[RFC PATCH] ALSA: compress: Fix gapless playback state machine

[Vinod Koul]

On 11-06-20, 12:40, Jaroslav Kysela wrote:
> Dne 11. 06. 20 v 11:44 Vinod Koul napsal(a):
> > On 11-06-20, 11:09, Jaroslav Kysela wrote:
> > > Dne 11. 06. 20 v 10:46 Charles Keepax napsal(a):
> > > > On Wed, Jun 10, 2020 at 04:28:20PM +0530, Vinod Koul wrote:
> > > > > On 10-06-20, 12:40, Jaroslav Kysela wrote:
> > > > > > Dne 10. 06. 20 v 12:07 Srinivas Kandagatla napsal(a):
> > > > > > > For gapless playback call to snd_compr_drain_notify() after
> > > > > > > partial drain should put the state to SNDRV_PCM_STATE_RUNNING
> > > > > > > rather than SNDRV_PCM_STATE_SETUP as the driver is ready to
> > > > > > > process the buffers for new track.
> > > > > > > 
> > > > > > > With existing code, if we are playing 3 tracks in gapless, after
> > > > > > > partial drain finished on previous track 1 the state is set to
> > > > > > > SNDRV_PCM_STATE_SETUP which is then moved to SNDRV_PCM_STATE_PREPARED
> > > > > > > after data write. With this state calls to snd_compr_next_track() and
> > > > > > > few other calls will fail as they expect the state to be in
> > > > > > > SNDRV_PCM_STATE_RUNNING.
> > > > > > > 
> > > > > > > Here is the sequence of events and state transitions:
> > > > > > > 
> > > > > > > 1. set_params (Track 1), state =  SNDRV_PCM_STATE_SETUP
> > > > > > > 2. set_metadata (Track 1), no state change, state = SNDRV_PCM_STATE_SETUP
> > > > > > > 3. fill and trigger start (Track 1), state = SNDRV_PCM_STATE_RUNNING
> > > > > > > 4. set_next_track (Track 2), state = SNDRV_PCM_STATE_RUNNING
> > > > > > > 5. partial_drain (Track 1), state = SNDRV_PCM_STATE_SETUP
> > > > > > > 6  snd_compr_drain_notify (Track 1), state = SNDRV_PCM_STATE_SETUP
> > > > > > > 7. fill data (Track 2), state = SNDRV_PCM_STATE_PREPARED
> > > > > > > 8. set_metadata (Track 3), no state change, state = SNDRV_PCM_STATE_PREPARED
> > > > > > > 9. set_next_track (Track 3), !! FAILURE as state != SNDRV_PCM_STATE_RUNNING
> > > > > > 
> > > > > > 
> > > > > > The snd_compr_drain_notify() is called only from snd_compr_stop(). Something
> > > > > > is missing in this sequence?
> > > > > 
> > > > > It is supposed to be invoked by driver when partial drain is complete..
> > > > > both intel and sprd driver are calling this. snd_compr_stop is stop
> > > > > while draining case so legit
> > > > > 
> > > > 
> > > > Not sure I follow this statement, could you elaborate a bit?
> > > > snd_compr_stop putting the state to RUNNING seems fundamentally
> > > > broken to me, the whole point of snd_compr_stop is to take the
> > > > state out of RUNNING.
> > > 
> > > Yes. I agree. It seems that the acknowledge for the partial drain should be
> > > handled differently.
> > 
> > Yeah sorry I overlooked that case and was thinking of it being invoked
> > from driver!
> > 
> > Yes this would make the snd_compr_stop() behave incorrectly.. so this
> > cant be done as proposed.
> > 
> > But we still need to set the draining stream state properly and I am
> > thinking below now:
> > 
> > diff --git a/sound/core/compress_offload.c b/sound/core/compress_offload.c
> > index 509290f2efa8..9aba851732d7 100644
> > --- a/sound/core/compress_offload.c
> > +++ b/sound/core/compress_offload.c
> > @@ -929,7 +929,9 @@ static int snd_compr_partial_drain(struct snd_compr_stream *stream)
> >          }
> >          stream->next_track = false;
> > -       return snd_compress_wait_for_drain(stream);
> > +       retval = snd_compress_wait_for_drain(stream);
> > +       stream->runtime->state = SNDRV_PCM_STATE_RUNNING;
> > +       return retval;
> >   }
> 
> I see a race possibility when the last track is too small and the driver
> signals the end-of-track twice. In this case the partial drain should not
> end with the running state. It would be probably better to separate partial
> / last track acknowledgements.

I completely agree that we should have separate acknowledgements here,
and going to rethink all state transitions for gapless here..

Thanks for the help
-- 
~Vinod