Intel GVT (vGPU) and QEMU example

qemu.sh

#!/bin/bash -e

# https://www.kraxel.org/blog/2018/04/vgpu-display-support-finally-merged-upstream/

# You should set up your system as following:
# 1) Add 'i915.enable_gvt=1 intel_iommu=on iommu=pt' to boot arguments
# 1.1) If you get Windows guest working somehow, then add 'kvm.ignore_msrs=1' to avoid guest BSOD-ing
#       (Windows does weird stuff).
# 2) `MODULES=(kvm kvmgt i915)` in /etc/mkinitcpio.conf
# 3) `echo -e "options i915 enable_gvt=1\nsoftdep i915 pre: kvmgt" > /etc/modprobe.d/20-kvmgt.conf` for sure
# 4) Make sure that `ulimit -l` returns `unlimited` (probably needs less, but better safe than sorry)
#       Hint: Add 'yourusername  soft  memlock unlimited' and
#                 'yourusername  hard  memlock unlimited' to /etc/security/limits.conf 
# 5) Create new udev rule to access vfio as non-root:
#       echo 'SUBSYSTEM=="vfio", OWNER="root", GROUP="kvm"' > /etc/udev/rules.d/10-vfio.rules

# Edit this to match to your device
MDEV_DEVICE="/sys/class/mdev_bus/0000:00:02.0/mdev_supported_types/i915-GVTg_V4_4"

# Creates vGPU
UUID="$(uuidgen)"
echo "${UUID}" | sudo tee "${MDEV_DEVICE}/create"

# Notes:
# - This requires quite new QEMU, something like 2.12.0+ (it's in testing repos for Arch Linux at the time of writing)
# - Also this example uses OVMF, OS must support UEFI. You might want to save a copy of OVMF_VARS.fd somewhere writable
#       and remove `readonly` flag to configure EFI vars properly.
# - Windows 7 didn't like my 5th gen 5500 GT2 vGPU. Windows 10 complained about GPU having different ROM than boot time.
#       Setting rombar=0 helped getting it that far. Couldn't copy vBIOS either because Linux complained about failing
#       sanity check.
# - Kernel log gets spammed with gvt warnings of invalid addresses
#       It doesn't bother me (as I do 'cat /proc/kmsg | grep -v "gvt"'), but there is no real solution for that.
#       If you have specified '-video qxl', then don't open vfio-pci output before you've made sure that Intel vGPU is
#       initialized properly.
# - Host kernel threw bunch of gvt warnings and caused lag in VM when QEMU mouse emulation was used.
#       Binding USB mouse directly to VM helped to get rid of that issue.
# - If you are using NVIDIA Optimus like me (official NVIDIA way, iow disabling Intel, enabling modesetting and
#       giving whole display for discrete GPU to deal with), then you have to disable it somehow temporarily.
#       Otherwise you'll get no output from QEMU and/or crash Intel GPU completely (NVIDIA kernel module misbehaves and/or
#       NVIDIA OpenGL stack is missing few features on Linux).
#       Heck, NVIDIA OpenGL stack sucks anyway: https://i.imgur.com/smUCSLg.png (however it performs better
#       than nouveau so... that's why I'm stuck on it.)
#       I'm using https://github.com/mikroskeem/dot/blob/master/bin/bin/toggle-nvidia.sh for switching between GPUs without
#       reboots.
#       Probably doesn't work on non-Optimus setups either, I haven't confirmed it because of lacking hardware for testing.
qemu-system-x86_64 \
    -enable-kvm \
    -m 2G \
    -nodefaults \
    -M graphics=off \
    -monitor stdio \
    -soundhw ac97 \
    -cpu host \
    -drive if=pflash,format=raw,readonly,file=/usr/share/ovmf/x64/OVMF_CODE.fd \
    -drive if=pflash,format=raw,readonly,file=/usr/share/ovmf/x64/OVMF_VARS.fd \
    -smp cpus=2,cores=1,threads=1 \
    -netdev user,id=vmnet0,net=192.168.76.0/24,dhcpstart=192.168.76.9 \
    -display gtk,gl=on \
    -device nec-usb-xhci \
    -device e1000,netdev=vmnet0 \
    -device vfio-pci,sysfsdev=/sys/bus/mdev/devices/"${UUID}",rombar=0,display=on \
    -cdrom ~/Downloads/Fedora-Workstation-Live-x86_64-27-1.6.iso

#     -vga qxl \
#     -hda ~/win7.qcow2 \
#     -cdrom ~/.local/share/libvirt/images/W7.X86X64.OEM.ESD.M6.NOV2016.iso

# Deletes vGPU
echo 1 | sudo tee "${MDEV_DEVICE}/devices/${UUID}/remove"