Re: [PATCH 3/4] blk-mq: use hw tag for scheduling if hw tag space is big enough

[Omar Sandoval <osandov@osandov.com>]

On Mon, May 01, 2017 at 03:06:16PM +0000, Bart Van Assche wrote:
> On Sat, 2017-04-29 at 18:35 +0800, Ming Lei wrote:
> > On Fri, Apr 28, 2017 at 06:09:40PM +0000, Bart Van Assche wrote:
> > > On Fri, 2017-04-28 at 23:15 +0800, Ming Lei wrote:
> > > > +static inline bool blk_mq_sched_may_use_hw_tag(struct request_queue *q)
> > > > +{
> > > > +	if (q->tag_set->flags & BLK_MQ_F_TAG_SHARED)
> > > > +		return false;
> > > > +
> > > > +	if (blk_mq_get_queue_depth(q) < q->nr_requests)
> > > > +		return false;
> > > > +
> > > > +	return true;
> > > > +}
> > > 
> > > The only user of shared tag sets I know of is scsi-mq. I think it's really
> > > unfortunate that this patch systematically disables BLK_MQ_F_SCHED_USE_HW_TAG
> > > for scsi-mq.
> > 
> > In previous patch, I actually allow driver to pass this flag, but this
> > feature is dropped in this post, just for making it simple & clean.
> > If you think we need it for shared tag set, I can add it in v1.
> > 
> > For shared tag sets, I suggest to not enable it at default, because
> > scheduler is per request queue now, and generaly more requests available,
> > better it performs.  When tags are shared among several request
> > queues, one of them may use tags up for its own scheduling, then
> > starve others. But it should be possible and not difficult to allocate
> > requests fairly for scheduling in this case if we switch to per-hctx
> > scheduling.
> 
> Hello Ming,
> 
> Have you noticed that there is already a mechanism in the block layer to
> avoid starvation if a tag set is shared? The hctx_may_queue() function
> guarantees that each user that shares a tag set gets at least some tags.
> The .active_queues counter keeps track of the number of hardware queues
> that share a tag set.
> 
> Bart.

The scheduler tags are there to abstract away the hardware, and
USE_HW_TAG should just be an optimization for when that abstraction is a
noop. That's not the case when there are shared tags, and I doubt that
the overhead of the scheduler tags is significant for scsi-mq. Let's
stick with the behavior Ming had here.