Re: [RFC PATCH] blk-mq: fix potential I/O hang caused by batch wakeup

[YangYang]

On 2024/5/21 2:11, Bart Van Assche wrote:
> On 5/19/24 20:38, Yang Yang wrote:
> > The depth is 62, and the wake_batch is 8. In the following situation,
> > the task would hang forever.
> >
> >    t1:                 t2:                          t3:
> >    blk_mq_get_tag      .                            .
> >    io_schedule         .                            .
> >                        elevator_switch              .
> >                        blk_mq_freeze_queue          .
> >                        blk_freeze_queue_start       .
> >                        blk_mq_freeze_queue_wait     .
> >                                                     blk_mq_submit_bio
> >                                                     __bio_queue_enter
> >
> > Fix this issue by waking up all the waiters sleeping on tags after
> > freezing the queue.
>
> Shouldn't blk_mq_alloc_request() be mentioned in t1 since that is the function
> that calls blk_queue_enter()?

 t1:                      t2:                          t3:
 blk_mq_submit_bio        .                            .
 __blk_mq_alloc_requests  .                            .
 blk_mq_get_tag           .                            .
 io_schedule              .                            .
                          elevator_switch              .
                          blk_mq_freeze_queue          .
                          blk_freeze_queue_start       .
                          q->mq_freeze_depth=1         .
                          blk_mq_freeze_queue_wait     .
                                                       blk_mq_submit_bio
                                                       __bio_queue_enter
                                                       wait_event(!q->mq_freeze_depth)

> > diff --git a/block/blk-core.c b/block/blk-core.c
> > index a16b5abdbbf5..e1eacfad6e5b 100644
> > --- a/block/blk-core.c
> > +++ b/block/blk-core.c
> > @@ -298,8 +298,6 @@ void blk_queue_start_drain(struct request_queue *q)
> >        * prevent I/O from crossing blk_queue_enter().
> >        */
> >       blk_freeze_queue_start(q);
> > -    if (queue_is_mq(q))
> > -        blk_mq_wake_waiters(q);
> >       /* Make blk_queue_enter() reexamine the DYING flag. */
> >       wake_up_all(&q->mq_freeze_wq);
> >   }
>
> Why has blk_queue_start_drain() been modified? I don't see any reference
> in the patch description to blk_queue_start_drain(). Am I perhaps missing
> something?

blk_mq_wake_waiters() has already been called in blk_freeze_queue_start(),
so I thought it can be removed from blk_queue_start_drain().

> > diff --git a/block/blk-mq.c b/block/blk-mq.c
> > index 4ecb9db62337..9eb3139e713a 100644
> > --- a/block/blk-mq.c
> > +++ b/block/blk-mq.c
> > @@ -125,8 +125,10 @@ void blk_freeze_queue_start(struct request_queue *q)
> >       if (++q->mq_freeze_depth == 1) {
> >           percpu_ref_kill(&q->q_usage_counter);
> >           mutex_unlock(&q->mq_freeze_lock);
> > -        if (queue_is_mq(q))
> > +        if (queue_is_mq(q)) {
> > +            blk_mq_wake_waiters(q);
> >               blk_mq_run_hw_queues(q, false);
> > +        }
> >       } else {
> >           mutex_unlock(&q->mq_freeze_lock);
> >       }
>
> Why would the above change be necessary? If the blk_queue_enter() call
> by blk_mq_alloc_request() succeeds and blk_mq_get_tag() calls
> io_schedule(), io_schedule() will be woken up indirectly by the
> blk_mq_run_hw_queues() call because that call will free one of the tags
> that the io_schedule() call is waiting for.

This patch is a workaround solution. I think the hang is caused by
a lost wakeup, so after blk_mq_run_hw_queues(), t1 is still waiting
for the tag.

    bt = 0xFFFFFF802F9C6790 -> (
    sb = (
      depth = 62,
      shift = 6,
      map_nr = 1,
      round_robin = FALSE,
      map = 0xFFFFFF803BF97000,
      alloc_hint = 0x00000049119B3F4C),
    wake_batch = 6,
    wake_index = (counter = 0),
    ws = 0xFFFFFF803BEBCA00,
    ws_active = (counter = 1),
    min_shallow_depth = 48,
    completion_cnt = (counter = 1),
    wakeup_cnt = (counter = 0))

Upon analyzing the coredump, it was noticed that sbq->completion_cnt=1,
and I can't figure out why.
blk_mq_put_tag->sbitmap_queue_clear->sbitmap_queue_wake_up(sbq, 1) should
be called multiple times, considering that sbq->ws_active=1,
sbq->completion_cnt should be greater than 1.
Looking forward to some advice from block layer experts.

Thanks.

> Thanks,
>
> Bart.