UtkarshVerma / qemu-vfio-win10

Shell script for launching a Windows 10 guest with GPU passthrough using Looking Glass.

qemu-vfio-win10

This repo aims to serve as an example for getting VFIO passthrough working on a Windows 10 guest on Linux. Plain QEMU has been used to keep things minimal, and distro-specific utilities have been avoided to keep things general.

Features

The script has the following features:

• Hugepages
• HyperV enlightenments
• Dynamic VFIO passthrough for NVIDIA GPU, i.e. GPU is usable by host after guest exits.
• Respects NVIDIA Runtime D3 power management
• evdev passthrough using persistent-evdev.py
• Looking Glass
  • SPICE input and clipboard sharing
  • DMA buffer support
  • Auto shared memory creation
• virtio devices
• pulseaudio audio device passthrough
• CPU pinning with qemu-affinity
• CPU governor configuration
• Runs QEMU with jemalloc() memory allocator

Usage

This script assumes you already have a working Windows 10 guest set up for Looking Glass and GPU passthrough. If you are new to VFIO passthrough, then these links should get you started:

• PCI passthrough via OVMF - Arch Wiki
• GPU passthrough with libvirt qemu kvm
• Looking Glass Docs

Once you have a working Windows 10 guest, move the disk image file to this folder and rename it to hdd.qcow2. After that copy the OVMF file descriptors OVMF_CODE.fd and OVMF_VARS.fd next to the script as well so that you get the following tree:

.
├── LICENSE
├── OVMF_CODE.fd
├── OVMF_VARS.fd
├── README.md
├── SSDT1.dat
├── hdd.qcow2
├── launch.sh
├── persistent-evdev.json
├── persistent-evdev.py
├── qemu-affinity
└── vfio

0 directories, 11 files

You'll also need to configure a few variables in the script according to your machine.

USER=subaru                     # Your username
RAM=5                           # RAM, in GB
CORES=3                         # No. of cores
THREADS=2                       # No. of threads per core
GPU=01:00.0                     # PCI address of your NVIDIA GPU
GPU_AUDIO=01:00.1               # PCI address of your NVIDIA GPU's audio function
SHARED_FOLDER=/media/stuff      # Folder to be shared to guest

Once that is out of the way, make sure you have jemalloc installed on your machine, and persistent-evdev.py configured through persistent-evdev.json if you plan on using evdev passthrough. Also make sure you have installed the qemu-affinity python package.

pip install qemu-affinity

Once that's done, run the launch script as root, and you're good to go.

sudo ./launch.sh

Running this command starts the guest and opens Looking Glass, with the QEMU monitor socket opened on /tmp/qemuwin.sock.

The script supports some command line flags:

Flags	Usage
-s / --spice	Use SPICE for input instead of evdev. Defaults to evdev.
-h / --hugepages	Use hugepages. Disabled by default.
-f / --fullscreen	Open the guest in fullscreen. Disabled by default.
-nd / --nodmabuf	Disable DMABUF for Looking Glass, which is enabled by default.

What to expect from the guest

The guest performs really well for a VM. I don't experience any freezes. For normal usage and light gaming (Brawlhalla), I couldn't notice any issues with this setup on my laptop (ASUS TUF FX505DT).

Credits

This workflow would not have been possible without the following tools.

• QEMU
• Looking Glass
• jemalloc
• persistent-evdev.py
• qemu-affinity