[pulseaudio-discuss] Expand the simple API?

[Tanu Kaskinen]

On Wed, 2013-03-20 at 10:06 +0100, David Henningsson wrote:
> On 03/20/2013 09:49 AM, Tanu Kaskinen wrote:
> > On Wed, 2013-03-20 at 09:32 +0100, David Henningsson wrote:
> >> On 03/20/2013 08:57 AM, Tanu Kaskinen wrote:
> >>> On Tue, 2013-03-19 at 13:20 +0100, Thomas Martitz wrote:
> >>>> Am 19.03.2013 10:56, schrieb Tanu Kaskinen:
> >>>>> I'd be OK with adding pa_simple_writable/readable_size(). I believe the
> >>>>> multi-threading aspect of adding the write/read callbacks unavoidably
> >>>>> makes writing applications not-simple, so the simple API probably
> >>>>> shouldn't have those callbacks.
> >>>>>
> >>>>
> >>>> It's not unlikely that an application that uses blocking or polling
> >>>> methods uses an extra thread for that, which it wouldn't do if it could
> >>>> use the callback method. So in this case the callback would simplify the
> >>>> application, because the multi threading aspect is not in the
> >>>> application logic but in pulseaudio. I don't know how common this case
> >>>> is, but I would expect it's rather common.
> >>>
> >>> The applications would still need to be aware of the threading.
> >>> pa_simple_read/write() is expected to be done from the main thread, so
> >>> the application necessarily needs to implement some communication method
> >>> between the two threads.
> >>
> >>   From what I can see, there is no such restriction of the simple API. It
> >> does not have a concept of a "main thread", so as long as you don't call
> >> into the simple API from two threads at the same time you should be just
> >> fine, e g, it would be perfectly valid to do pa_simple_new in one
> >> thread, and later call pa_simple_write from another thread.
> >
> > pa_simple_write() calls pa_threaded_mainloop_wait(), and
> > pa_threaded_mainloop_wait() has this assertion:
> >
> > pa_assert(!m->thread || !pa_thread_is_running(m->thread) || !in_worker(m));
> >
> > The in_worker() check will fail, if pa_simple_write() was called from
> > the callback context.
> 
> Sure. You said "pa_simple_read/write() is expected to be done from the 
> main thread", but you meant more like "pa_simple_read/write() is 
> expected to be done from any other thread than the callback context"?
> 
> A pa_simple callback API, if implemented, I would envision would supply 
> a data pointer and length, and ask for the callback function to either 
> read that out (record) or fill it in (playback), and to do so without 
> calling into the PulseAudio API at all.
> 
> And with a big fat note that it is being called from a separate thread 
> and that the pa_simple API cannot be used from the callback function.

OK, I can see this being useful in the use case that Thomas talked
about: it makes it easier to put all audio processing into a separate
thread, since you save the thread and mainloop setup/teardown code, at
the cost of losing the ability to add any other event sources to the
mainloop. This is actually how the JACK API works, I think.

I'm not against adding this new functionality.

-- 
Tanu