[Spice-devel] [PATCH RFC v4 0/1] Virtio Video Device Specification

[Michael S. Tsirkin]

On Fri, Jul 03, 2020 at 02:45:15PM +0900, Alexandre Courbot wrote:
> Hi Dmitry,
> 
> On Thu, Jul 2, 2020 at 10:47 PM Dmitry Sepp <dmitry.sepp at opensynergy.com> wrote:
> >
> > Hi Keiichi,
> >
> > Thanks for the clarification. I believe we should explicitly describe this in
> > the VIRTIO_VIDEO_CMD_RESOURCE_ATTACH section. And I also still see a problem
> > there. If it is a guest allocated resource, we cannot consider it to be free
> > until the device really releases it. And it won't happen until we issue the
> > next ATTACH call. Also, as we discussed before, it might be not possible to
> > free individual buffers, but the whole queue only.
> 
> In the case of the encoder, a V4L2 driver is not supposed to let
> user-space dequeue an input frame while it is used as reference for
> the encoding process. So if we add a similar rule in the virtio-video
> specification, I suppose this would solve the problem?
> 
> For the decoder case, we have a bigger problem though. Since DMABUFs
> can be arbitrarily attached to any V4L2 buffer ID, we may end up
> re-queueing the DMABUF of a decoded frame that is still used as
> reference under a different V4L2 buffer ID. In this case I don't think
> the driver has a way to know that the memory resource should not be
> overwritten, and it will thus happily use it as a decode target. The
> easiest fix is probably to update the V4L2 stateful specification to
> require that reused DMABUFs must always be assigned to the same V4L2
> buffer ID, and must be kept alive as long as decoding is in progress,
> or until a resolution change event is received. We can then apply a
> similar rule to the virtio device.


Sounds like a generic kind of problem - how do other devices solve it?