Re: [PATCH v9 10/15] vb2: add explicit fence user API

[Ezequiel Garcia <ezequiel@collabora.com>]

On Wed, 2018-05-09 at 17:33 +0100, Brian Starkey wrote:
> On Wed, May 09, 2018 at 12:52:26PM -0300, Ezequiel Garcia wrote:
> > On Wed, 2018-05-09 at 11:33 +0100, Brian Starkey wrote:
> > > Hi Ezequiel,
> > > 
> > > On Fri, May 04, 2018 at 05:06:07PM -0300, Ezequiel Garcia wrote:
> > > > From: Gustavo Padovan <gustavo.padovan@collabora.com>
> > > > 
> > > > Turn the reserved2 field into fence_fd that we will use to send
> > > > an in-fence to the kernel or return an out-fence from the kernel to
> > > > userspace.
> > > > 
> > > > Two new flags were added, V4L2_BUF_FLAG_IN_FENCE, that should be used
> > > > when sending an in-fence to the kernel to be waited on, and
> > > > V4L2_BUF_FLAG_OUT_FENCE, to ask the kernel to give back an out-fence.
> > > > 
> > > > v7: minor fixes on the Documentation (Hans Verkuil)
> > > > 
> > > > v6: big improvement on doc (Hans Verkuil)
> > > > 
> > > > v5: - keep using reserved2 field for cpia2
> > > >    - set fence_fd to 0 for now, for compat with userspace(Mauro)
> > > > 
> > > > v4: make it a union with reserved2 and fence_fd (Hans Verkuil)
> > > > 
> > > > v3: make the out_fence refer to the current buffer (Hans Verkuil)
> > > > 
> > > > v2: add documentation
> > > > 
> > > > Signed-off-by: Gustavo Padovan <gustavo.padovan@collabora.com>
> > > > ---
> > > > Documentation/media/uapi/v4l/buffer.rst         | 45 +++++++++++++++++++++++--
> > > > drivers/media/common/videobuf2/videobuf2-v4l2.c |  2 +-
> > > > drivers/media/v4l2-core/v4l2-compat-ioctl32.c   |  4 +--
> > > > include/uapi/linux/videodev2.h                  |  8 ++++-
> > > > 4 files changed, 52 insertions(+), 7 deletions(-)
> > > > 
> > > > diff --git a/Documentation/media/uapi/v4l/buffer.rst b/Documentation/media/uapi/v4l/buffer.rst
> > > > index e2c85ddc990b..be9719cf5745 100644
> > > > --- a/Documentation/media/uapi/v4l/buffer.rst
> > > > +++ b/Documentation/media/uapi/v4l/buffer.rst
> > > > @@ -301,10 +301,22 @@ struct v4l2_buffer
> > > > 	elements in the ``planes`` array. The driver will fill in the
> > > > 	actual number of valid elements in that array.
> > > >     * - __u32
> > > > -      - ``reserved2``
> > > > +      - ``fence_fd``
> > > >       -
> > > > -      - A place holder for future extensions. Drivers and applications
> > > > -	must set this to 0.
> > > > +      - Used to communicate a fence file descriptors from userspace to kernel
> > > > +	and vice-versa. On :ref:`VIDIOC_QBUF <VIDIOC_QBUF>` when sending
> > > > +	an in-fence for V4L2 to wait on, the ``V4L2_BUF_FLAG_IN_FENCE`` flag must
> > > > +	be used and this field set to the fence file descriptor of the in-fence.
> > > > +	If the in-fence is not valid ` VIDIOC_QBUF`` returns an error.
> > > > +
> > > > +        To get an out-fence back from V4L2 the ``V4L2_BUF_FLAG_OUT_FENCE``
> > > > +	must be set, the kernel will return the out-fence file descriptor in
> > > > +	this field. If it fails to create the out-fence ``VIDIOC_QBUF` returns
> > > > +        an error.
> > > > +
> > > > +	For all other ioctls V4L2 sets this field to -1 if
> > > > +	``V4L2_BUF_FLAG_IN_FENCE`` and/or ``V4L2_BUF_FLAG_OUT_FENCE`` are set,
> > > > +	otherwise this field is set to 0 for backward compatibility.
> > > >     * - __u32
> > > >       - ``reserved``
> > > >       -
> > > > @@ -648,6 +660,33 @@ Buffer Flags
> > > >       - Start Of Exposure. The buffer timestamp has been taken when the
> > > > 	exposure of the frame has begun. This is only valid for the
> > > > 	``V4L2_BUF_TYPE_VIDEO_CAPTURE`` buffer type.
> > > > +    * .. _`V4L2-BUF-FLAG-IN-FENCE`:
> > > > +
> > > > +      - ``V4L2_BUF_FLAG_IN_FENCE``
> > > > +      - 0x00200000
> > > > +      - Ask V4L2 to wait on the fence passed in the ``fence_fd`` field. The
> > > > +	buffer won't be queued to the driver until the fence signals. The order
> > > > +	in which buffers are queued is guaranteed to be preserved, so any
> > > > +	buffers queued after this buffer will also be blocked until this fence
> > > > +	signals. This flag must be set before calling ``VIDIOC_QBUF``. For
> > > > +	other ioctls the driver just reports the value of the flag.
> > > > +
> > > > +        If the fence signals the flag is cleared and not reported anymore.
> > > > +	If the fence is not valid ``VIDIOC_QBUF`` returns an error.
> > > > +
> > > > +
> > > > +    * .. _`V4L2-BUF-FLAG-OUT-FENCE`:
> > > > +
> > > > +      - ``V4L2_BUF_FLAG_OUT_FENCE``
> > > > +      - 0x00400000
> > > > +      - Request for a fence to be attached to the buffer. The driver will fill
> > > > +	in the out-fence fd in the ``fence_fd`` field when :ref:`VIDIOC_QBUF
> > > > +	<VIDIOC_QBUF>` returns. This flag must be set before calling
> > > > +	``VIDIOC_QBUF``. For other ioctls the driver just reports the value of
> > > > +	the flag.
> > > > +
> > > > +        If the creation of the out-fence fails ``VIDIOC_QBUF`` returns an
> > > > +	error.
> > > > 
> > > 
> > > I commented similarly on some of the old patch-sets, and it's a minor
> > > thing, but I still think the ordering of this series is off. It's
> > > strange/wrong to me document all this behaviour, and expose the flags
> > > to userspace, when the functionality isn't implemented yet.
> > > 
> > > If I apply this patch to the kernel, then the kernel doesn't do what
> > > the (newly added) kernel-doc says it will.
> > > 
> > 
> > This has never been a problem, and it has always been the canonical
> > way of doing things.
> > 
> > First the required macros, stubs, documentation and interfaces are added,
> > and then they are implemented.
> 
> If you say so, I don't know what sets the standard but that seems
> kinda backwards.
> 
> I'd expect the "flick the switch, expose to userspace" to always be
> the last thing, but I'm happy to be shown examples to the contrary.
> 
> > 
> > I see no reason to go berserk here, unless you see an actual problem?
> > Or something actually broken?
> > 
> 
> The only "broken" thing, is as I said - I can apply this patch to a
> kernel (any kernel, because there's no dependencies in the code), and
> it won't do what the kernel-doc says it will.
> 

I don't think we've ever honored that, but I can be wrong too.

> Maybe I'm crazy, but shouldn't comments at least be correct at the
> point where they are added, even if they become incorrect later
> through neglect?
> 

This is the best example I can give, a similar policy in the
devicetree bindings [1]:

""
  3) The Documentation/ portion of the patch should come in the series before
     the code implementing the binding.
""

[1]  Documentation/devicetree/bindings/submitting-patches.txt