[FFmpeg-devel] [PATCH v4 3/9] avformat/utils: Move the reference to the packet list

[James Almer]

On 9/25/2019 7:06 PM, Andreas Rheinhardt wrote:
> James Almer:
>> On 9/20/2019 5:39 PM, Andreas Rheinhardt wrote:
>>> Up until now, ff_packet_list_put had a flaw: When it moved a packet to
>>> the list (meaning, when it ought to move the reference to the packet
>>> list instead of creating a new one via av_packet_ref), it did not reset
>>> the original packet, confusing the ownership of the data in the packet.
>>> This has been done because some callers of this function were not
>>> compatible with resetting the packet.
>>>
>>> This commit changes these callers and fixes this flaw. In order to
>>> indicate that the ownership of the packet has moved to the packet list,
>>> pointers to constant AVPackets are used whenever the target of the
>>> pointer might already be owned by the packet list.
>>>
>>> Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt at gmail.com>
>>> ---
>>>  libavformat/utils.c | 32 +++++++++++++++++---------------
>>>  1 file changed, 17 insertions(+), 15 deletions(-)
>>>
>>> diff --git a/libavformat/utils.c b/libavformat/utils.c
>>> index e348df3269..58e0db61da 100644
>>> --- a/libavformat/utils.c
>>> +++ b/libavformat/utils.c
>>> @@ -460,10 +460,7 @@ int ff_packet_list_put(AVPacketList **packet_buffer,
>>>              return ret;
>>>          }
>>>      } else {
>>> -        // TODO: Adapt callers in this file so the line below can use
>>> -        //       av_packet_move_ref() to effectively move the reference
>>> -        //       to the list.
>>> -        pktl->pkt = *pkt;
>>> +        av_packet_move_ref(&pktl->pkt, pkt);
>>
>> I think it would be a good idea ensuring the packet is reference counted
>> here, to have more robust code and not depend on current callers having
>> done it themselves.
>> It's a matter of adding an av_packet_make_refcounted() call before this
>> line, much like it's done in av_bsf_send_packet(). It's essentially a
>> no-op when the packet is already reference counted.
>>
> I can add a commit doing this. But there is at least one potential
> use-case where using av_packet_make_refcounted() is not good: Uncoded
> frames. They aren't currently used in conjunction with this function,
> but that may change of course. Should I add a check for this scenario
> or another FF_PACKETLIST_FLAG_* flag that says that the packet should
> not be made refcounted, but simply moved?

No, moving a packet to the list should give full ownership of the packet
to the list, and the data described in them shouldn't just suddenly
become invalid because some buffer was freed by some other function.
That's precisely what happened with the bitstream filter API, generating
filtering and decoding failures because of how ffmpeg.c used to handle
packets internally, until an av_packet_make_refcounted() call was added
to av_bsf_send_packet().

And i personally don't consider adding a check like the one you suggest
something worth doing. If you want you can add a mention to
ff_packet_list_put() doxy that it will ensure the packet will be refcounted.