Re: [PATCH -next RFC v2 8/8] sbitmap: wake up the number of threads based on required tags

[yukuai (C)]

在 2022/04/09 12:16, Bart Van Assche 写道:
> On 4/8/22 19:17, yukuai (C) wrote:
> > I think the reason to wake up 'wake_batch' waiters is to make sure
> > wakers will use up 'wake_batch' tags that is just freed, because each
> > wakers should aquire at least one tag. Thus I think if we can make sure
> > wakers will use up 'wake_batch' tags, it's ok to wake up less waiters.
>
> Hmm ... I think it's up to you to (a) explain this behavior change in 
> detail in the commit message and (b) to prove that this behavior change 
> won't cause trouble (I guess this change will cause trouble).

Hi, Bart

Sorry that the commit message doesn't explain clearly.

There are only two situations that wakers will be less than 'wake_batch'
after this patch:

(a) some wakers will acquire multipul tags, as I mentioned above, this
is ok because wakers will use up 'wake_batch' tags.

(b) the total number of waiters is less than 'wake_batch', this is
problematic if tag preemption is disabled, because io concurrency will
be declined.(patch 5 should fix the problem)

For the race that new threads are waited after get_wake_nr() and before
wake_up_nr() in situation (b), I can't figure out how this can be
problematic, however, this can be optimized by triggering additional
wake up:

@@ -623,15 +623,17 @@ static unsigned int get_wake_nr(struct 
sbq_wait_state *ws, unsigned int nr_tags)
        spin_lock_irq(&ws->wait.lock);
        list_for_each_entry(entry, &ws->wait.head, entry) {
                wait = container_of(entry, struct sbq_wait, wait);
-               if (nr_tags <= wait->nr_tags)
+               if (nr_tags <= wait->nr_tags) {
+                       nr_tags = 0;
                        break;
+               }

                nr++;
                nr_tags -= wait->nr_tags;
        }
        spin_unlock_irq(&ws->wait.lock);

-       return nr;
+       return nr + nr_tags;
 }

What do you think?

Thanks,
Kuai

> Thanks,
>
> Bart.
> .