[alsa-devel] [PATCH 0/3] Make pcm_ioplug check lock status before locking (fixes pcm_jack lockups)

[Ben Russell]

On Tue, 24 Sep 2019 12:18:24 +0200
Takashi Iwai <tiwai at suse.de> wrote:

> On Tue, 24 Sep 2019 09:45:25 +0200,
> Ben Russell wrote:
> > 
> > On Sun, 22 Sep 2019 22:25:48 +0200
> > Takashi Iwai <tiwai at suse.de> wrote:
> > 
> > > On Sun, 22 Sep 2019 05:28:50 +0200,
> > > Ben Russell wrote:
> > > > 
> > > > This is my first time contributing a patch to a mailing list. If I've made a mess, please let me know so I can learn how to avoid it in future.
> > > > 
> > > > The purpose of this patchset is to fix a specific common lockup in the pcm_jack plugin (from alsa-plugins). I'm not sure if this is the right approach to take, but at the very least it should make the pcm_ioplug code a bit more resilient.
> > > > 
> > > > The problem is this: When using the pcm_jack plugin, if a program attempts to play audio using the SND_PCM_FORMAT_FLOAT format, said program locks up.
> > > > 
> > > > This should be enough to reproduce the bug:
> > > > 
> > > >     pcm.rawjack {
> > > >         type jack
> > > >         playback_ports {
> > > >             0 system:playback_1
> > > >             1 system:playback_2
> > > >         }
> > > >         capture_ports {
> > > >             0 system:capture_1
> > > >             1 system:capture_2
> > > >         }
> > > >     }
> > > >     
> > > >     pcm.!default {
> > > >         type plug
> > > >         slave.pcm "rawjack"
> > > >     }
> > > > 
> > > > What's happening is that several snd_pcm_ioplug_* functions assume that the pcm mutex is locked already. It then proceeds to unlock the mutex, call a function, and then relock the mutex. When the mutex isn't locked already, the initial unlock results in a silently ignored pthread error, and the lock results in the program eventually deadlocking as it doesn't expect the mutex to be held at that point.
> > > > 
> > > > Patch 2 modifies pcm_ioplug to check if the mutex is held before doing the unlock-act-lock sequence, and if the mutex is not held then it skips the unlock and lock stages. This depends on Patch 1, which adds a snd_pcm_is_locked function to give the state of the mutex.
> > > > 
> > > > Patch 3 is completely optional. It adds assertions which make sure that all uses of snd_pcm_lock/snd_pcm_unlock are correct. On one hand this will likely result in crashes in some of the less refined parts of the code. On the other hand, when that happens, you'll know which parts need a bit more love. I know it was useful for finding this issue in the first place.
> > > > 
> > > > These patches fix the problems I am having, but if you have a more suitable approach to fixing this problem then please let me know.
> > > 
> > > Thanks for the analysis and patches.  It's indeed a serious problem we
> > > need to address.
> > > 
> > > The current semantics of locking in alsa-lib code assumes that the
> > > lock/unlock never fails.  So the "right-and-quick" fix would be just
> > > to take the patch 3 to assert() upon pthread_mutex_lock/unlock
> > > errors.  Then we need to paper over the actual invalid locks.
> > > 
> > > I do wonder, though, exactly which code path triggers the pthread
> > > mutex error?  You should be able to catch it via gdb after the patch.
> > > 
> > > 
> > > thanks,
> > > 
> > > Takashi
> > 
> > Sure thing. With patch #3 against commit 1d7a131f, this trips up on an assert:
> > 
> > $ gdb --args mpv --ao=alsa --audio-format=float RESPECOGNIZE.mp3
> > 
> > The relevant backtrace:
> > 
> > #4  0x00007ffff4b06cf3 in snd_pcm_unlock (pcm=0x5555558625c0) at pcm_local.h:1187
> > #5  0x00007ffff4b075b3 in snd_pcm_unlock (pcm=0x5555558625c0) at pcm_local.h:1187
> > 	if (pcm->lock_enabled && pcm->need_lock) {
> > 		unlock_err = pthread_mutex_unlock(&pcm->lock);
> > 		assert(unlock_err == 0); <-- this line
> > 	}
> > #6  snd_pcm_ioplug_prepare (pcm=0x5555558625c0) at pcm_ioplug.c:167
> > 	snd_pcm_ioplug_reset(pcm);
> > 	if (io->data->callback->prepare) {
> > 		snd_pcm_unlock(pcm); /* to avoid deadlock */ <-- this line
> > 		err = io->data->callback->prepare(io->data);
> > 		snd_pcm_lock(pcm);
> > 	}
> > #7  0x00007ffff4abcba0 in snd_pcm_prepare (pcm=0x5555558627c0) at pcm.c:1214
> > 	snd_pcm_lock(pcm);
> > 	if (pcm->fast_ops->prepare)
> > 		err = pcm->fast_ops->prepare(pcm->fast_op_arg); <-- this line
> > 	else
> > 		err = -ENOSYS;
> > #8  0x00005555555a2efa in ?? ()
> > 
> > Using --audio-format=s16 does not deadlock, as it goes via pcm_plugin before it can go via pcm_ioplug (it's most likely for sample format conversion):
> > 
> > #0  0x00007ffff4b07522 in snd_pcm_unlock (pcm=<optimized out>) at pcm_ioplug.c:166
> > #1  snd_pcm_ioplug_prepare (pcm=0x555555862640) at pcm_ioplug.c:167
> > #2  0x00007ffff4abcba0 in snd_pcm_prepare (pcm=0x555555862640) at pcm.c:1214
> > #3  0x00007ffff4ad2ed7 in snd_pcm_plugin_prepare (pcm=0x555555862e50) at pcm_plugin.c:159
> > 	snd_pcm_plugin_t *plugin = pcm->private_data;
> > 	int err;
> > 	err = snd_pcm_prepare(plugin->gen.slave); <-- this line
> > 	if (err < 0)
> > 		return err;
> > #4  0x00007ffff4abcba0 in snd_pcm_prepare (pcm=0x555555862840) at pcm.c:1214
> > #5  0x00005555555a2efa in ?? ()
> 
> Thanks, after reading the traces carefully, finally I figured out what
> went wrong.  In some cases, pcm->fast_op_arg isn't pcm itself but has
> a different value.  Then we call snd_pcm_lock() with pcm, but call
> fast_op with fast_op_arg as the PCM object.  This leads to the
> unexpected behavior for the plugin that tries to unlock / relock from
> the given PCM object.
> 
> So, we'd need to convert snd_pcm_lock()/unlock() with fast_op_arg or
> op_arg in most places.
> 
> A test fix patch is below.  Could you test it quickly?
> I'm going to traveling from tomorrow, so I'd like to have confirmation
> before merging into git repo.
> 
> 
> thanks,
> 
> Takashi
> 
> -- 8< --

The patch works fine. mpv works for float and s16, Wine works, Mumble works for capture and playback.

Thanks for tracking this down so quickly.

Regards,
Ben R