Xe Native Context Enablement for virtio-gpu

Patches to enable Intel Xe GPU acceleration in QEMU/KVM virtual machines via drm_native_context. This allows OpenGL, Vulkan, and VA-API (hardware video decode/encode) to work inside a guest VM without full GPU passthrough.

What This Is

The Linux virtio-gpu driver supports a mechanism called native context (drm_native_context) that allows a guest VM to submit GPU commands directly to the host kernel driver, with the host-side virglrenderer acting as a thin relay rather than doing full GPU command translation. This is already supported for i915 (legacy Intel) and other drivers.

These patches extend native context support to the Intel Xe kernel driver, which is the modern replacement for i915 on recent Intel GPUs.

Status

Functional and tested. OpenGL (via iris/Gallium), Vulkan (via ANV), and VA-API (hardware video acceleration) are all working.

• Tested hardware:
  • Intel Arc B390 (iGPU on Core X7 358H / Panther Lake) -- original development platform
  • 13th Gen Intel Core i7-13700H with Intel Iris Xe Graphics -- confirmed working unmodified (test report, Unigine Heaven: 2928 native-context vs 1557 virgl translation)
• Testers welcome for other Xe-based Intel GPUs (Arc A-series, Meteor Lake, Arrow Lake, Battlemage, etc.) -- please open an issue with your results.

Performance varies by workload and rendering path. Some paths work very well; others may show overhead compared to bare metal. This is experimental.

Quick start

If you'd rather not build from source:

• Arch / CachyOS: prebuilt x86-64-v3 pacman repositories are published at cmspam/xe-virt-repo. Add [xe-virt-host-v3] on the host for patched virglrenderer, and [xe-virt-guest-v3] in the guest for patched mesa, lib32-mesa, and intel-media-driver. The repos rebuild automatically as new Arch packaging tags land. See that repo's README for the pacman.conf snippets.
• NixOS: issue #1 contains a self-contained flake.nix from an external tester that overlays all three patches, builds a runnable guest VM, and exposes runVM-native and runVM-virgl apps for side-by-side comparison.

Upstreaming (help wanted)

These patches are not yet upstream in virglrenderer or Mesa, and I'd like them to be. I've opened virglrenderer issue #662 asking for someone with the time and experience submitting changes to virglrenderer / Mesa to lead the upstreaming effort, but nobody has picked it up yet.

I'm not a regular contributor to these projects and currently lack the bandwidth to navigate the full Merge Request and revision process for a component this critical. I also relied on AI assistance to fill some of my knowledge gaps during development; the result is functional and I kept it as close to the i915 implementation as I could, but it would benefit from review by someone more familiar with the internal architecture of virglrenderer / Mesa.

If you have the experience and time to take this on, please consider this an open invitation. I'm happy to answer questions about the design decisions, but I'd be thrilled if an experienced contributor felt comfortable leading the upstreaming. New Intel hardware no longer supports i915, so getting Xe drm_native_context upstream matters for everyone running virtio-gpu going forward.

Patches

Patch	Apply To	Where
virglrenderer-xe-native-context.patch	virglrenderer	Host
mesa-01-xe-native-context-plus-iris-upload-fix.patch	Mesa	Guest
intel-media-driver-xe-native-context.patch	intel-media-driver	Guest (only needed for VA-API)

virglrenderer (host)

Adds an Xe renderer backend to virglrenderer. This is the host-side component that receives Xe DRM commands from the guest over the virtio-gpu transport and replays them against the real Xe kernel driver. Build with -Ddrm-renderers=xe-experimental (or add xe-experimental to your existing list).

Mesa (guest)

Adds Xe virtio ccmd support to Mesa's shared Intel device/virtio layer. This plumbing is used by both the iris Gallium driver (OpenGL) and ANV (Vulkan) to issue Xe ioctls over the vdrm protocol instead of talking to a real kernel driver. Also includes a targeted fix for upload buffer caching on virtio Xe (forces WC for staging BOs that exhibit CPU-write stalls under WB caching).

The Mesa patch must be applied on guest (where the actual GPU drivers run) but it's also important to have a recent version of Mesa on the host.

intel-media-driver (guest, optional)

A self-contained vdrm client shim for Intel's iHD media driver. This transparently intercepts Xe DRM ioctls from the media driver and routes them through the virtio-gpu vdrm protocol. Only needed if you want hardware video decode/encode (VA-API) in the guest. Tested and working with ffmpeg, vlc, brave browser, and others.

How to Apply

These patches were developed against upstream git as of early April 2026. They may need minor rebasing if the upstream projects have changed significantly since then.

# virglrenderer (host)
cd virglrenderer
git apply virglrenderer-xe-native-context.patch

# Mesa (guest and host)
cd mesa
git apply mesa-01-xe-native-context-plus-iris-upload-fix.patch

# intel-media-driver (guest, only for VA-API)
cd media-driver
git apply intel-media-driver-xe-native-context.patch

QEMU Configuration

You need a recent QEMU build from git (the required honor-guest-pat and drm_native_context options are not available in older releases).

Required QEMU flags:

-accel kvm,honor-guest-pat=on
-device virtio-vga-gl,blob=on,hostmem=4G,drm_native_context=on

Key requirements:

• drm_native_context=on -- enables the native context path (without this, you get software rendering or virgl translation, not Xe passthrough)
• blob=on -- enables blob resources for zero-copy buffer sharing
• hostmem=4G -- host memory window size for blob resources (adjust as needed)
• honor-guest-pat=on -- required for correct guest memory caching behavior on Intel

The host must be running the patched virglrenderer. The guest kernel must have virtio-gpu and drm support enabled.

Building

virglrenderer (host)

meson setup build -Ddrm-renderers=xe-experimental
ninja -C build
sudo ninja -C build install

If you also use i915 native context, combine them: -Ddrm-renderers=i915-experimental,xe-experimental.

Mesa (guest and host)

Build Mesa as normal for your distribution, but with -D intel-virtio-experimental=true enabled (this is also needed for i915).

intel-media-driver (guest)

Build as normal per the upstream instructions. The virtio shim is compiled automatically when the patched source is built. In the guest, set:

export LIBVA_DRIVER_NAME=iHD

Architecture Overview

Guest VM                              Host
+--------------------------+          +---------------------------+
| Application              |          |                           |
|   |                      |          |                           |
| iris/ANV (Mesa)          |          |                           |
|   |                      |          |                           |
| xe virtio ccmd layer     |  virtio  | virglrenderer             |
|   (vdrm protocol)       <----------->  xe renderer backend     |
|   |                      |   gpu    |   |                       |
| virtio-gpu kernel driver |          | Xe kernel driver (real)   |
+--------------------------+          +---------------------------+
                                              |
                                         Intel Xe GPU

Constraints/Limitations

• The banned_exec_queue_mask in shared memory tracks up to 64 exec queues (one bit per queue in a uint64_t). This is sufficient for normal usage.
• VM binds are synchronous on the host side (the host waits for each bind to complete before returning to the guest). This is by design to guarantee mapping visibility before dependent submissions.
• The intel-media-driver shim uses a single-instance global vdrm connection, matching iHD's one-device-per-process model.

License

The modified files in these patches are MIT/X11 licensed (see SPDX headers in each patched file).