A couple of patches are required for the kernel and qemu on the host side. You should use the latest version of mesa (host and guest) and virglrenderer libraries (host) from the upstream main branches.

Series

What

https://patchew.org/QEMU/20240623152343.328436-1-dmitry.osipenko@collabora.com/

QEMU support for blob memory and Venus

https://lore.kernel.org/lkml/20240726235234.228822-1-seanjc@google.com/

KVM support to avoid pinning device memory

Dependencies

Host

You will also need to build the latest mesa and virglrenderer with support. The graphics stacks loads a lot of libraries dynamically so it’s important to make sure you set all the environment variables so things point at the right place. Specifically:

Env Variable

Purpose

PKG_CONFIG_PATH

Configure can find the newer libs

$INSTALL/lib/pkconfig and $INSTALL/lib/$TRIPLE/pkgconfig

LD_LIBRARY_PATH

So dlopen() searchs the right place

$INSTALL/lib/

$INSTALL/lib/$TRIPLE

$INSTALL/lib/$TRIPLE/dri

VK_ICD_FILENAMES

So the latest Vulkan driver is loaded

$INSTALL/share/vulkan/icd.d/$GPU.$ARCH.json

Guest Dependencies

Debian Sid now has the latest Mesa bits (24.2) which should get promoted to Testing (Trixie) in due couse. However if you do want to install a newer Mesa you will want to ensure you boot up in console mode and then only start the GUI making sure you are using the newer Mesa and not the system version.

There is a test kernel+initramfs here: https://fileserver.linaro.org/s/ce5jXBFinPxtEdx

Build virglrenderer

$ sudo apt install libdrm-dev libepoxy-dev pkg-config cmake meson libvulkan-dev libva-dev
$ meson setup build \
  -Dvenus=true \
  -Drender-server=true \
  -Drender-server-worker=thread \
  -Dbuildtype=release \
  -Dprefix=${INSTDIR}
$ ninja -C build install

Build mesa

LLVM is required to build mesa libaray. If llvm doesn't enable cpp_rtti, add the following line to meson.build.

-rtti = false

and add "-cpp_rtti=false" to the following meson setup.

$ sudo apt install \
yacc \
bison \
cmake \
llvm-15 \
glslang-tools \
libglvnd-dev \
libwayland-egl-backend-dev \
libxext-dev \
libxfixes-dev \
libxcb-glx0-dev \
libxcb-shm0-dev \
libx11-xcb-dev \
libxcb-dri2-0-dev \
libxcb-dri3-dev \
libxcb-present-dev \
libxshmfence-dev \
libxxf86vm-dev \
libxrandr-dev \
libudev-dev \
libelf-dev \
libdrm-dev
$ meson setup build \
  -Dgallium-nine=false -Dgallium-xa=disabled -Dglx=dri \
  -Dshared-glapi=enabled -Ddri3=enabled -Degl=enabled -Dgbm=enabled \
  -Dglvnd=true -Dgallium-drivers=radeonsi,virgl -Dplatforms=x11,wayland \
  -Dgles1=disabled -Dgles2=enabled -Dvulkan-drivers=virtio,amd \
  -Dlibunwind=disabled -Dbuildtype=release \
  -Dprefix=${INSTDIR}
$ ninja -C build install

For guest, use the following command to configure

(TBD)

Build qemu

Download the source code[1] from Collabora or the patch[2] from AMD should also work. Replacee “x86_64-linux-gnu” with “aarch64-linux-gnu” on AVA.

$ sudo apt install libglib2.0-dev \
libpixman-1-dev \
libaio-dev (optional) \
liburing-dev (optional) \
libslirp-dev \
libudev-dev (optional) \
libsdl2-dev \
libgtk-3-dev
$ PKG_CONFIG_PATH=${INSTDIR}/lib/x86_64-linux-gnu/pkgconfig
$ mkdir build
$ cd build
$ ../configure \
  --prefix=${INSTDIR} \
  --extra-cflags="-I${INSTDIR}/include " \
  --extra-ldflags="-L${INSTDIR}/lib/x86_64-linux-gnu" \
  --target-list=x86_64-softmmu \
  --enable-kvm \
  --disable-werror \
  --enable-slirp \
  --enable-opengl \
  --enable-virglrenderer \
  --enable-gtk \
  --enable-sdl
$ ninja install

[1] https://gitlab.freedesktop.org/digetx/qemu.git branch:venus-v4

[2] https://patchew.org/QEMU/20230915111130.24064-1-ray.huang@amd.com/

Build kernel

As of v6.10.5 the kernel should have everything you need for the guest (baring any host specific hacks e.g. for AVA memory consistancy). The PFNMAP patches should only be needed on the host and only when you have a GPU with on-card memory.

You will need the following options enabled:

Kernel Flag

Reason

CONFIG_DRM_VIRTIO_GPU

Enable VirtIO GPU support

CONFIG_UDMABUF

Userspace access to DMA buffers

Run the guest on x86

$ sudo usermod -a -G kvm ${YOURID} # for /dev/udmabuf
$ export VK_DRIVER_FILES=${INSTDIR}/share/vulkan/icd.d/radeon_icd.x86_64.json
$ {INSTDIR}/bin/qemu-system-x86_64 \
  -enable-kvm \
  -smp 1 \
  -m 4G \
  -hda ${ROOTFS_IMG} \
  -serial mon:stdio \
  -device virtio-vga-gl,iommu_platform=on,hostmem=4G,blob=true,context_init=true \   
  -display gtk,gl=on \
  -object memory-backend-memfd,id=mem1,size=4G,share=on \
  -machine q35,accel=kvm,kernel-irqchip=split,memory-backend=mem1 \
  -device intel-iommu,intremap=on \
  -device virtio-balloon \
  -net nic,model=virtio -net user,hostfwd=tcp::2222-:22 \
  -d guest_errors \
  -boot d

(Note for Intel platform only)

If you see a problem similar to [4] when you run vkmark or vulkan applications, apply the following tweak to the host kernel.

diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
index dd92361f41b3..f0a9ae22f691 100644
--- a/arch/x86/kvm/vmx/vmx.c
+++ b/arch/x86/kvm/vmx/vmx.c
@@ -7556,7 +7556,7 @@ static u8 vmx_get_mt_mask(struct kvm_vcpu *vcpu, gfn_t gfn, bool is_mmio)
                return MTRR_TYPE_UNCACHABLE << VMX_EPT_MT_EPTE_SHIFT;

        if (!kvm_arch_has_noncoherent_dma(vcpu->kvm))
-               return (MTRR_TYPE_WRBACK << VMX_EPT_MT_EPTE_SHIFT) | VMX_EPT_IPAT_BIT;
+               return (MTRR_TYPE_WRBACK << VMX_EPT_MT_EPTE_SHIFT);

        if (kvm_read_cr0(vcpu) & X86_CR0_CD) {
                if (kvm_check_has_quirk(vcpu->kvm, KVM_X86_QUIRK_CD_NW_CLEARED))

[4] https://drive.google.com/file/d/1gaF2FCv8lmD_tsVDbrdvGgaxCK-RdDkX/view?usp=sharing

Run the guest on AVA

Host and Guest Kernel

This kernel tree contains the PFN patches and a few local hacks for AVA hardware weirdness: https://git.linaro.org/people/alex.bennee/linux.git/log/?h=review/pfn-references-v12-with-altra-tweaks