[FFmpeg-devel] [PATCH] avcodec/avcodec: don't free AVOption settable fields in avcodec_close()

[Andreas Rheinhardt]

James Almer:
> On 3/20/2022 8:34 PM, Andreas Rheinhardt wrote:
>> James Almer:
>>>
>>>
>>> On 3/20/2022 8:26 PM, Andreas Rheinhardt wrote:
>>>> James Almer:
>>>>> It can uninitialize fields that may still be used after the context
>>>>> was closed,
>>>>> so do it instead in avcodec_free_context().
>>>>>
>>>>> Signed-off-by: James Almer <jamrial at gmail.com>
>>>>> ---
>>>>>    libavcodec/avcodec.c | 1 -
>>>>>    libavcodec/options.c | 2 +-
>>>>>    2 files changed, 1 insertion(+), 2 deletions(-)
>>>>>
>>>>> diff --git a/libavcodec/avcodec.c b/libavcodec/avcodec.c
>>>>> index 38bdaad4fa..122d09b63a 100644
>>>>> --- a/libavcodec/avcodec.c
>>>>> +++ b/libavcodec/avcodec.c
>>>>> @@ -524,7 +524,6 @@ av_cold int avcodec_close(AVCodecContext *avctx)
>>>>>          if (avctx->priv_data && avctx->codec &&
>>>>> avctx->codec->priv_class)
>>>>>            av_opt_free(avctx->priv_data);
>>>>> -    av_opt_free(avctx);
>>>>>        av_freep(&avctx->priv_data);
>>>>>        if (av_codec_is_encoder(avctx->codec)) {
>>>>>            av_freep(&avctx->extradata);
>>>>> diff --git a/libavcodec/options.c b/libavcodec/options.c
>>>>> index 33f11480a7..91335415c1 100644
>>>>> --- a/libavcodec/options.c
>>>>> +++ b/libavcodec/options.c
>>>>> @@ -172,7 +172,7 @@ void avcodec_free_context(AVCodecContext **pavctx)
>>>>>        av_freep(&avctx->intra_matrix);
>>>>>        av_freep(&avctx->inter_matrix);
>>>>>        av_freep(&avctx->rc_override);
>>>>> -    av_channel_layout_uninit(&avctx->ch_layout);
>>>>> +    av_opt_free(avctx);
>>>>>          av_freep(pavctx);
>>>>>    }
>>>>
>>>> This will lead to memleaks for users that use avcodec_close(avctx) +
>>>> av_free(avctx) to free an AVCodecContext (e.g. our frame-threaded
>>>> encoders do this). Notice that avcodec_free_context() violates the
>>>> documentation of AVCodecContext.extradata (documented to not be freed
>>>> for decoders) and AVCodecContext.subtitle_header and
>>>> AVCodecContext.rc_override (documented to not be freed by lavc for
>>>> encoders), so there is a reason for using it instead of
>>>> avcodec_free_context() (even when not reusing the context).
>>>
>>> That's an absolute mess of a situation. av_free(avctx) should not be an
>>> allowed or supported scenario when avcodec_free_context() exists. And
>>> why is the latter violating its own documentation?
>>>
>>
>> It is not violating its own documentation, but the documentation of the
>> relevant AVCodecContext fields. IIRC Anton wanted a function that just
>> frees the whole context, even if this meant that fields which are
>> documented as being owned by the user are freed. Even documenting the
>> current state of affairs in avcodec.h doesn't change the fact that there
>> is a valid reason to use avcodec_close()+av_free(), so we can't pretend
>> it doesn't happen.
>>
>> - Andreas
> 
> Ok, do i add a codecpar copy like i suggested in
> http://ffmpeg.org/pipermail/ffmpeg-devel/2022-March/294312.html, then?
> It works, but it feels really weird doing that in what's the cleanup
> portion of the function.
> Alternatively, add the dance from
> https://patchwork.ffmpeg.org/project/ffmpeg/patch/20220319030407.45503-1-jamrial@gmail.com/
> which should have the same effect and never fail, unlike param copy.

The other place where the internal context is reinitialized (in
read_frame_internal()) also calls avcodec_parameters_to_context(); yet
this code is reached when the demuxer updates the AVCodecParameters and
sets need_context_update accordingly whereas the other call to
avcodec_close() happens when no such updates were triggered. So I can
live with both.
(Is it actually intended for AVChannelLayout to be movable? If so, it
should be documented.)

- Andreas