[FFmpeg-devel] [PATCH v3 2/3] lavf/dashdec: Multithreaded DASH initialization

[Lukas Fellechner]

Gesendet: Montag, 05. September 2022 um 12:45 Uhr
Von: "Andreas Rheinhardt" <andreas.rheinhardt at outlook.com>
An: ffmpeg-devel at ffmpeg.org
Betreff: Re: [FFmpeg-devel] [PATCH v3 2/3] lavf/dashdec: Multithreaded DASH initialization
Lukas Fellechner:
>>> Moreover, older pthread standards did not allow to use
>>> PTHREAD_MUTEX_INITIALIZER with non-static mutexes, so I don't know
>>> whether we can use that. Also our pthreads-wrapper on top of
>>> OS/2-threads does not provide PTHREAD_COND_INITIALIZER (which is used
>>> nowhere in the codebase).
>>
>> I missed that detail about the initializer macro. Thank you for clearing
>> that up.
>>
>> After looking more into threads implementation in ffmpeg, I wonder if I
>> really need to check any results of init/destroy or other functions.
>> In slicethreads.c, there is zero checking on any of the lock functions.
>> The pthreads-based implementation does internally check the results of all
>> function calls and calls abort() in case of errors ("strict_" wrappers).
>> The Win32 implementation uses SRW locks which cannot even return errors.
>> And the OS2 implementation returns 0 on all calls as well.
>>
>> So right now, I think that I should continue with normal _init() calls
>> (no macros) and drop all error checking, just like slicethread does.
>> Are you fine with that approach?
> 
> Zero checking is our old approach; the new approach checks for errors
> and ensures that only mutexes/condition variables that have been
> properly initialized are destroyed. See ff_pthread_init/free in
> libavcodec/pthread.c (you can't use this in libavformat, because these
> functions are local to libavcodec).
> 
> - Andreas

I was able to switch to using the slicethread implementation. It has a
very minor delay on init, because it waits for all threads to fully start
up before continueing. But it is only few ms and not worth adding a new
implementation just for that.

I also changed the initialization and release of mutex and conds, with
full return code checking and safe release.

There was one cross-thread issue I needed to address.

A multi hour duration test (connecting in endless loop) did not show
any issues after fixing the avio_opts cross thread access.

Please see the v4 patch for all the changes.

- Lukas