Re: sound: deadlock involving snd_hrtimer_callback

From: Dmitry Vyukov
Date: Mon Apr 25 2016 - 04:04:01 EST


On Sun, Apr 24, 2016 at 11:31 PM, Takashi Iwai <tiwai@xxxxxxx> wrote:
> On Sun, 24 Apr 2016 19:09:48 +0200,
> Dmitry Vyukov wrote:
>>
>> On Sun, Apr 24, 2016 at 6:16 PM, Takashi Iwai <tiwai@xxxxxxx> wrote:
>> > On Sun, 24 Apr 2016 17:16:32 +0200,
>> > Dmitry Vyukov wrote:
>> >>
>> >> On Sat, Apr 23, 2016 at 11:02 PM, Takashi Iwai <tiwai@xxxxxxx> wrote:
>> >> > On Sat, 23 Apr 2016 15:40:21 +0200,
>> >> > Dmitry Vyukov wrote:
>> >> >>
>> >> >> Hi Takashi,
>> >> >>
>> >> >> I've incorporated your hrtimer fixes (but also updated to
>> >> >> ddce192106e4f984123884f8e878f66ace94b573) and now I am seeing lots of
>> >> >> the following deadlock messages:
>> >> >>
>> >> >>
>> >> >> [ INFO: possible circular locking dependency detected ]
>> >> >> 4.6.0-rc4+ #351 Not tainted
>> >> >> -------------------------------------------------------
>> >> >> swapper/0/0 is trying to acquire lock:
>> >> >> (&(&timer->lock)->rlock){-.-...}, at: [<ffffffff8537a749>]
>> >> >> snd_timer_interrupt+0xa9/0xd30 sound/core/timer.c:701
>> >> >>
>> >> >> but task is already holding lock:
>> >> >> (&(&stime->lock)->rlock){-.....}, at: [<ffffffff85383d3f>]
>> >> >> snd_hrtimer_callback+0x4f/0x2b0 sound/core/hrtimer.c:54
>> >> >>
>> >> >> which lock already depends on the new lock.
>> >> >
>> >> > Oh crap, my second patch is buggy, it leads to ABBA lock, indeed.
>> >> > The first patch is still OK, as it just adds a new behavior mode.
>> >> >
>> >> > Could you replace the second patch with the below one?
>> >>
>> >>
>> >> I've replaced the second path with this one. The deadlocks has gone,
>> >> but I've hit these two hangs that look related:
>> >>
>> >> https://gist.githubusercontent.com/dvyukov/805718ea249c49d17ae759d1b0160684/raw/20891f7e87fe9af3967565559d465d296469244b/gistfile1.txt
>> >> https://gist.githubusercontent.com/dvyukov/7f397ea4aeb9e35596e0c8053cf35a11/raw/3fc22f24f7bab5941e47bab604f96487b5f1944d/gistfile1.txt
>> >
>> > Hmm, so it wasn't a good idea to call hrtimer_cancel() in the
>> > spinlock, in anyway. Scratch the previous one.
>> >
>> > OK, below is the yet revised two patches. One is the simplified
>> > version of the patch, and another is to call hrtimer_cancel() in a new
>> > timer op without spinlock. Apply these after the first patch
>> > "ALSA: timer: Allow backend disabling start/stop from handler".
>>
>> Done. I will let you know if I see any failures.
>
> After reconsideration, I rewrote the whole patchset again.
> Could you scratch all the previous three patches and replace with the
> single patch below? Sorry for inconvenience!


I did not yet reapply the patch, but I hit this with over night:
https://gist.githubusercontent.com/dvyukov/85531fe7923ebc9be4376e009d6c960a/raw/9ea4c9dda31aea9bf20538585089083d66687f74/gistfile1.txt