vhost-device/vhost-device-gpu

vhost-device-gpu - GPU emulation backend daemon

Synopsis

vhost-device-gpu --socket-path <SOCKET> --gpu-mode <GPU_MODE>

Description

A virtio-gpu device using the vhost-user protocol.

Options

  -s, --socket-path <SOCKET>
          vhost-user Unix domain socket

  -g, --gpu-mode <GPU_MODE>
          The mode specifies which backend implementation to use
          
          [possible values: virglrenderer, gfxstream]

  -c, --capset <CAPSET>
          Comma separated list of enabled capsets

          Possible values:
          - virgl:            [virglrenderer] OpenGL implementation, superseded by Virgl2
          - virgl2:           [virglrenderer] OpenGL implementation
          - gfxstream-vulkan: [gfxstream] Vulkan implementation (partial support only)
             NOTE: Can only be used for 2D display output for now, there is no hardware acceleration yet
          - gfxstream-gles:   [gfxstream] OpenGL ES implementation (partial support only)
             NOTE: Can only be used for 2D display output for now, there is no hardware acceleration yet

      --use-egl <USE_EGL>
          Enable backend to use EGL
          
          [default: true]
          [possible values: true, false]

      --use-glx <USE_GLX>
          Enable backend to use GLX
          
          [default: false]
          [possible values: true, false]

      --use-gles <USE_GLES>
          Enable backend to use GLES
          
          [default: true]
          [possible values: true, false]

      --use-surfaceless <USE_SURFACELESS>
          Enable surfaceless backend option
          
          [default: true]
          [possible values: true, false]

  -h, --help
          Print help (see a summary with '-h')

  -V, --version
          Print version

NOTE: Option -g, --gpu-mode can only accept the gfxstream value if the crate has been built with the gfxstream feature, which is the default.

Limitations

This device links native libraries (because of the usage of Rutabaga) compiled with GNU libc, so the CI is setup to not build this device for musl targets. It might be possible to build those libraries using musl and then build the gpu device, but this is not tested.

We are currently only supporting sharing the display output to QEMU through a socket using the transfer_read operation triggered by VIRTIO_GPU_CMD_TRANSFER_FROM_HOST_3D to transfer data from and to virtio-gpu 3D resources. It'll be nice to have support for directly sharing display output resource using dmabuf.

This device does not yet support the VIRTIO_GPU_CMD_RESOURCE_CREATE_BLOB, VIRTIO_GPU_CMD_SET_SCANOUT_BLOB and VIRTIO_GPU_CMD_RESOURCE_ASSIGN_UUID features. This requires https://github.com/rust-vmm/vhost/pull/251, which in turn requires QEMU API stabilization. Because blob resources are not yet supported, some capsets are limited:

• Venus (Vulkan implementation in virglrenderer project) support is not available at all.
• gfxstream-vulkan and gfxstream-gles support are exposed, but can practically only be used for display output, there is no hardware acceleration yet.

Features

The device leverages the rutabaga_gfx crate to provide rendering with virglrenderer and gfxstream.

This crate supports two GPU backends: gfxstream (default) and virglrenderer. Both require the system-provided virglrenderer and minigbm libraries due to the dependence on rutabaga_gfx.

Install the development packages for your distro, then build with:

CROSVM_USE_SYSTEM_VIRGLRENDERER=1 \
CROSVM_USE_SYSTEM_MINIGBM=1 \
cargo build

gfxstream support is compiled by default, it can be disabled by not building with the gfxstream feature flag, for example:

CROSVM_USE_SYSTEM_VIRGLRENDERER=1 \
CROSVM_USE_SYSTEM_MINIGBM=1 \
cargo build --no-default-features

With Virglrenderer, Rutabaga translates OpenGL API and Vulkan calls to an intermediate representation and allows for OpenGL acceleration on the host.

With the gfxstream rendering mode, GLES and Vulkan calls are forwarded to the host with minimal modification.

Examples

First start the daemon on the host machine using either of the 2 gpu modes:

1. virglrenderer
2. gfxstream (if the crate has been compiled with the feature gfxstream)

host# vhost-device-gpu --socket-path /tmp/gpu.socket --gpu-mode virglrenderer

With QEMU, there are two device front-ends you can use with this device. You can either use vhost-user-gpu-pci or vhost-user-vga, which also implements VGA, that allows you to see boot messages before the guest initializes the GPU. You can also use different display outputs (for example gtk or dbus). By default, QEMU also adds another VGA output, use -vga none to make sure it is disabled.

1. Using vhost-user-gpu-pci

Start QEMU with the following flags:

-chardev socket,id=vgpu,path=/tmp/gpu.socket \
-device vhost-user-gpu-pci,chardev=vgpu,id=vgpu \
-object memory-backend-memfd,share=on,id=mem0,size=4G, \
-machine q35,memory-backend=mem0,accel=kvm \
-display gtk,gl=on,show-cursor=on \
-vga none

2. Using vhost-user-vga

Start QEMU with the following flags:

-chardev socket,id=vgpu,path=/tmp/gpu.socket \
-device vhost-user-vga,chardev=vgpu,id=vgpu \
-object memory-backend-memfd,share=on,id=mem0,size=4G, \
-machine q35,memory-backend=mem0,accel=kvm \
-display gtk,gl=on,show-cursor=on \
-vga none

License

This project is licensed under either of

• Apache License, Version 2.0
• BSD-3-Clause License