Hi Frédéric,

On 08/30/2013 05:20 AM, Frédéric DALLEAU wrote:
> Hi Denis,
>
>>> +    /* Ask main loop to join */
>>> +    g_idle_add(thread_cleanup, userdata);
>>
>> This part makes no sense to me.  We have two broad cases being covered:
>> 1. Cleaning up the threads on exit, in which case this g_idle_add is
>> pointless since there is no more main loop to run
>> 2. Cleaning up the thread when an exceptional condition occurs, e.g.
>> socket is closed.
>
> There would be a race condition if the thread had to remove itself from
> the threads list while, at same time, the main loop would parse the
> threads list. This is possible if exit and socket close occurs
> simultaneously.
> Thus the threads list is handled from main thread only. This way,
> hfp_thread structure is guaranted to be accessible during thread's life.
>

Of course there would be a race condition if you do not lock the list 
and try to modify it from multiple threads.  What I'm saying is trigger 
the appropriate handler code for the appropriate condition in the main 
thread.

Right now you are always calling g_idle_add when the main loop has 
exited and being cleaned up.  g_idle_add triggers at least two mutex 
locks / unlocks and adding an event to the main loop queue that will 
never be called.  It also makes the code way less readable or 
understandable.

In fact it might be easier to simply use a lock around the thread list 
instead of trying to play main event tricks.  You're not gaining much by 
doing that in the first place.

>> You might want to separate the logic for each appropriately.
>>
>> A more broad question is, why are you using threads in the first place?
>>   Can't we interact with ALSA via regular FD polling?
>
> In theory, snd_pcm_poll_descriptors can be used for that.
> But this depends on ALSA driver implementation and I got some weird
> behavior.
> The sound was choppy using glib polling, so I used threads.

Are you sure this has anything to do with the driver implementation? 
Quickly browsing the ALSA documentation tells me that you're using basic 
read/write operations on an fd, so there should be nothing preventing 
you from using poll/select here.  There are no warnings about driver 
support, only about properly using the revents structure:

"...do not use only returned file descriptors, but handle events member 
as well - see snd_pcm_poll_descriptors function description for more 
details and snd_pcm_poll_descriptors_revents for events demangling)."

Regards,
-Denis