komorebi/docs/linux/dev/utm.md

Plasma guests to develop in

The VM tests boot a machine, run a scenario and power it off. To try komorebi by hand against a real KWin, the flake has NixOS guests with Plasma 6 that stay up. Each guest is made for one kind of host, because the accelerator of a host (HVF on a Mac, KVM on Linux) runs guests of the host's own architecture:

Guest	Host	Runs under	What it is
plasma-vm	Apple silicon Mac	QEMU from nixpkgs, with HVF	Plasma installed, no login: a console and ssh, to run kwin_wayland --virtual and build komorebi from the repository
plasma-vm-session	Apple silicon Mac	QEMU from nixpkgs, with HVF	the same, plus a Plasma Wayland session that logs in by itself, komorebi running in it, and an RDP server
plasma-utm	Apple silicon Mac	UTM, with a GPU	the session guest as a disk image that boots by UEFI
plasma-vm-x86_64, plasma-vm-session-x86_64	Linux on x86_64	QEMU from nixpkgs, with KVM	the first two, as x86_64-linux guests
plasma-vm-aarch64, plasma-vm-session-aarch64	Linux on aarch64	QEMU from nixpkgs, with KVM	the first two, as aarch64-linux guests
run-kde-x86_64, run-kde-aarch64	Linux on x86_64, on aarch64	QEMU from nixpkgs, with KVM	the session guest with a demo on it, which run-kde starts (below)

The guests of a Mac are aarch64-linux. A guest of a Linux host that has no KVM does not start: the runner refuses, because Plasma under software emulation is too slow to try anything.

They are nixosConfigurations of the flake, made from nix/linux/dev/plasma-vm.nix, nix/linux/dev/plasma-utm.nix and nix/linux/dev/krdp.nix, by mkPlasmaGuest in flake.nix, which takes the system of the guest and the packages of the machine that runs QEMU. Each one installs komorebi with the flake's NixOS module, services.komorebi, so the komorebi in a guest is the one of the commit the guest was built from, and in the session guests the module's unit starts it at login. They carry no key, no password and no configuration of anyone's.

In the QEMU guests the guest has a plain virtio-gpu, which has no 3D, so KWin composites on the CPU (llvmpipe) and every frame then has to be encoded to reach an RDP client. That is slow. On a Mac the QEMU in nixpkgs cannot do better: it has no 3D at all. UTM's QEMU has virtio-gpu-gl, which gives the guest OpenGL through Mesa's VirGL, rendered on the Mac's GPU, and UTM shows the guest in a window of its own, so there is no RDP in the path at all. The QEMU of an x86_64 Linux host has and a window (GTK), at time of writing (QEMU 11.1.2 from nixpkgs, 2026-10-28), which these guests do not use.

What a builder has to be, on a Mac

A Mac cannot build a Linux system, so the guests are built on a builder: a machine with Nix that builds for aarch64-linux and that the Mac can reach as a remote store, such as ssh-ng://builder@localhost. Any builder will do: the one nix run nixpkgs#darwin.linux-builder starts, a Linux machine on the network, or the one the VM tests use on a Mac (vm-tests.md). A Plasma system is a closure of about 10 GB, and the builder needs room for it and for its build: the defaults of darwin.linux-builder (one core, 3 GB of memory, a 20 GB disk) are too small. Six cores and 10 GB of memory worked, at time of writing (2026-10-27).

The examples below use the same variables as scripts/linux/vm-mac:

export KOMOREBI_BUILDER=ssh-ng://builder@localhost   # the store URL of your builder
export NIX_SSHOPTS="-p 31022 -i $HOME/.ssh/builder_ed25519 -o IdentitiesOnly=yes"   # if it needs them

The QEMU guests on a Mac

The system is built on the builder and copied back, and only the QEMU runner, which is a program for the Mac, is built on the Mac. A QEMU guest runs the system of its vmVariant (which adds the disk and the shared directory), not the system of the configuration itself:

guest=plasma-vm-session   # or plasma-vm
config=".#nixosConfigurations.$guest.config"
system=$(nix build "$config.virtualisation.vmVariant.system.build.toplevel" \
  --eval-store auto --store "$KOMOREBI_BUILDER" --no-link --print-out-paths)
nix copy --from "$KOMOREBI_BUILDER" --substitute-on-destination --no-check-sigs "$system"
nix build "$config.system.build.vm" --builders '' -o .vm/result-$guest

The guest mounts the directory KOMOREBI_REPO names at /repo, and makes its disk, plasma-vm.qcow2 or plasma-vm-session.qcow2, in the directory it is started from on its first start, and keeps it there. .vm is ignored by git:

cd .vm && KOMOREBI_REPO=$PWD/.. ./result-$guest/bin/run-$guest-vm

The terminal is the guest's console, logged in as the user komorebi, who can use sudo with no password. ssh is forwarded from localhost:2222 (plasma-vm) or localhost:2223 (plasma-vm-session) and takes keys only: put yours in ~/.ssh/authorized_keys from the console once, and the disk keeps it. Both guests can run at once.

In the guest, /repo is the repository and CARGO_TARGET_DIR is in the guest's home, since a target directory on the shared directory is slow. The session guest shows nothing on the Mac; it is reached over RDP (below).

The QEMU guests on a Linux machine

A Linux host builds a guest itself, so there is no builder to set up. It needs Nix with flakes, KVM (a /dev/kvm that the user can open, which on most distributions means the group kvm), and room for the closure (about 10 GB) and for the guest's disk (20 GB, sparse). The runner is built with its system, and started the same way as on a Mac:

guest=plasma-vm-session-x86_64   # or plasma-vm-x86_64, or the -aarch64 pair on an aarch64 host
nix build ".#nixosConfigurations.$guest.config.system.build.vm" -o .vm/result-$guest
cd .vm && KOMOREBI_REPO=$PWD/.. ./result-$guest/bin/run-*-vm

The console, the user, the ssh ports (2222 and 2223), the RDP port, the disk and /repo are as described for a Mac. What differs is what nixpkgs' runner does on a Linux host:

• It starts one virtiofsd for each shared directory, as the user who runs the guest, and the guest mounts them by virtiofs where a Mac's guest mounts them by 9p. One of them is the host's /nix/store, so the guest has the closure without a copy of it.
• It refuses to start when /dev/kvm is missing or cannot be opened by the user, since the guests set virtualisation.qemu.forceAccel.

A guest has one display device, a virtio-gpu, so that KWin has a DRM device to draw on. The machine of an x86_64 guest has a VGA of its own, which would be a second display next to the virtio-gpu, so the guest has virtio-vga (a virtio-gpu that is also a VGA) in its place and -vga none. The session guest shows nothing on the host, as on a Mac: it is reached over RDP (below).

The demo guest

run-kde-x86_64 and run-kde-aarch64 are the session guest with a demo on it, made by nix/linux/dev/run-kde.nix on top of plasma-vm.nix. komorebi is configured by the module's typed config (eight workspaces with a layout each, borders and animation) and by the sample key bindings, docs/skhdrc.sample, in the module's default mode. At login the guest opens two Konsole windows, Kate and Dolphin, which komorebi tiles, and a cheat sheet that it leaves alone: the main key bindings, komorebic commands to try, what KWin draws with, and the output of komorebic check for the configuration the daemon runs. The VM suite run-kde boots the same module on the session node.

The closure is the same for everyone. No password, key or configuration of a person is in it, and what a person's run needs comes from the environment of the runner:

Variable	What the guest does with it
KOMOREBI_REPO	the directory mounted at /repo
RUN_KDE_STATE	the directory mounted at /run/run-kde, with rdp-password (the password for RDP) and ssh_key.pub (the public key that may log in over ssh)
RUN_KDE_MEMORY	the memory in MiB, which has to be the -m of the same run
QEMU_NET_OPTS	the port forwards, as hostfwd=tcp:127.0.0.1:2223-:22,hostfwd=tcp:127.0.0.1:3389-:3389
QEMU_OPTS	the display device: -vga none -device virtio-vga,xres=1280,yres=800, or -display gtk,gl=on -vga none -device virtio-vga-gl,xres=1280,yres=800 for a GPU
NIX_DISK_IMAGE	the disk image, which the runner makes when it is not there

run-kde sets them (the guide says what a person sees and does); to start the guest by hand, on a Linux machine:

nix build .#nixosConfigurations.run-kde-x86_64.config.system.build.vm -o .vm/result-run-kde
mkdir -m 700 .vm/state && cd .vm
ssh-keygen -q -t ed25519 -N '' -f state/ssh_key
printf %s "a password" > state/rdp-password
KOMOREBI_REPO=$PWD/.. RUN_KDE_STATE=$PWD/state RUN_KDE_MEMORY=6144 \
  QEMU_NET_OPTS=hostfwd=tcp:127.0.0.1:2223-:22,hostfwd=tcp:127.0.0.1:3389-:3389 \
  QEMU_OPTS='-cpu host -vga none -device virtio-vga,xres=1280,yres=800' \
  ./result-run-kde/bin/run-komorebi-demo-vm -m 6144 -smp 4

The guest differs from plasma-vm-session in what the runner of nixpkgs would otherwise fix when the closure is built:

• The display device is not part of the guest (komorebi.plasmaVm.display is false). The same closure and the same disk run with the software virtio-vga or with virtio-vga-gl, which renders on the host's GPU through VirGL, as the runner's environment says.
• The ports are not forwarded by the guest, so that a person can choose them.
• The runner does not insist on KVM (qemu.forceAccel is false): run-kde checks for it and lets --tcg start a guest without.
• The shared memory that virtiofs needs is a memory backend of the guest's memory size, which nixpkgs fixes at memorySize in the command line of the runner. The guest makes its own, from RUN_KDE_MEMORY, so that a run can have less memory than the closure's 8 GB.
• The directory /run/run-kde is mounted after the boot and not in the initrd, where a mount under /run is hidden by the /run that the booted system mounts.
• The files in it can belong to a user of the host who is not the guest's user, and the guest cannot read the directory then. The unit run-kde-credentials copies the two files as root, to /run/run-kde-guest, where the RDP unit (komorebi.krdp.passwordFile) and sshd (AuthorizedKeysFile) read them.

The command

run-kde starts the demo guest on a Linux machine, from a checkout: nix run .#run-kde -- [options], or just run-kde [options]. It is scripts/linux/run-kde, which the flake wraps with the runner of the machine's guest (run-kde-x86_64 or run-kde-aarch64) written into it, as packages.<system>.run-kde and apps.<system>.run-kde. The package names the runner and does not depend on it, so nix run does not fetch 4 GB before the command can say what the machine lacks.

In order, it:

1. Checks the machine. --doctor prints every check, a tick or a cross or a !, with the fix under each cross and each !: WSL 2 or not, Nix, KVM (/dev/kvm, the group kvm, the CPU's svm or vmx flag, the module kvm_amd or kvm_intel, and on WSL 2 the line nestedVirtualization=true of .wslconfig), virtiofs (a virtiofsd, the runner's own when it is built, is started as this user with the runner's options and must make its socket: a cross when it does not; /dev/fuse is a !, since virtiofsd never opens it), memory, cores, disk, a window (WSLg: WAYLAND_DISPLAY, DISPLAY), Windows interop for --rdp, the GPU (/dev/dxg on WSL 2, a render node elsewhere, and the OpenGL renderer of the host from glxinfo), and the networking mode of .wslconfig. A cross stops the command and --doctor exits with 1; a ! does not. A missing KVM is a cross, unless --tcg allows software emulation, which is too slow to try anything.
2. Builds the guest when its runner is not in the store, with nix build of the checkout's flake and the build log shown. There is no binary cache: the first run builds komorebi and downloads about 4 GB of Plasma, which are 11 GB in the Nix store. The first run says how long that took where it was measured.
3. Makes the state directory, $XDG_STATE_HOME/komorebi-run-kde (~/.local/state/komorebi-run-kde), when it is not there: the guest's disk (20 GB, or --size, sparse), a random RDP password and an ed25519 key for ssh, all of mode 0600 in a directory of mode 0700. --reset deletes it. Nothing of it is in the Nix store.
4. Prints a banner and starts the runner with the environment of the demo guest: -cpu host with KVM, the repository at /repo (the checkout, or KOMOREBI_REPO), the state directory at /run/run-kde, the ports on 127.0.0.1 (ssh 2223, RDP 3389: --ssh-port, --rdp-port), and -m and -smp from what the machine has (--memory, --cores): the machine's memory less 2048 MB, between 4096 and 8192, and its cores up to 6.

The terminal is the guest's console, logged in as komorebi. What else the guest shows depends on the display and on the GPU.

--display	What it does
gtk	a window of QEMU; the default when WAYLAND_DISPLAY or DISPLAY is set
rdp	no window; the guest is reached over RDP; the default otherwise
vnc	no window; QEMU's VNC server on 127.0.0.1:5900
none	no window; --shell is this with the GPU off

--gpu auto uses the GPU when the display is a window, the host has a GPU device and the OpenGL of the host (glxinfo) is not llvmpipe. Otherwise it does not, and the banner says why. The GPU is -display gtk,gl=on with -device virtio-vga-gl, which renders the guest's OpenGL on the host through VirGL; without it the guest has virtio-vga and renders on its CPUs. --gpu on with no window starts -display egl-headless, which needs a DRM render node. The guest says what KWin draws with in its cheat sheet and on the console (GPU (virgl) or software (llvmpipe)): the guest's driver, not the host's.

On WSL 2, --rdp starts Windows' own RDP client once ssh says that the unit krdp is running: cmdkey.exe stores the password for TERMSRV/localhost, mstsc.exe connects to localhost:3389, and the credential is deleted when run-kde ends.

--dry-run prints the environment and the command that would be started, and starts nothing. The script reads the machine through variables, so that a test can make up a machine: RUN_KDE_KVM_DEVICE, RUN_KDE_DXG_DEVICE, RUN_KDE_DRI_DIR, RUN_KDE_MEMINFO, RUN_KDE_CPUINFO, RUN_KDE_OSRELEASE, RUN_KDE_SYS_MODULE_DIR, RUN_KDE_STORE_DIR, RUN_KDE_FUSE_DEVICE, RUN_KDE_VIRTIOFSD, RUN_KDE_FREE_GB, RUN_KDE_CPUS, RUN_KDE_HOST_RENDERER and RUN_KDE_WSLCONFIG, besides WAYLAND_DISPLAY and DISPLAY; and RUN_KDE_GUEST and RUN_KDE_RUNNER_BIN, which the package sets, and RUN_KDE_QEMU_OPTS, which adds to the options of QEMU. The checks run-kde-dry-run and run-kde-doctor (nix/linux/tests/run-kde.nix) run it against a WSL 2 with WSLg and a GPU, one with WSLg and a software OpenGL, one with WSLg and no GPU device, a Linux machine with no window, --tcg with no KVM, and machines that cannot run the guest (no KVM, a KVM the user cannot open, a virtiofsd that exits), a /dev/fuse that is missing or closed, and compare the output with the texts in nix/linux/tests/run-kde. The check run-kde builds the package, which shellchecks the script.

The UTM disk

UTM boots plasma-utm from a disk image. The image is made by installing the system onto a blank disk from a running Linux system with nixos-install, which needs no virtualization on the builder: here, the plasma-vm guest, which has the tools for it. Build both systems first: the plasma-vm runner as above, and the plasma-utm system, on the builder, copied back:

system=$(nix build .#nixosConfigurations.plasma-utm.config.system.build.toplevel \
  --eval-store auto --store "$KOMOREBI_BUILDER" --no-link --print-out-paths)
nix copy --from "$KOMOREBI_BUILDER" --substitute-on-destination --no-check-sigs "$system"

The closure goes into the guest as one file that nix-store --export writes, and not through the guest's view of the Mac's store: macOS stores files whose names differ only by case under other names (~nix~case~hack~1), which a guest reading the directory sees, and nix then rejects the hashes. An export undoes that. Then boot plasma-vm with a blank disk of 40 GB attached:

mkdir -p .vm/utm && cd .vm
nix-store --export $(nix-store -qR "$system") > utm/plasma-utm.closure
echo "$system" > utm/plasma-utm.system
nix shell nixpkgs#qemu --command qemu-img create -f raw utm/plasma.raw 40G
QEMU_OPTS="-drive file=$PWD/utm/plasma.raw,format=raw,if=none,id=utm -device virtio-blk-pci,drive=utm" \
  KOMOREBI_REPO=$PWD/.. ./result-plasma-vm/bin/run-plasma-vm-vm

In the guest's console, partition the disk, which is /dev/vdb, and install the system onto it:

sudo parted -s /dev/vdb -- mklabel gpt mkpart ESP fat32 1MiB 512MiB set 1 esp on mkpart nixos ext4 512MiB 100%
sudo mkfs.vfat -F 32 -n ESP /dev/vdb1
sudo mkfs.ext4 -q -L nixos /dev/vdb2
sudo mount /dev/vdb2 /mnt && sudo mkdir -p /mnt/boot && sudo mount /dev/vdb1 /mnt/boot
sudo nix-store --import --store /mnt < /repo/.vm/utm/plasma-utm.closure > /dev/null
sudo nixos-install --system "$(cat /repo/.vm/utm/plasma-utm.system)" --root /mnt --no-root-passwd --no-channel-copy
sudo umount /mnt/boot /mnt && sudo poweroff

Back on the Mac, in .vm, the raw disk becomes the qcow2 that UTM is pointed at (about 12 GB used of 40):

nix shell nixpkgs#qemu --command qemu-img convert -O qcow2 utm/plasma.raw utm/plasma.qcow2
rm utm/plasma.raw utm/plasma-utm.closure

Point UTM at it

1. Install UTM (brew install --cask utm, or from its website).
2. New VM → Virtualize → Linux. Leave the boot ISO empty. Give it 8192 MB and 6 cores. (The wizard insists on a new disk; delete it in step 3.)
3. Before the first start, open the VM's settings:
   • Drives: remove the new disk, then Import .vm/utm/plasma.qcow2 (interface VirtIO).
   • Display: Emulated Display Card virtio-gpu-gl-pci (GPU Supported). New Linux VMs get it by default. Renderer backend Default (ANGLE on Metal) or Apple Core OpenGL (GL 4.1).
   • Network: Shared Network (the guest gets an address the Mac can reach), or Emulated VLAN with port forwards 2224→22 and 3389→3389.
   • Name it plasma-utm.
4. Start it. sddm logs in by itself and Plasma comes up with komorebi running.

Using it

The VM's window is the desktop: click into it and use it. The pointer is an absolute tablet, so it moves in and out of the window without being grabbed; if the window does capture it, Control+Option releases it (UTM's default; Command+Option also works, and Control+Option collides with VoiceOver). The Option key is Alt in the guest. macOS takes a few combinations before UTM sees them (Control-arrow switches Spaces, and anything with Command), and those never reach the guest. Resizing the window resizes the guest's screen, and the clipboard is shared (spice-vdagent runs in the guest).

To check that the session is on the GPU, from a terminal in the guest (or over ssh, once a key of yours is in ~/.ssh/authorized_keys):

eglinfo 2>/dev/null | grep -i renderer                        # virgl (Apple M...)
journalctl --user -b -t kwin_wayland | grep -i "opengl\|renderer" | head
komorebic state | head

RDP

The session guests run KRDP, Plasma's RDP server, as the user unit krdp, on port 3389 (forwarded from localhost:3389 for plasma-vm-session). It logs in the session's user, komorebi, with a password of your choosing, which the unit reads from a file in the guest; until the file is there, the unit does not start. In the guest:

umask 077 && mkdir -p ~/.config/komorebi-vm && cat > ~/.config/komorebi-vm/rdp-password   # type it, then Ctrl-D
systemctl --user start krdp

The server makes a self-signed certificate the first time it starts, so a client warns about it. The password is passed to krdpserver on its command line, which every user of the guest can read in the process list. PLATFORM.md says what is known about RDP into the session, and what is not.

What is not known

At time of writing (2026-10-28), the guests of a Mac, made from these files, have been built on an aarch64-linux machine, and not booted: no test covers them, and plasma-guests-eval only evaluates them. The x86_64 pair for a Linux host was built and started under KVM on an x86_64 machine. Both guests reached the login on their serial console and had one DRM device with a render node, and the session guest drew the Plasma desktop on its display (a screenshot taken through QEMU's monitor) with komorebi running in the session. The demo guest, run-kde-x86_64, was started under KVM with a state directory: ssh took the key from it, the RDP unit started with the password from it, komorebi tiled the four windows, and the cheat sheet opened. run-kde --doctor, --dry-run, --build-only and a start with --display none were run on the same machine, which has no WSL, no window and no GPU: the guest came up with the memory and cores asked for, and it said software (llvmpipe). The GPU path was started on that machine in QEMU's GTK window on an X server whose OpenGL is Mesa's llvmpipe. Nothing was run on WSL 2, with WSLg, with a GPU, or with mstsc.exe: what the command does there is not known yet. The aarch64 guests for a Linux host were evaluated and not built. The RDP server is the exception to all of this: the VM suite rdp runs the same unit in the test session. That KWin 6.8 runs well on VirGL under UTM is not known either: the UTM and Mesa notes say VirGL is the way to render a Linux desktop, and Plasma works on it for other people, but it has not been measured here.