Re: [PATCH v3 1/2] media: docs-rst: Document memory-to-memory video decoder interface

[Hans Verkuil]

On 1/31/19 1:44 PM, Philipp Zabel wrote:
> On Thu, 2019-01-31 at 13:30 +0100, Hans Verkuil wrote:
>> On 1/31/19 11:45 AM, Hans Verkuil wrote:
>>> On 1/24/19 11:04 AM, Tomasz Figa wrote:
>>>> Due to complexity of the video decoding process, the V4L2 drivers of
>>>> stateful decoder hardware require specific sequences of V4L2 API calls
>>>> to be followed. These include capability enumeration, initialization,
>>>> decoding, seek, pause, dynamic resolution change, drain and end of
>>>> stream.
>>>>
>>>> Specifics of the above have been discussed during Media Workshops at
>>>> LinuxCon Europe 2012 in Barcelona and then later Embedded Linux
>>>> Conference Europe 2014 in DÃsseldorf. The de facto Codec API that
>>>> originated at those events was later implemented by the drivers we already
>>>> have merged in mainline, such as s5p-mfc or coda.
>>>>
>>>> The only thing missing was the real specification included as a part of
>>>> Linux Media documentation. Fix it now and document the decoder part of
>>>> the Codec API.
>>>>
>>>> Signed-off-by: Tomasz Figa <tfiga@xxxxxxxxxxxx>
>>>> ---
>>>>  Documentation/media/uapi/v4l/dev-decoder.rst  | 1076 +++++++++++++++++
>>>>  Documentation/media/uapi/v4l/dev-mem2mem.rst  |    5 +
>>>>  Documentation/media/uapi/v4l/pixfmt-v4l2.rst  |    5 +
>>>>  Documentation/media/uapi/v4l/v4l2.rst         |   10 +-
>>>>  .../media/uapi/v4l/vidioc-decoder-cmd.rst     |   40 +-
>>>>  Documentation/media/uapi/v4l/vidioc-g-fmt.rst |   14 +
>>>>  6 files changed, 1135 insertions(+), 15 deletions(-)
>>>>  create mode 100644 Documentation/media/uapi/v4l/dev-decoder.rst
>>>>
>>>
>>> <snip>
>>>
>>>> +4.  **This step only applies to coded formats that contain resolution information
>>>> +    in the stream.** Continue queuing/dequeuing bitstream buffers to/from the
>>>> +    ``OUTPUT`` queue via :c:func:`VIDIOC_QBUF` and :c:func:`VIDIOC_DQBUF`. The
>>>> +    buffers will be processed and returned to the client in order, until
>>>> +    required metadata to configure the ``CAPTURE`` queue are found. This is
>>>> +    indicated by the decoder sending a ``V4L2_EVENT_SOURCE_CHANGE`` event with
>>>> +    ``V4L2_EVENT_SRC_CH_RESOLUTION`` source change type.
>>>> +
>>>> +    * It is not an error if the first buffer does not contain enough data for
>>>> +      this to occur. Processing of the buffers will continue as long as more
>>>> +      data is needed.
>>>> +
>>>> +    * If data in a buffer that triggers the event is required to decode the
>>>> +      first frame, it will not be returned to the client, until the
>>>> +      initialization sequence completes and the frame is decoded.
>>>> +
>>>> +    * If the client sets width and height of the ``OUTPUT`` format to 0,
>>>> +      calling :c:func:`VIDIOC_G_FMT`, :c:func:`VIDIOC_S_FMT`,
>>>> +      :c:func:`VIDIOC_TRY_FMT` or :c:func:`VIDIOC_REQBUFS` on the ``CAPTURE``
>>>> +      queue will return the ``-EACCES`` error code, until the decoder
>>>> +      configures ``CAPTURE`` format according to stream metadata.
>>>
>>> I think this should also include the G/S_SELECTION ioctls, right?
>>
>> I've started work on adding compliance tests for codecs to v4l2-compliance and
>> I quickly discovered that this 'EACCES' error code is not nice at all.
>>
>> The problem is that it is really inconsistent with V4L2 behavior: the basic
>> rule is that there always is a format defined, i.e. G_FMT will always return
>> a format.
>>
>> Suddenly returning an error is actually quite painful to handle because it is
>> a weird exception just for the capture queue of a stateful decoder if no
>> output resolution is known.
>>
>> Just writing that sentence is painful.
>>
>> Why not just return some default driver defined format? It will automatically
>> be updated once the decoder parsed the bitstream and knows the new resolution.
>>
>> It really is just the same behavior as with a resolution change.
>>
>> It is also perfectly fine to request buffers for the capture queue for that
>> default format. It's pointless, but not a bug.
>>
>> Unless I am missing something I strongly recommend changing this behavior.
> 
> I just wrote the same in my reply to Nicolas, the CODA driver currently
> sets the capture queue width/height to the output queue's crop rectangle
> (rounded to macroblock size) without ever having seen the SPS.

And thinking about the initial 0x0 width/height for the output queue:

that too is an exception, although less of a problem than the EACCES behavior.

It should be fine for an application to set width/height to 0 when calling
S_FMT for the output queue of the decoder, but I would also prefer that it is
just replaced by the driver with some default resolution. It really doesn't
matter in practice, since you will wait for the SOURCE_CHANGE event regardless.

Only then do you start to configure the CAPTURE queue.

Using 0x0 and EACCES looks good on paper, but in the code it is a hassle and
I'm not convinced there is any benefit.

I like generic APIs and no where else do we ever return a 0 value for width
or height, except in this corner case. It's just awkward.

Regards,

	Hans