Description
Kiri on Windows today runs the x86-64 Linux build inside a WSL 2 distro, and kiri-win.exe carries the engine API over npipe through one wsl.exe … kiri bridge. That keeps the host side small, but it also means the guest Kiri runs on Windows is Microsoft's kernel, not the 10 MB kernel in guest/kernel that the macOS path is built around. A fair number of the open Windows issues trace back to that one fact:
- Windows: containers fail on a stock WSL2 — Kiri Linux boots a hybrid cgroup tree, so kiri-init exits 1 #277 — the WSL kernel boots a hybrid cgroup tree, so
kiri-initexits 1 - Bridge-network containers have no outbound connectivity on Windows/WSL 2 (no MASQUERADE rule for bridge subnets) #44 — no MASQUERADE rule, so bridge-network containers have no egress
- Security: Linux, WSL: containers reach the internet and are firewalled on kernels before 6.3 #60 — containers reach the internet unfirewalled on kernels before 6.3
- Windows: docker build hangs when a RUN step needs network (BuildKit has no usable DNS) #209 — BuildKit has no usable DNS
- Windows: Stop leaves the WSL VM (vmmemWSL) running and its memory unclaimed #189 — Stop leaves
vmmemWSLrunning with its memory unclaimed - Windows: setup fails with WSL_E_WSL_OPTIONAL_COMPONENT_REQUIRED - the "Subsystem for Linux" optional component is never enabled #297, xtask windows-dev: pick a real WSL distro on Windows #314, Windows: Docker works again after WSL restarts #318 — the distro itself as an install and lifecycle dependency
- and the engine API has to stay on
npipe, because WSL's localhost relay drops half-closes (docs/features.md)
So: would you consider giving Windows a VM backend of Kiri's own — libkrun and the Windows Hypervisor Platform — instead of a WSL 2 distro?
What that would bring to the architecture Kiri already has:
- Your own guest kernel.
krun_set_kerneltakes a rawImage, soguest/out/Imageand its four patches could be the Windows kernel too — cgroup v2, netfilter, the GSO/GRO fixes — instead of whatever WSL ships. - A real vsock channel. libkrun provides
virtio-vsock, andkrun_add_vsockhands the host plain listen/connect fds: the same shape as thekiri-protolink on macOS, and it removes the reason the Windows API sits onnpipeat all. - virtio-net with host port forwarding, instead of a relay that drops half-closes, plus TSI for containers that only need outbound.
- virtio-fs and virtio-block — the same devices
crates/kiri-vmmalready drives on Virtualization.framework, so Windows would be a sibling backend rather than a separate engine. - Memory that goes back to Windows.
virtio-balloonplus idle vCPUs sleeping in the hypervisor is what Windows: Stop leaves the WSL VM (vmmemWSL) running and its memory unclaimed #189 asks for. libkrun's own balloon is free-page reporting only, so this one needs checking before it counts. - Less install surface. One signed
.msi, no distro, no Store, no separatewsl --updateto keep working.
What it costs, as far as I can tell: WHP is still a Hyper-V platform feature, so the virtualization prerequisite does not go away (the WSL distro does); libkrun's own README still lists WHP as upcoming; and the WHP path is x86-64 only as far as I can see, so Kiri's Windows on Arm build would be the odd one out or at least unproven. The work is a new backend, not a port.
The only existing evidence I have that libkrun on WHP runs at all is smolvm, which I am mentioning as a data point, not as something for Kiri to depend on. What I would like to know:
- Is WSL 2 the intended long-term Windows story, or a stepping stone?
- Have you looked at libkrun/WHP and set it aside? If so, on what ground — vsock on Windows, USB passthrough, signing, or the maintainer cost?
- Would you review a proposal for a second VM backend? If yes, how would you want the Windows guest assets (kernel, data disk) built and shipped?
Thanks — the macOS and Linux paths are genuinely impressive, and Windows is the one place where it does not feel like the same engine.
Activity