[FFmpeg-devel] [PATCH] avcodec/bsf: make sure the AVBSFInternal stored packet is reference counted

[James Almer]

On 3/23/2018 8:15 PM, James Almer wrote:
> On 3/23/2018 7:46 PM, wm4 wrote:
>> On Fri, 23 Mar 2018 18:38:22 -0300
>> James Almer <jamrial at gmail.com> wrote:
>>
>>> Some bitstream filters may buffer said packet in their own contexts
>>> for latter use.
>>> The documentation for av_bsf_send_packet() doesn't forbid feeding
>>> it non-reference counted packets, which depending on the way said
>>> packets were internally buffered by the bsf it may result in the
>>> data described in them to become invalid or unavailable at any time.
>>>
>>> This was the case with vp9_superframe after commit e1bc3f4396, which
>>> was then promptly fixed in 37f4a093f7 and 7a02b364b6. It is still the
>>> case even today with vp9_reorder_raw.
>>>
>>> With this change the bitstream filters will not have to worry how to
>>> store or consume the packets fed to them.
>>>
>>> Signed-off-by: James Almer <jamrial at gmail.com>
>>> ---
>>> Regarding vp9_raw_reorder, see "[PATCH] avcodec/vp9_raw_reorder: cache
>>> input packets using new references" for a local fix similar to what
>>> vp9_superframe got with 37f4a093f7 and 7a02b364b6.
>>>
>>> A simple reproducer if you're curious is:
>>>
>>> ffmpeg -i INPUT -c:v copy -bsf:v vp9_raw_reorder -f null -
>>>
>>> Which segfauls with current git master.
>>>
>>> "[PATCH 2/2] ffmpeg: pass reference counted packets on codec copy
>>> when possible" also works around this in most cases by doing what its
>>> subject describes, but only affects the ffmpeg CLI only and not the
>>> API itself, of course.
>>>
>>>  libavcodec/bsf.c | 10 +++++++++-
>>>  1 file changed, 9 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/libavcodec/bsf.c b/libavcodec/bsf.c
>>> index 38b423101c..25f7a20ad6 100644
>>> --- a/libavcodec/bsf.c
>>> +++ b/libavcodec/bsf.c
>>> @@ -188,7 +188,15 @@ int av_bsf_send_packet(AVBSFContext *ctx, AVPacket *pkt)
>>>          ctx->internal->buffer_pkt->side_data_elems)
>>>          return AVERROR(EAGAIN);
>>>  
>>> -    av_packet_move_ref(ctx->internal->buffer_pkt, pkt);
>>> +    if (pkt->buf)
>>> +        av_packet_move_ref(ctx->internal->buffer_pkt, pkt);
>>> +    else {
>>> +        int ret = av_packet_ref(ctx->internal->buffer_pkt, pkt);
>>> +
>>> +        if (ret < 0)
>>> +            return ret;
>>> +        av_packet_unref(pkt);
>>> +    }
>>>  
>>>      return 0;
>>>  }
>>
>> Fine, but we should probably really provide an API function that
>> ensures that a packet is refcounted (and made refcounting if it isn't).
>> av_dup_packet() does this, but it's deprecated and has a bad name.
> 
> av_packet_ref() ensures that, and so does av_packet_make_writable(), but
> as a side effect of their main intended use. The documentation even
> states to use av_packet_ref() for that purpose.
> 
> If you want one exactly like av_dup_packet() but in a less hacky way
> that exclusively does "Make this packet's data ref counted if it's not,
> do nothing else", we could add an av_packet_make_refcounted() function
> or whatever. It should be trivial, so just tell me a name you'd like for
> it and I'll write it.

Patch pushed in the meantime.