[alsa-devel] [RFC PATCH] ALSA: core: do not update tstamp when hw ptr has not moved

[Henrik Eriksson]

On Thu, Nov 16, 2017 at 19:49:36 +0100, Takashi Iwai wrote:
> On Thu, 16 Nov 2017 19:11:37 +0100,
> Pierre-Louis Bossart wrote:
> > 
> > On 11/16/17 8:48 AM, Takashi Iwai wrote:
> > > On Thu, 16 Nov 2017 13:43:57 +0100,
> > > Henrik Eriksson wrote:
> > >>
> > >> Hello
> > >>
> > >> This reverts a change that was part of a larger commit.  That change
> > >> broke timestamps in some application code for us as outlined in the log,
> > >> and reverting it obviously fixes this.  Since the original change did
> > >> not explain why the behaviour was modified I can't speak for the effects
> > >> on that use case.
> > >>
> > >> Regards,
> > >> /henrik
> > >>> 8------------------------------------------------------8<
> > >> Commit 3179f62001880e588e229db3006a59ad87b7792a ("ALSA: core: add
> > >> .get_time_info") made an undocumented change to the behaviour of the
> > >> PCM runtime tstamp.  Prior to this change the tstamp was not updated
> > >> by snd_pcm_update_hw_ptr0() unless the hw ptr had moved, after this
> > >> change the tstamp is always updated.
> > >>
> > >> For an application, using alsa-lib, doing snd_pcm_readi() followed by
> > >> snd_pcm_status() to estimate the age of the read samples by subtracting
> > >> status->avail * [sample rate] from status->tstamp this change degraded
> > >> the accuracy of that estimate on devices where the pcm hw does not
> > >> provide a granular hw pointer.  On a device using
> > >> soc-generic-dmaengine-pcm.c and a dma-engine with residue_granularity
> > >> DMA_RESIDUE_GRANULARITY_DESCRIPTOR the accuracy of the estimate depended
> > >> on the latency between the PCM hw completing a period and when the
> > >> driver called snd_pcm_period_elapsed() to notify ALSA core, typically
> > >> determined by interrupt handling latency.  After this change the
> > >> accuracy of the estimate is determined by the latency between the PCM hw
> > >> completing a period and the application calling snd_pcm_status(),
> > >> determined by the scheduling of the application process.
> > >>
> > >> Revert the part of commit 3179f62001880e588e229db3006a59ad87b7792a
> > >> ("ALSA: core: add .get_time_info") that changed be behaviour of
> > >> snd_pcm_update_hw_ptr0() when the PCM hw pointer has not moved.
> > >>
> > >> Fixes: 3179f6200188 ("ALSA: core: add .get_time_info")
> > >> Signed-off-by: Henrik Eriksson <henrik.eriksson at axis.com>
> > >
> > > So this is a regression and we need to address it.
> > >
> > > I *guess* that the reason to update tstamp event at the same hwptr is
> > > that now we can get a more fine-grained tstamp.  OTOH, we have to
> > > avoid the regression by that.
> > >
> > > What about the (untested) patch like below?  It skips the tstamp
> > > update as default but follows the new standard when a better tstamp
> > > mode is deployed.
> > 
> > update_audio_tstamp() only updates the audio timestamp when the type
> > is DEFAULT already, but indeed update runtime->status->tstamp is
> > updated unconditionally.
> > 
> > I have a bit of heartburn with the suggested solutions because the
> > delay can change even with the same hw_ptr. Not updating the
> > timestamps would be going back to the days where timing updates was
> > dependent on DMA granularity.
> > 
> > Maybe we could make the tstamp update dependent on a change in
> > audio_tstamp, e.g.
> > 
> > 
> > diff --git a/sound/core/pcm_lib.c b/sound/core/pcm_lib.c
> > index a93a4235a332..b606f3ea7a17 100644
> > --- a/sound/core/pcm_lib.c
> > +++ b/sound/core/pcm_lib.c
> > @@ -248,8 +248,10 @@ static void update_audio_tstamp(struct
> > snd_pcm_substream *substream,
> >                                 runtime->rate);
> >                 *audio_tstamp = ns_to_timespec(audio_nsecs);
> >         }
> > -       runtime->status->audio_tstamp = *audio_tstamp;
> > -       runtime->status->tstamp = *curr_tstamp;
> > +       if (runtime->status->audio_tstamp != *audio_tstamp) {
> > +               runtime->status->audio_tstamp = *audio_tstamp;
> > +               runtime->status->tstamp = *curr_tstamp;
> > +       }
> 
> OK for me as long as it works.
> Henrik?

It seems reasonable at first glance.  I'll try it out.

BTW, I ran your patch variant overnight and, as expected, it has the
same behaviour as my original take since we do not have get_time_info
in the pcm driver.  In real terms that means that the result of
status->tstamp - status->avail * [rate] on consequtive buffers stayed
within 64 +- 3 ms, on a 64 ms period length.

Thanks.

/henrik