[alsa-devel] Fixing snd_pcm_indirect_playback_transfer to support rewinds?

[Takashi Iwai]

On Tue, 02 Jan 2018 22:46:16 +0100,
Floris Bos wrote:
> 
> On 01/02/2018 02:49 PM, Takashi Iwai wrote:
> >
> >> It seems it is simply denying rewinds instead of making them work?
> >> Isn't there any way to make them work, instead of disabling
> >> functionality userspace seems to be expecting to have?
> > The rewind can't work with the indirect PCM mode due to its nature.
> > In the indirect PCM operation, the data has been already copied, thus
> > you can't go back beyond the point.
> 
> Ah, didn't know it consumed all data right away inside the function.
> The construct with sw_ready/hw_ready/hw_queue_size variables gave me
> the impression it could also only process a part and leave the rest
> for a next call.

It's not always copying the whole, but in principle, the indirect PCM
tries to copy the data from sw buffer to hw buffer as much as
possible.  That, of course, limits the availability of rewind in most
cases.

> BTW another thing that I noticed, is that it currently is only
> possible for a sound driver to report either success or failure from
> the ack callback.
> While the userspace documentation seems to suggest that partial
> success -in which a lower number of frames than requested is rewinded-
> is also a possibility.

Yes, but it's far more difficult to code in the current design.  The
whole rewind procedure is done only in appl_ptr and hw_ptr
manipulation, and the only thing we can perform in addition is the
call of ack callback, and it can give only success or failure.

Basically, if the error from rewind doesn't work, it's clearly an
application bug that heavily relies on the feature that is never
guaranteed to work.  So the best would be to fix the application.

Meanwhile, PulseAudio is almost a single application that is actually
using the rewind.  Breaking it doesn't sound like a good idea.

If the workaround with SNDRV_PCM_INFO_BATCH doesn't seem working,
maybe the patch like below would be a compromise.

Let me know if it works.


thanks,

Takashi

-- 8< --
From: Takashi Iwai <tiwai at suse.de>
Subject: [PATCH] ALSA: pcm: Workaround for weird PulseAudio behavior on rewind
 error

The commit 9027c4639ef1 ("ALSA: pcm: Call ack() whenever appl_ptr is
updated") introduced the possible error code returned from the PCM
rewind ioctl.  Basically the change was for handling the indirect PCM
more correctly, but ironically, it caused rather a side-effect:
PulseAudio gets pissed off when receiving an error from rewind, throws
everything away and stops processing further, resulting in the
silence.

It's clearly a failure in the application side, so the best would be
to fix that bug in PA.  OTOH, PA is mostly the only user of the rewind
feature, so it's not good to slap the sole customer.

This patch tries to mitigate the situation: instead of returning an
error, now the rewind ioctl returns zero when the driver can't rewind.
It indicates that no rewind was performed, so the behavior is
consistent, at least.

Fixes: 9027c4639ef1 ("ALSA: pcm: Call ack() whenever appl_ptr is updated")
Cc: <stable at vger.kernel.org>
Signed-off-by: Takashi Iwai <tiwai at suse.de>
---
 sound/core/pcm_native.c | 9 +++++++--
 1 file changed, 7 insertions(+), 2 deletions(-)

diff --git a/sound/core/pcm_native.c b/sound/core/pcm_native.c
index a4d92e46c459..f08772568c17 100644
--- a/sound/core/pcm_native.c
+++ b/sound/core/pcm_native.c
@@ -2580,7 +2580,7 @@ static snd_pcm_sframes_t forward_appl_ptr(struct snd_pcm_substream *substream,
 	return ret < 0 ? ret : frames;
 }
 
-/* decrease the appl_ptr; returns the processed frames or a negative error */
+/* decrease the appl_ptr; returns the processed frames or zero for error */
 static snd_pcm_sframes_t rewind_appl_ptr(struct snd_pcm_substream *substream,
 					 snd_pcm_uframes_t frames,
 					 snd_pcm_sframes_t avail)
@@ -2597,7 +2597,12 @@ static snd_pcm_sframes_t rewind_appl_ptr(struct snd_pcm_substream *substream,
 	if (appl_ptr < 0)
 		appl_ptr += runtime->boundary;
 	ret = pcm_lib_apply_appl_ptr(substream, appl_ptr);
-	return ret < 0 ? ret : frames;
+	/* NOTE: we return zero for errors because PulseAudio gets depressed
+	 * upon receiving an error from rewind ioctl and stops processing
+	 * any longer.  Returning zero means that no rewind is done, so
+	 * it's not absolutely wrong to answer like that.
+	 */
+	return ret < 0 ? 0 : frames;
 }
 
 static snd_pcm_sframes_t snd_pcm_playback_rewind(struct snd_pcm_substream *substream,
-- 
2.15.1