[vlc-devel] [PATCH 00/20] MR: Finishing vout life cycle work

[Thomas Guillem]

On Thu, May 9, 2019, at 13:25, Rémi Denis-Courmont wrote:
> It will still deadlock because vout_Request() and aout_DecNew() still need to be interlocked, for the same reason as Laurent put the lock in the first place. Without it, we cannot sequence outputs creation/deletion as expected.

I don't understand, what do you mean by interlocked ?
aout_DecNew() is called not locked. It's called directly from the input/decoder.c

> 
> Well probably it could be separate for aout and vout. But it's still needed.
> 
> Le 9 mai 2019 14:10:33 GMT+03:00, Thomas Guillem <thomas at gllm.fr> a écrit :
>> I backported my "merge both locks" commit into 3.0 to reproduce the deadlock.
>> 
>> Indeed, it protected a deadlock when the following happened:
>> 
>> The GUI (Qt here) hold vouts while a vout is created from an other thread.
>> cf. https://code.videolan.org/tguillem/vlc/snippets/955/raw for the backtrace
>> 
>> 1/ A decoder thread request a vout, LOCK the input_resource mutex, and create a vout
>> 2/ The GUI is on its mainloop and WAIT the input_resource mutex.
>> 3/ From the decoder thread, The vout_Request() load a "vout window" plugin, that will emit and wait for the getVideoSlot() that need to be executed in the GUI thread
>> 
>> => Deadlock
>> 
>> But this won't happen on VLC 4.0 (I tested it to made sure) since vout_Request() is called unlocked. Only the vout_Create() is locked, but this call only load the "vout window" plugin without enabling it.
>> 
>> I updated my vout-life/6 branch to fix some issues: https://code.videolan.org/tguillem/vlc/commits/vout-life/6
>> 
>> Mainly related to the "es_out: terminate free vout in more places" commit. The window was not disabled when I was disabled video tracks from the UI.
>> 
>> Also, this branch do revert a commit from Rémi: "resource: atomically return vout on failure (fixes #22284)" only to fix the issue in a diffrent way from the "resource: respect vout order" commit.
>> 
>> On Thu, May 9, 2019, at 12:01, Thomas Guillem wrote:
>>> 
>>> On Thu, May 9, 2019, at 11:52, Rémi Denis-Courmont wrote:
>>>> No. Plugins are supposed to be reentrant, and at least the Qt window provider was before you disabled it. Besides, the resource lock cannot prevent reentry into a vout window activation callback anyway.
>>> 
>>> The Qt window provider is protected by the resource lock since it's created from vout_Create() that is locked. Indeed this lock is not enough since there can be several input_resource.
>>> 
>>>> -- 
>>>> Envoyé de mon appareil Android avec Courriel K-9 Mail. Veuillez excuser ma brièveté. 
>>>> _______________________________________________
>>>> vlc-devel mailing list
>>>> To unsubscribe or modify your subscription options:
>>>> https://mailman.videolan.org/listinfo/vlc-devel
>>> 
>>> _______________________________________________
>>> vlc-devel mailing list
>>> To unsubscribe or modify your subscription options:
>>> https://mailman.videolan.org/listinfo/vlc-devel
>> 
> 
> -- 
> Envoyé de mon appareil Android avec Courriel K-9 Mail. Veuillez excuser ma brièveté. 
> _______________________________________________
> vlc-devel mailing list
> To unsubscribe or modify your subscription options:
> https://mailman.videolan.org/listinfo/vlc-devel
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.videolan.org/pipermail/vlc-devel/attachments/20190509/d834250c/attachment.html>