[vlc-devel] [PATCH] darwin/thread: Ignore vlc_mutex_unlock failures
remi at remlab.net
Tue Oct 29 08:24:24 CET 2019
Did you investigate all crash reports to verify that none of them were not caused by the unspecified constraints that you claim exist? Did you check what the implementations on affected macOS versions actually does apply those constraints? All that I see is that the Darwin doc contradicts your statement.
Le 29 octobre 2019 09:10:36 GMT+02:00, Thomas Guillem <thomas at gllm.fr> a écrit :
>OK, it's an UB, but on a constrained environment, that is macOS. For
>now, this UB seem to have 0 impact on users. As Marvin said, we had
>this issue since 7 years, VLC 3.x and 2.x users don't seem to be
>bothered by this.
>We can't abort VLC for an UB that won't have any impact.
>I'm OK with this patch and with a 3.0 backport. It is was I do locally
>every time I develop on macOS (in fact, I just define NDEBUG for this
>One other argument in favor: this may be only temporary since Marvin
>finally reported this issue to Apple.
>On Tue, Oct 29, 2019, at 04:26, Rémi Denis-Courmont wrote:
>> Le maanantaina 28. lokakuuta 2019, 22.35.49 EET Marvin Scholz a écrit
>> > On 28 Oct 2019, at 18:46, Rémi Denis-Courmont wrote:
>> > > Le maanantaina 28. lokakuuta 2019, 19.24.44 EET Marvin Scholz a
>> > >
>> > >> macOS pthread implementation for pthread_cond_wait is buggy,
>> > >> sometimes the mutex to not be locked when running thread
>> > >> cleanup handlers.
>> > >> This causes random failures of vlc_mutex_unlock, as the
>> > >> cleanup handler does not hold a lock on the mutex in rare cases.
>> > >
>> > > This won't work:
>> > >
>> > > 1) Darwin does not define unlocking a mutex by a thread that did
>> > > lock it:
>> > > "Calling pthread_mutex_unlock with a mutex that the calling
>> > > does not
>> > > hold will result in undefined behavior." So you cannot simply
>> > > that it
>> > > will return an error and have no side effects; it probably only
>> > > that way
>> > > on sunny day cases.
>> > >
>> > > 2) Even if Darwin did provide that guarantee, this would cause
>> > > races if
>> > > the calling thread actually relied on the mutex being locked to
>> > > lock-
>> > > protected data.
>> > Hi, thanks for the review.
>> > I am aware of these issues, the intention of this is just to make
>> > enabled
>> > builds on macOS work again without randomly crashing all the time.
>> No, you don't seem to be aware of what UB means. You cannot recover
>> Period. This is too late in code flow to put a workaround; we always
>> workarounds before the bug, not after it.
>> The only ways to handle UB is to 1) avoid getting there in the first
>> 2) abort as simply and quickly as possible if it happens anyway. It's
>> about debugging, it's about security/safety.
>> If your argument is that the error actual happens, the correct
>solution is to
>> replace the assert with an abort. Otherwise you concretely might as
>> deadlock, or fail the following lock or destroy, which are even worst
>> vlc-devel mailing list
>> To unsubscribe or modify your subscription options:
>vlc-devel mailing list
>To unsubscribe or modify your subscription options:
Envoyé de mon appareil Android avec Courriel K-9 Mail. Veuillez excuser ma brièveté.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the vlc-devel