Bug 55135

Oh, this is on pulseaudio 2.0 (wouldn't have just filed this but I thought I was running latest release...)

Had some debugging help... It looks like gstreamer does something weird and pulseaudio decides to assert because of that. If I use paplay everything works, but gstreamer using apps will assert PA with module-match.

Here's a pa_sink_input_new_data* in pa_sink_input_new() -- notice how the sample_spec is just empty, e.g. sample_spec.channels = 0:

Breakpoint 1, pa_sink_input_new (_i=_i@entry=0x7fffb2fd4508, core=0x21ea130, data=data@entry=0x7fffb2fd4680)
    at pulsecore/sink-input.c:239
239	pulsecore/sink-input.c: No such file or directory.
(gdb) print *data
$1 = {flags = PA_SINK_INPUT_START_CORKED, proplist = 0x2208780, 
  driver = 0x7f681ee55920 "pulsecore/protocol-native.c", module = 0x2272550, client = 0x2204350, sink = 0x0, 
  origin_sink = 0x0, resample_method = PA_RESAMPLER_INVALID, sync_base = 0x0, sample_spec = {format = PA_SAMPLE_U8, 
    rate = 0, channels = 0 '\000'}, channel_map = {channels = 0 '\000', map = {
      PA_CHANNEL_POSITION_MONO <repeats 32 times>}}, format = 0x0, req_formats = 0x2200e80, nego_formats = 0x0, 
  volume = {channels = 0 '\000', values = {0 <repeats 32 times>}}, volume_factor = {channels = 0 '\000', values = {
      0 <repeats 32 times>}}, volume_factor_sink = {channels = 0 '\000', values = {0 <repeats 32 times>}}, 
  muted = false, sample_spec_is_set = false, channel_map_is_set = false, volume_is_set = false, 
  volume_factor_is_set = false, volume_factor_sink_is_set = false, muted_is_set = false, volume_is_absolute = false, 
  volume_writable = true, save_sink = false, save_volume = false, save_muted = false}


I don't think PA should assert when this happens.

I couldn't reproduce this, but the information posted so far is enough to find the error. module-match uses the SINK_INPUT_NEW hook, and then relies on the sample spec being properly initialized. This is clearly wrong: if the sink input uses the FIX_CHANNELS flag, the number of channels will be unknown until the sink input has been routed to some sink.

In your case the FIX_CHANNELS flag is not being used. I'm not sure if there's some other bug involved in your case (I wouldn't be surprised), but fixing module-match should make things work for you. If the fix doesn't turn out to be hard to do, I'll post a patch soon.

Created attachment 67767 [details] [review]
Fix candidate

Does this patch fix the problem for you?

(In reply to comment #4)
> Created attachment 67767 [details] [review] [review]
> Fix candidate
> 
> Does this patch fix the problem for you?

It does. No asserts and seems to work as expected.

Thanks.

Although I did not test at the same time with stream-restore... Is this bound to break the functionality refered in this comment:
     /* hook EARLY - 1, to match before stream-restore */

(I'm not familiar with the internals so maybe that makes no sense)

This patch actually makes the comment correct :) I think earlier module-stream-restore used the NEW hook too, and the comment was written at that time. Then module-stream-restore was changed to use the FIXATE hook, possibly to fix a bug similar to this, so the comment became outdated. Now the comment is relevant again, because the two modules use the same hook.

Fix applied in master: http://cgit.freedesktop.org/pulseaudio/pulseaudio/commit/?id=0a8634f03e64a84f594bc8533235998e969eda66