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@xxxxxxxxxx> 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 _______________________________________________ Virtualization mailing list Virtualization@xxxxxxxxxxxxxxxxxxxxxxxxxx https://lists.linuxfoundation.org/mailman/listinfo/virtualization
