[FFmpeg-devel] [RESEND PATCH] avcodec/v4l2_context: always log POLLERR when buffers are uninitialized

[Richard Acayan]

On Sun, Jul 23, 2023 at 10:15:50AM +0200, Marton Balint wrote:
>
>
> On Fri, 21 Jul 2023, Richard Acayan wrote:
>
>> Will this patch be applied or receive any comments? I have been waiting
>> more than 2 weeks since the original submission
>> (https://ffmpeg.org/pipermail/ffmpeg-devel/2023-July/311580.html) and
>> have not received a response since.
>> 
>> I am trying to get mpv to work again with the v4l2_m2m codec, as it
>> started emitting segmentation faults. This is partly due to some
>> incorrect usage of libavcodec in mpv, and partly due to the
>> error/warning handling in libavcodec. An mpv maintainer dismissed this
>> as a bug in ffmpeg, and this change is simpler than what mpv would need
>> for a fix, so I am hoping that this patch goes in.
>> 
>> I don't have write access so I can't apply it myself, even if there are
>> no objections.
>> 
>> On Tue, Jul 18, 2023 at 06:00:17PM -0400, Richard Acayan wrote:
>>> The error handler for POLLERRs assumes that the timeout is only zero
>>> when v4l2_dequeue_v4l2buf is called by v4l2_getfree_v4l2buf. This
>>> assumption is incorrect, as ff_v4l2_context_dequeue_frame also calls
>>> this function with a timeout of zero. Do not check for unavailable
>>> buffers if the buffers are uninitialized.
>>> 
>>> See https://trac.ffmpeg.org/ticket/9957 for the original bug report.
>>> 
>>> Signed-off-by: Richard Acayan <mailingradian at gmail.com>
>>> ---
>>>  libavcodec/v4l2_context.c | 2 +-
>>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>> 
>>> diff --git a/libavcodec/v4l2_context.c b/libavcodec/v4l2_context.c
>>> index a40be94690..69ddf80723 100644
>>> --- a/libavcodec/v4l2_context.c
>>> +++ b/libavcodec/v4l2_context.c
>>> @@ -327,7 +327,7 @@ start:
>>>      if (pfd.revents & POLLERR) {
>>>          /* if we are trying to get free buffers but none have been queued yet
>>>             no need to raise a warning */
>>> -        if (timeout == 0) {
>>> +        if (timeout == 0 && ctx->buffers) {
>
> This would warn if no buffers are allocated yet, but this is not a warning
> condition as far as I understand the issue, simply a different API usage
> pattern.

I assumed that the usage was "incorrect" because the V4L2 API didn't
like it.

> So maybe it is better to simply do a
>
> if (!ctx->buffers)
>    return NULL;
>
> in the beginning of the if (timeout == 0) block.

The sequence in mpv seems to be supported by libavcodec though, so I
will do this in v2.

>
> Thanks,
> Marton
>
>>>              for (i = 0; i < ctx->num_buffers; i++) {
>>>                  if (ctx->buffers[i].status != V4L2BUF_AVAILABLE)
>>>                      av_log(logger(ctx), AV_LOG_WARNING, "%s POLLERR\n", ctx->name);
>>> --
>>> 2.41.0
>>>