Today's VirtIO-GPU driver always has to wait for in-fences on the guest side, creating bottleneck for workloads that produce lots of job submissions or require lower submission latency. This bottleneck is caused by a need to do extra round trips between host and guest for the fence signalling. The solution is to move fence signalling and waiting entierly to the host side, where it can be done more efficiently, without holding the guest side. The proposed patch introduces new "fence passing" VirtIO-GPU protocol feature, which extends VIRTIO_GPU_CMD_SUBMIT_3D with array of in-fence IDs that are passed from guest to host along with the 3D commands. Host then takes the in-fence IDs and passes them to virglrender, where the fence IDs are resolved into sync file fences and handled (waited) by a corresponding virglrender context. The patch is RFC because we want to do more testing with crosvm and CrOS first, support fence-passing for Venus and virtio-wl. The new virglrender experimental fence-passing API already has been merged [1] and crosvm support began to pick it up [2]. Qemu patch is available at [3]. Fence passing was tested using WIP amdgpu and intel native contexts, where it proved to be useful by increasing FPS up to 3x times. [1] https://gitlab.freedesktop.org/virgl/virglrenderer/-/merge_requests/1138 [2] https://chromium-review.googlesource.com/c/crosvm/crosvm/+/4679609 [3] https://gitlab.freedesktop.org/digetx/qemu/-/commits/native-context-iris Dmitry Osipenko (1): drm/virtio: Support fence-passing feature drivers/gpu/drm/virtio/virtgpu_drv.c | 1 + drivers/gpu/drm/virtio/virtgpu_drv.h | 11 ++- drivers/gpu/drm/virtio/virtgpu_fence.c | 15 +++- drivers/gpu/drm/virtio/virtgpu_ioctl.c | 11 ++- drivers/gpu/drm/virtio/virtgpu_kms.c | 8 +- drivers/gpu/drm/virtio/virtgpu_submit.c | 99 ++++++++++++++++++++++++- drivers/gpu/drm/virtio/virtgpu_vq.c | 7 +- include/uapi/drm/virtgpu_drm.h | 3 + include/uapi/linux/virtio_gpu.h | 11 ++- 9 files changed, 152 insertions(+), 14 deletions(-) -- 2.41.0
