The current default value for prebuf is very high, almost the full virtual ALSA buffer. This breaks some application especially where low latency is involved.

This patch makes pcm_pulse implement the sw_params callback and get the prebuf value from the ALSA software parameters. Thus the trigger latency is much more like what an ALSA application should expect from an ALSA PCM device. --- pulse/pcm_pulse.c | 25 +++++++++++++++++++++++++ 1 file changed, 25 insertions(+)

diff --git a/pulse/pcm_pulse.c b/pulse/pcm_pulse.c index 0165120..08b6cea 100644 --- a/pulse/pcm_pulse.c +++ b/pulse/pcm_pulse.c @@ -859,6 +859,30 @@ static int pulse_hw_params(snd_pcm_ioplug_t * io, return err; }

+static int pulse_sw_params(snd_pcm_ioplug_t *io, snd_pcm_sw_params_t *params) +{ + snd_pcm_pulse_t *pcm = io->private_data; + snd_pcm_uframes_t start_threshold; + + assert(pcm); + + if (!pcm->p || !pcm->p->mainloop) + return -EBADFD; + + pa_threaded_mainloop_lock(pcm->p->mainloop); + + snd_pcm_sw_params_get_start_threshold(params, &start_threshold); + + if (start_threshold < io->period_size) + start_threshold = io->period_size; + + pcm->buffer_attr.prebuf = start_threshold * pcm->frame_size; + + pa_threaded_mainloop_unlock(pcm->p->mainloop); + + return 0; +} + static int pulse_close(snd_pcm_ioplug_t * io) { snd_pcm_pulse_t *pcm = io->private_data; @@ -925,6 +949,7 @@ static const snd_pcm_ioplug_callback_t pulse_playback_callback = { .poll_revents = pulse_pcm_poll_revents, .prepare = pulse_prepare, .hw_params = pulse_hw_params, + .sw_params = pulse_sw_params, .close = pulse_close, .pause = pulse_pause };

Le jeudi 15 novembre 2012 10:00:03, David Henningsson a écrit :

On 11/14/2012 07:07 PM, Rémi Denis-Courmont wrote:

...

The current default value for prebuf is very high, almost the full virtual ALSA buffer. This breaks some application especially where low latency is involved.

This patch makes pcm_pulse implement the sw_params callback and get the prebuf value from the ALSA software parameters. Thus the trigger latency is much more like what an ALSA application should expect from an ALSA PCM device.

This seems reasonable I believe; see review comments below.

...

---

pulse/pcm_pulse.c | 25 +++++++++++++++++++++++++ 1 file changed, 25 insertions(+)

diff --git a/pulse/pcm_pulse.c b/pulse/pcm_pulse.c index 0165120..08b6cea 100644 --- a/pulse/pcm_pulse.c +++ b/pulse/pcm_pulse.c @@ -859,6 +859,30 @@ static int pulse_hw_params(snd_pcm_ioplug_t * io,

return err;

}

+static int pulse_sw_params(snd_pcm_ioplug_t *io, snd_pcm_sw_params_t *params) +{

• snd_pcm_pulse_t *pcm = io->private_data;
• snd_pcm_uframes_t start_threshold;
• assert(pcm);
• if (!pcm->p || !pcm->p->mainloop)

• return -EBADFD;
  

• pa_threaded_mainloop_lock(pcm->p->mainloop);
• snd_pcm_sw_params_get_start_threshold(params, &start_threshold);
• if (start_threshold < io->period_size)

• start_threshold = io->period_size;
  

Why do we need the above constraint?

In my experience, PulseAudio really does not like having less than a period worth of buffered samples. I guess the code assumes there is always at least that much, and may underrun if not.

But sure enough, we could remove the constraint and blame the ALSA application for using inadequate parameters if a problem occurs. What do you prefer?

Hi Remi, and thanks for your patch, this is indeed a thorny area where you change behaviour and fix one application and break another...that has at least been my experience with the alsa-pulse plugin...

On 11/14/2012 07:07 PM, Rémi Denis-Courmont wrote:

There are no samples to play at that point so this is useless.

I'm not following this reasoning. The start callback corresponds to the _trigger callback in kernel, which wants the playback to start immediately.

For playback streams, you should not call the start callback (or any alsa-lib function that ends up calling it) from your application unless there are already samples in the buffer, or you will get an immediate underrun.

For recording streams, starting with an empty buffer is expected and not useless at all.

The non-zero buffer_attr.prebuf ensures that the PulseAudio daemon will start streaming automatically (like ALSA sw_params start_threshold).

The start callback is for the cases where the application wants to explicitly start the stream before prebuf has been reached.

---

pulse/pcm_pulse.c | 12 +++--------- 1 file changed, 3 insertions(+), 9 deletions(-)

diff --git a/pulse/pcm_pulse.c b/pulse/pcm_pulse.c index 24fd4da..0165120 100644 --- a/pulse/pcm_pulse.c +++ b/pulse/pcm_pulse.c @@ -204,8 +204,8 @@ static void stream_success_cb(pa_stream * p, int success, void *userdata) static int pulse_start(snd_pcm_ioplug_t * io) { snd_pcm_pulse_t *pcm = io->private_data;

• pa_operation *o, *u;
• int err = 0, err_o = 0, err_u = 0;

• pa_operation *o;

• int err = 0, err_o = 0;

  assert(pcm);

@@ -224,18 +224,12 @@ static int pulse_start(snd_pcm_ioplug_t * io) goto finish; }

• u = pa_stream_trigger(pcm->stream, stream_success_cb, pcm);

• pcm->underrun = 0; err_o = pulse_wait_operation(pcm->p, o);

• if (u)

• err_u = pulse_wait_operation(pcm->p, u);
  
  

  pa_operation_unref(o);

• if (u)

• pa_operation_unref(u);
  
  

• if (err_o < 0 || err_u < 0) {

• if (err_o < 0) { err = -EIO; goto finish; }