 0);
> >          if (node == NUMA_NO_NODE)
> >                  node = set->numa_node;
> >          hctx->numa_node = node;
> > diff --git a/include/linux/blk-mq.h b/include/linux/blk-mq.h
> > index 66711c7..ea1ddb1 100644
> > --- a/include/linux/blk-mq.h
> > +++ b/include/linux/blk-mq.h
> > @@ -139,6 +139,10 @@ struct blk_mq_hw_ctx {
> >           * shared across request queues.
> >           */
> >          atomic_t                nr_active;
> > +       /**
> > +        * @elevator_queued: Number of queued requests on hctx.
> > +        */
> > +       atomic_t                elevator_queued;
> > 
> >          /** @cpuhp_online: List to store request if CPU is going to die */
> >          struct hlist_node       cpuhp_online;
> > 
> > 
> > 
> Would it make sense to move it into the elevator itself?

That is my initial suggestion, and the counter is just done for bfq &
mq-deadline, then we needn't to pay the cost for others.

Thanks,
Ming