Given a QEMU Windows 10 image (you can create one as you normally would, name it emugaming.qcow2), this will run it on Docker with GPU/vfio passthrough. Works great for near-native gaming.

Runs on Linux and with some changes should also be able to run on Windows or MacOS hosts (I think MacOS supports VT-d?). Note that you'll need a ROM dump of your GPU if you're passing through the primary boot video card. Also you'll need the kernel parameters iommu=pt intel_iommu=on to enable IOMMU/VT-d.

Still working on making stuff actually configurable, but make your modifications in start.sh for configuring your components. Ideally in the future you'll be able to just run docker run lg/emugaming with hardly any parameters and all will be automatic, including generating the Windows 10 image from MS.

DOCKER_BUILDKIT=1 docker build -t emugaming .

1. Create a disk image using qemu-img create -f qcow2 emugaming.qcow2 100G on the host. You might need the qemu-img package.
2. Run the docker run command below but appending ash at the end to enter the docker image. This will open ash, Alpine's bash equivalent.
3. Download the latest Windows 10 ISO from Microsoft.
4. Download the latest stable virtio-win driver pack from Fedora or their github.
5. Run a qemu command similar to start.sh, except:
   • I'd recommend not using the GPU passthrough just yet. Remove the vfio-pci, usb-host, vga and nographic lines, and instead pass through -vnc 0.0.0.0:0 -vga qxl -device usb-tablet and using VNC to install Windows and getting the image stable.
   • Note you need to mount the two ISOs you downloaded -cdrom win10.iso -cdrom virtio.iso
   • When you install Windows it won't see your hard drive, click Have Disk and select the viostor driver in the VirtIO driver ISO
   • You can use the quit or system_powerdown commands in qemu to do a hard shutdown or a soft one
6. Once Windows is installed, I'd recommend some things to change:
   • Disable hibernation and standby (its ok to keep monitor standby on though)
   • Set bcdedit /set '{current}' recoveryenabled No since it can get confusing when debugging video card passthrough problems
   • Install the remainder of the Virtio drivers, plus other missing drivers too (like video card)

docker run \
  --volume $PWD/emugaming.qcow2:/emugaming.qcow2 `# the persistent volume` \
  --interactive --tty \
  \
  --device /dev/kvm `# use hardware acceleration` \
  --device /dev/vfio/vfio ` # vfio is used for PCIe passthrough` \
  --device /dev/vfio/1 `# the vfio IOMMU group` \
  --ulimit memlock=-1:-1 `# so DMA can happen for the vfio passthrough` \
  --device /dev/bus/usb `# since we use usb-host device passthrough (note you can specify specific devices too)` \
  --volume /dev/bus/usb:/dev/bus/usb `# to allow for hot-plugging of USB devices` \
  --volume /lib/modules:/lib/modules `# needed for loading vfio` \
  --privileged `# needed for allowing hot-plugging of USB devices, but should be able to replace with cgroup stuff? also needed for modprobe commands` \
  emugaming

To use the QEMU Monitor command-line:

docker exec -it <CONTAINER_ID> socat - UNIX-CONNECT:/var/run/qemu_monitor

To start this docker image at boot:

docker run \
  ...
  --detach \
  --restart unless-stopped \
  ...

• Make GPU params configurable (like address of devices and ROM file)
• Make USB devices passed-through configurable
• Make CPU cores configurable (or autodetect max?)
• Detect and complain when IOMMU isnt on or devices not passthrough-able
• Generate Windows 10 image automatically based on official Microsoft images
• Remove the need for --privileged and broad permissions. Should be possible...
• Remove the sleep statement in start.sh
• Fix windows launching recover mode
• Get this built for k3s or another Kubernetes distro