On Sun, 7 Oct 2007, Mike Gorse wrote:
I posted last week because I am getting lockups when portaudio calls snd_pcm_drain, and snd_pcm_rate_drain calls snd_pcm_wait on its slave.
Now that I've looked at this more, I'm seeing that it involves dmix, since a dmix plugin is behind the rate plugin.
The dmix code in alsa-lib is setting a xrun state because snd_pcm_mmap_playback_avail returns a large value.
snd_pcm_wait is then called on the slave of the rate pcm, which is a pcm with a type of SND_PCM_TYPE_ROUTE whose slave is the dmix, but the route pcm is returning a different value for snd_pcm_mmap_avail than the dmix pcm, so it is not triggering snd_pcm_wait to check for a xrun.
I think that better fix is to change the code in snd_pcm_rate_drain(). You can set avail_min to 1 there, so condition in snd_pcm_wait() will match.
Does this patch help?
diff -r 68d71239d651 src/pcm/pcm_rate.c --- a/src/pcm/pcm_rate.c Thu Sep 20 13:20:03 2007 +0200 +++ b/src/pcm/pcm_rate.c Fri Sep 21 10:14:41 2007 +0200 @@ -1076,10 +1076,10 @@ static int snd_pcm_rate_drain(snd_pcm_t snd_pcm_uframes_t size, ofs, saved_avail_min; snd_pcm_sw_params_t sw_params;
/* temporarily set avail_min to one period */
/* temporarily set avail_min to one */sw_params = rate->sw_params; saved_avail_min = sw_params.avail_min;
sw_params.avail_min = rate->gen.slave->period_size;
sw_params.avail_min = 1;snd_pcm_sw_params(rate->gen.slave, &sw_params);
size = rate->appl_ptr - rate->last_commit_ptr;
Jaroslav
Unfortunately, I just discovered that this patch (now in alsa-lib 1.0.15-rc3) does not always solve my problem. It does most of the time, but snd_pcm_avail on the route occasionally returns 0, which is less than 1, so once in a while I still get a lock-up. I can't change the 1 to a 0 in snd_pcm_drain since snd_pcm_sw_params aborts if sw_params->avail_min == 0.
My original approach of checking for SND_PCM_TYPE_ROUTE in snd_pcm_wait() would still work.
It looks as though there is a buffer to fill which is included when snd_pcm_mmap_avail looks at the dmix pcm but not when it looks at the route in between the rate and the dmix.
I think we're looking bug on wrong place. If dmix is in XRUN state, snd_pcm_poll_descriptors_revents() should return POLLERR, thus the check for XRUN should be done in snd_pcm_wait_nocheck() after revents checking.
Could you trace what occurs in your scenario in snd_pcm_wait_nocheck() ?
Thanks, Jaroslav
----- Jaroslav Kysela perex@suse.cz Linux Kernel Sound Maintainer ALSA Project, SUSE Labs