[FFmpeg-devel] [PATCH] avcodec/encode: only allow undersized audio frames if they are the last

[Marton Balint]

On Sun, 4 Aug 2019, Marton Balint wrote:

>
>
> On Sat, 3 Aug 2019, Lynne wrote:
>
>> Aug 3, 2019, 9:40 PM by cus at passwd.hu:
>>
>>> Otherwise the user might get a silence padded frame in the beginning or in 
> the
>>> middle of the encoding.
>>>
>>> Some other bug uncovered this:
>>>
>>> ./ffmpeg -loglevel verbose -y -f data -i /dev/zero \
>>> -filter_complex "nullsrc=s=60x60:d=10[v0];sine=d=10[a]" \
>>> -map '[v0]' -c:v:0 rawvideo \
>>> -map '[a]'  -c:a:0 mp2 \
>>> -f mpegts out.ts
>>>
>>> Signed-off-by: Marton Balint <cus at passwd.hu>
>>> ---
>>>  libavcodec/encode.c | 10 ++++++++--
>>>  1 file changed, 8 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/libavcodec/encode.c b/libavcodec/encode.c
>>> index d12c42526b..d81b32b983 100644
>>> --- a/libavcodec/encode.c
>>> +++ b/libavcodec/encode.c
>>> @@ -174,8 +174,14 @@ int attribute_align_arg 
> avcodec_encode_audio2(AVCodecContext *avctx,
>>>  goto end;
>>>  }
>>>  } else if (!(avctx->codec->capabilities & 
> AV_CODEC_CAP_VARIABLE_FRAME_SIZE)) {
>>> -            if (frame->nb_samples < avctx->frame_size &&
>>> -                !avctx->internal->last_audio_frame) {
>>> +            /* if we already got an undersized frame, that must have been 
> the last */
>>> +            if (avctx->internal->last_audio_frame) {
>>> +                av_log(avctx, AV_LOG_ERROR, "frame_size (%d) was not 
> respected for a non-last frame (avcodec_encode_audio2)\n", 
> avctx->frame_size);
>>> +                ret = AVERROR(EINVAL);
>>> +                goto end;
>>> +            }
>>> +
>>> +            if (frame->nb_samples < avctx->frame_size) {
>>>  ret = pad_last_frame(avctx, &padded_frame, frame);
>>>  if (ret < 0)
>>>  goto end;
>>>
>>
>> You haven't fixed the bug, you've just made it an error. You have to fix 
> the encoder instead.
>> The check might be useful if its made a warning instead though.
>
> The command line was also failing before the patch with an invalid frame 
> size error at the end of the stream. So ffmpeg.c / libavfilter not 
> providing fixed frame sizes is a different bug to the one this patch 
> supposed to fix.
>
> This fix disallows padding a frame with silence in the middle of the 
> encoding if the user provides an undersized frame. libavcodec should 
> reject such encoding attempts and not corrupt the output with silence.
>

Ping, will apply soon.

Regards,
Marton