Linux Virtualization - Feb 2019 - [PATCH v2 5/6] drm/virtio: drop fencing in virtio_gpu_resource_create_ioctl

On Thu, Jan 31, 2019 at 11:04:31AM -0800, Gurchetan Singh
wrote:
> On Wed, Jan 30, 2019 at 1:43 AM Gerd Hoffmann <kraxel at redhat.com>
wrote:
> >
> > There is no need to wait for completion here.
> >
> > The host will process commands in submit order, so commands can
> > reference the new resource just fine even when queued up before
> > completion.
> 
> Does virtio_gpu_execbuffer_ioctl also wait for completion for a host
response?
No.

But you pass in a list of objects (drm_virtgpu_execbuffer->bo_handles)
used.  They will all get reserved, so you can use DRM_IOCTL_VIRTGPU_WAIT
on any of these objects to wait for completion.

Recently the driver got support for returning a fence fd for the
execbuffer, which you can also use to wait for completion in case your
kernel is new enough.
> From the guest driver perspective, a fence is just implemented has a
> virtio 3D resource.
> 
>
https://gitlab.freedesktop.org/mesa/mesa/blob/master/src/gallium/winsys/virgl/drm/virgl_drm_winsys.c#L787
> 
> The DRM_IOCTL_VIRTGPU_WAIT ioctl essentially waits for the reservation
> objects associated with that fence resource to become available.  So
> the flow is:
> 
> virtio_gpu_execbuffer_ioctl
> virtio_gpu_resource_create_ioctl with fence resource
> virtio_gpu_wait_ioctl with that fence resource --> associated with a
> GL wait on the host side
Oh.  /me looks surprised.
Wasn't aware that userspace is doing *that*.
> Does this change modify this sequence of events?
Yes.  DRM_IOCTL_VIRTGPU_WAIT will not wait any more.  Guess I have to
scratch the patch then ...

cheers,
  Gerd