[Spice-devel] [virtio-dev] Re: [PATCH v2 1/1] virtio-video: Add virtio video device specification

[Gerd Hoffmann]

On Wed, Dec 18, 2019 at 11:08:37PM +0900, Tomasz Figa wrote:
> On Wed, Dec 18, 2019 at 10:40 PM Gerd Hoffmann <kraxel at redhat.com> wrote:
> >
> >   Hi,
> >
> > > +The device MUST mark the last buffer with the
> > > +VIRTIO_VIDEO_BUFFER_F_EOS flag to denote completion of the drain
> > > +sequence.
> >
> > No, that would build a race condition into the protocol.  The device
> > could complete the last buffer after the driver has sent the drain
> > command but before the device saw it.  So the flag would not be
> > reliable.
> >
> > I also can't see why the flag is needed in the first place.  The driver
> > should know which buffers are queued still and be able to figure
> > whenever the drain is complete or not without depending on that flag.
> > So I'd suggest to simply drop it.
> 
> Unfortunately video decoders are not that simple. There are always
> going to be some buffers on the decoder side used as reference frames.
> Only the decoder knows when to release them, as it continues decoding
> the stream.

Not clearly defined in the spec:  When is the decoder supposed to send
the response for a queue request?  When it finished decoding (i.e. frame
is ready for playback), or when it doesn't need the buffer any more for
decoding (i.e. buffer can be re-queued or pages can be released)?

> How we defined this in the V4L2 stateful decoder interface is that if
> the decoder happened to return the last framebuffer before the drain
> request arrived, it would return one more, empty.

Ok.  That is not clear from the spec.  I've read the drain request as as
"please stop decoding and complete all queue requests which are in the
virtqueue, even when you didn't process them yet".  In which case
completing the last pending queue request would clearly indicate the
draining request is complete.  Also completing the drain request should
only happen when the operation is finished.

I think the various states a buffer can have and how queuing & deciding
& draining changes these states should be clarified in the
specification.

cheers,
  Gerd