[pulseaudio-discuss] [PATCH] tunnel-{sink, source}-new: Fix assertion when used with module-loopback

[Georg Chini]

On 24.07.2017 18:58, Tanu Kaskinen wrote:
> On Sat, 2017-07-22 at 10:21 +0200, Georg Chini wrote:
>> On 20.07.2017 15:48, Tanu Kaskinen wrote:
>>> On Mon, 2017-07-17 at 19:53 +0200, Georg Chini wrote:
>>>> On 17.07.2017 19:32, Tanu Kaskinen wrote:
>>>>> On Sun, 2017-07-16 at 11:42 +0200, Georg Chini wrote:
>>>>>> Currently pulseaudio crashes with an assertion in pa_rtpoll_item_new_asyncmsgq_read()
>>>>>> if a loopback is applied to a tunnel-new sink/source because tunnel-{sink,source}-new
>>>>>> do not set thread_info.rtpoll. Bug was reported on IRC.
>>>>>>
>>>>>> This patch fixes the problem by initializing thread_info.rtpoll properly.
>>>>> Did you test this patch? The tunnel devices don't run the rtpoll that
>>>>> you create, so I would expect that the loopback won't work.
>>>>>
>>>>> This is a known bug that I started working on in the past:
>>>>> https://bugs.freedesktop.org/show_bug.cgi?id=73429
>>>>>
>>>>> I made some patches, issues got pointed out in review, and then I never
>>>>> finished v2 of the patches. I haven't given up on that, but it's been a
>>>>> year since I last worked on it... In case you're interested in the
>>>>> current state of the v2 patches, I pushed the code to
>>>>> git://people.freedesktop.org/~tanuk/pulseaudio branch "rtpoll-mainloop-
>>>>> v2".
>>>>>
>>>> I haven't tested it myself, but the bug reporter on IRC (tar-dingens)
>>>> tested them and said it worked. The patch looked like the easiest way to
>>>> make module-loopback happy.
>>> I don't understand how it can work. module-loopback uses the rtpoll to
>>> set up the internal asyncmsgq. If the tunnel device doesn't run the
>>> rtpoll, the messages shouldn't get processed. One of the messages is
>>> used to send memchunks from the source to the sink, so there should be
>>> no audio moving if the messages aren't processed.
>>>
>> I think I know why it works. The messages are processed in
>> sink_input_pop_cb(). module-loopback calls
>> pa_asyncmsgq_process_one() from there. So it is not necessary
>> to run the rtpoll at all. From what I know about the loopback code
>> it should be fine this way.
> Yep.
>
>> The question is, if we can hit other issues with the "fake" rtpoll.
>> If not, I think we could generally go with the patch (and add some
>> comments in the tunnel modules that the rtpoll is not used at all
>> and only there to make module-loopback happy).
> I think this is ok as a temporary band-aid. At least module-combine-
> sink and module-rtp-recv are affected too, and they might not work with
> this patch, but since their previous behaviour was to crash, it's hard
> to imagine how they could work any worse than that.
>
I'll send a new version with clear comments then. But before I do
so, I will test it myself and also check at least the behavior of
combine-sink.