Re: [RFC PATCH v3 3/3] blk-mq: Lockout tagset iterator when exiting elevator

[John Garry]

On 06/03/2021 04:43, Bart Van Assche wrote:
> On 3/5/21 7:14 AM, John Garry wrote:
> > diff --git a/block/blk-mq-tag.c b/block/blk-mq-tag.c
> > index 7ff1b20d58e7..5950fee490e8 100644
> > --- a/block/blk-mq-tag.c
> > +++ b/block/blk-mq-tag.c
> > @@ -358,11 +358,16 @@ void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset,
> >   {
> >   	int i;
> >   
> > +	if (!atomic_inc_not_zero(&tagset->iter_usage_counter))
> > +		return;
> > +
> >   	for (i = 0; i < tagset->nr_hw_queues; i++) {
> >   		if (tagset->tags && tagset->tags[i])
> >   			__blk_mq_all_tag_iter(tagset->tags[i], fn, priv,
> >   					      BT_TAG_ITER_STARTED);
> >   	}
> > +
> > +	atomic_dec(&tagset->iter_usage_counter);
> >   }
> >   EXPORT_SYMBOL(blk_mq_tagset_busy_iter);

Hi Bart,

> This changes the behavior of blk_mq_tagset_busy_iter(). What will e.g.
> happen if the mtip driver calls blk_mq_tagset_busy_iter(&dd->tags,
> mtip_abort_cmd, dd) concurrently with another blk_mq_tagset_busy_iter()
> call and if that causes all mtip_abort_cmd() calls to be skipped?

I'm not sure that I understand this problem you describe. So if 
blk_mq_tagset_busy_iter(&dd->tags, mtip_abort_cmd, dd) is called, either 
can happen:
a. normal operation, iter_usage_counter initially holds >= 1, and then 
iter_usage_counter is incremented in blk_mq_tagset_busy_iter() and we 
iter the busy tags. Any parallel call to blk_mq_tagset_busy_iter() will 
also increase iter_usage_counter.
b. we're switching IO scheduler. In this scenario, first we quiesce all 
queues. After that, there should be no active requests. At that point, 
we ensure any calls to blk_mq_tagset_busy_iter() are finished and block 
(or discard may be a better term) any more calls. Blocking any more 
calls should be safe as there are no requests to iter. atomic_cmpxchg() 
is used to set iter_usage_counter to 0, blocking any more calls.

> > +	while (atomic_cmpxchg(&set->iter_usage_counter, 1, 0) != 1);
> Isn't it recommended to call cpu_relax() inside busy-waiting loops?

Maybe, but I am considering changing this patch to use percpu_refcnt() - 
I need to check it further.

> >   	blk_mq_sched_free_requests(q);
> >   	__elevator_exit(q, e);
> >   
> > +	atomic_set(&set->iter_usage_counter, 1);
> Can it happen that the above atomic_set() call happens while a
> blk_mq_tagset_busy_iter() call is in progress?

No, as at this point it should be ensured that iter_usage_counter holds 
0 from atomic_cmpxchg(), so there should be no active processes in 
blk_mq_tagset_busy_iter() sensitive region. Calls to 
blk_mq_tagset_busy_iter() are blocked when iter_usage_counter holds 0.

> Should that atomic_set()
> call perhaps be changed into an atomic_inc() call?

They have the same affect in practice, but we use atomic_set() in 
blk_mq_alloc_tag_set(), so at least consistent.

Thanks,
John