On Thu, 14 Oct 2021 at 05:57, Chen-Yu Tsai [off-list ref] wrote:

On Thu, Oct 14, 2021 at 4:46 PM Ezequiel Garcia
[off-list ref] wrote:

quoted

On Thu, 14 Oct 2021 at 04:31, Chen-Yu Tsai [off-list ref] wrote:

quoted

On Wed, Oct 13, 2021 at 9:40 PM Nicolas Dufresne [off-list ref] wrote:

quoted

Le mercredi 13 octobre 2021 à 15:05 +0800, Chen-Yu Tsai a écrit :

quoted

Hi,

On Fri, Oct 8, 2021 at 11:42 PM Nicolas Dufresne [off-list ref] wrote:

quoted

Hi Chen-Yu,

thanks for looking into this.

Le vendredi 08 octobre 2021 à 18:04 +0800, Chen-Yu Tsai a écrit :

quoted

Hi everyone,

While working on the rkvdec H.264 decoder for ChromeOS, I noticed some
behavioral differences compared to Hantro and Cedrus:

1. The driver always overrides the sizeimage setting given by userspace
   for the output format. This results in insufficient buffer space when
   running the ChromeOS video_decode_accelerator_tests test program,
   likely due to a small initial resolution followed by dynamic
   resolution change.

2. Doesn't support dynamic resolution change.

This small series fixes both and aligns the behavior with the other two
stateless decoder drivers. This was tested on the downstream ChromeOS
5.10 kernel with ChromeOS. Also compiled tested on mainline but I don't
have any other RK3399 devices set up to test video stuff, so testing
would be very much appreciated.

Also, I'm not sure if user applications are required to check the value
of sizeimage upon S_FMT return. If the value is different or too small,
what can the application do besides fail? AFAICT it can't split the
data of one frame (or slice) between different buffers.

While most software out there just assumes that driver will do it right and
crash when it's not the case, application that do map the buffer to CPU must
read back the fmt structure as the drivers are all fail-safe and will modify
that structure to a set of valid value s for the context.

I believe what is happening in Chromium is that the decoder is opened with
some default settings, including the smallest viable resolution for the
output side, and the buffers allocated accordingly. When dynamic resolution
change happens, the decoder does not check if the current buffers are
sufficiently sized; it just assumes that they are. And when it starts
pushing data into the buffers, it realizes they are too small and fails.

The spec also says:

    Clients are allowed to set the sizeimage field for variable length
    compressed data flagged with V4L2_FMT_FLAG_COMPRESSED at ioctl
    VIDIOC_ENUM_FMT, but the driver may ignore it and set the value itself,
    or it may modify the provided value based on alignment requirements or
    minimum/maximum size requirements.

The spec only guarantees that the buffers are of sufficient size for the
resolution configured at the time they were allocated/requested.

So I think my first patch is a workaround for a somewhat broken userspace.
But it seems the other stateless drivers are providing similar behavior,
as I previously mentioned.

That's what I mean, this is not a driver bug strictly speaking (assuming it does
guaranty the buffer size is sufficient) but it is without your change
inconvenient, as userspace may be aware of the largest resolution it will
decode, and may want to allocate larger buffer upfront.

Thinking about this more, I think a few follow up fixes for each driver
are in order. The spec implies that the driver should override the value
should userspace give some unrealistic value, such as asking for a 256 byte
buffer for a 4K frame size.

Where is the spec implying that?

In Documentation/userspace-api/media/v4l/pixfmt-v4l2-mplane.rst:

    Clients are allowed to set the sizeimage field for variable length
    compressed data flagged with ``V4L2_FMT_FLAG_COMPRESSED`` at
    :ref:`VIDIOC_ENUM_FMT`, but the driver may ignore it and set the
    value itself, or it may modify the provided value based on
    alignment requirements or minimum/maximum size requirements.

I guess I read "minimum/maximum size requirements" a bit liberally.
Maybe it refers to how much buffer space the hardware can address or
program for each request?

My interpretation is that the driver may modify the value to adjust
for hardware requirements in terms of buffer alignment hardware requirements
and buffer size hardware requirements.

quoted

This is encoded content, so I'm really inclined to avoid this path.
Having the driver decide what is "unrealistic" would mean some
heuristics in the drivers for something that should really come from userspace.

And if the driver refuses to give adequate buffer space, then it's a bug?

Not sure what you have in mind here. S_FMT typically won't fail,
as the spec mentions
https://www.kernel.org/doc/html/v4.9/media/uapi/v4l/vidioc-g-fmt.html?highlight=s_fmt#c.VIDIOC_S_FMT.
So the only way a driver can refuse, as you mentioned,
is by overriding a parameter (which is what you are fixing).

"""
Drivers should not return an error code unless the type field is
invalid, this is a mechanism to fathom device capabilities and to
approach parameters acceptable for both the application and driver. On
success the driver may program the hardware, allocate resources and
generally prepare for data exchange.
"""

The driver should just use what application provided for the buffer sizeimage.
Let me quote your cover letter question:

"""
Also, I'm not sure if user applications are required to check the value
of sizeimage upon S_FMT return. If the value is different or too small,
what can the application do besides fail? AFAICT it can't split the
data of one frame (or slice) between different buffers.
"""

If the sizeimage is not adequate and the application knows this,
then it can only conclude the operation is not possible on this hardware.

However, if I understand correctly, that's why you are fixing rkvdec.
Because the rkvdec driver shouldn't override the value from the application,
since that may be wrong (besides being suboptimal).

Hope this makes sense...

Thanks,
Ezequiel

Regards
ChenYu

quoted

Thanks,
Ezequiel

quoted

Cedrus (CCing Jernej) comes close, but a 1K buffer might not be enough for
really large frames, even though it's slice based?

ChenYu

quoted

As per Chromium bug, this is being addressed already. Thanks for this driver
improvement.

quoted

quoted

As for opposite direction (output vs capture) format being changed, this should
be documented in the spec, if you find it too unclear or missing for sateless
codec (I know it's there for stateful but can't remember, would have to re-read,
for stateless) let us know.

AFAICT the capture side is working OK and to spec.


Regards
ChenYu

quoted

regards,
Nicolas

quoted

Andrzej, I believe the second patch would conflict with your VP9 series.


Regards
ChenYu

Chen-Yu Tsai (2):
  media: rkvdec: Do not override sizeimage for output format
  media: rkvdec: Support dynamic resolution changes

 drivers/staging/media/rkvdec/rkvdec-h264.c |  5 +--
 drivers/staging/media/rkvdec/rkvdec.c      | 40 +++++++++++-----------
 2 files changed, 23 insertions(+), 22 deletions(-)

_______________________________________________
Linux-rockchip mailing list
Linux-rockchip@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-rockchip