On Wed, 14 Nov 2012, Jens Axboe wrote:
> Date: Wed, 14 Nov 2012 08:21:41 -0700
> From: Jens Axboe <axboe@xxxxxxxxx>
> To: Lukáš Czerner <lczerner@xxxxxxxxxx>
> Cc: linux-kernel@xxxxxxxxxxxxxxx, linux-fsdevel@xxxxxxxxxxxxxxx,
> jmoyer@xxxxxxxxxx, akpm@xxxxxxxxxxxxxxxxxxxx
> Subject: Re: [PATCH v3] loop: Limit the number of requests in the bio list
>
> On 2012-11-14 02:02, Lukáš Czerner wrote:
> > On Tue, 13 Nov 2012, Jens Axboe wrote:
> >
> >> Date: Tue, 13 Nov 2012 09:42:58 -0700
> >> From: Jens Axboe <axboe@xxxxxxxxx>
> >> To: Lukas Czerner <lczerner@xxxxxxxxxx>
> >> Cc: linux-kernel@xxxxxxxxxxxxxxx, linux-fsdevel@xxxxxxxxxxxxxxx,
> >> jmoyer@xxxxxxxxxx, akpm@xxxxxxxxxxxxxxxxxxxx
> >> Subject: Re: [PATCH v3] loop: Limit the number of requests in the bio list
> >>
> >>> @@ -489,6 +491,12 @@ static void loop_make_request(struct request_queue *q, struct bio *old_bio)
> >>> goto out;
> >>> if (unlikely(rw == WRITE && (lo->lo_flags & LO_FLAGS_READ_ONLY)))
> >>> goto out;
> >>> + if (lo->lo_bio_count >= q->nr_congestion_on) {
> >>> + spin_unlock_irq(&lo->lo_lock);
> >>> + wait_event(lo->lo_req_wait, lo->lo_bio_count <
> >>> + q->nr_congestion_off);
> >>> + spin_lock_irq(&lo->lo_lock);
> >>> + }
> >>
> >> This makes me nervous. You are reading lo_bio_count outside the lock. If
> >> you race with the prepare_to_wait() and condition check in
> >> __wait_event(), then you will sleep forever.
> >
> > Hi Jens,
> >
> > I am sorry for being dense, but I do not see how this would be
> > possible. The only place we increase the lo_bio_count is after that
> > piece of code (possibly after the wait). Moreover every time we're
> > decreasing the lo_bio_count and it is smaller than nr_congestion_off
> > we will wake_up().
> >
> > That's how wait_event/wake_up is supposed to be used, right ?
>
> It is, yes. But you are checking the condition without the lock, so you
> could be operating on a stale value. The point is, you have to safely
> check the condition _after prepare_to_wait() to be completely safe. And
> you do not. Either lo_bio_count needs to be atomic, or you need to use a
> variant of wait_event() that holds the appropriate lock before
> prepare_to_wait() and condition check, then dropping it for the sleep.
>
> See wait_even_lock_irq() in drivers/md/md.h.
Ok I knew that much. So the only possibility to deadlock is when we
would process all the bios in the loop_thread() before the waiting
event would get to checking the condition after which we would read
the stale data where lo_bio_count is still < nr_congestion_off so we
get back to sleep, never to be woken up again. That sounds highly
unlikely. But fair enough, it make sense to make it absolutely bullet
proof.
I'll take a look at the wait_event_lock_irq.
Thanks!
-Lukas
