cellos-host-windows 0.6.0-pre

Windows host backend for CellOS — per-cell Job Object isolation (process-tree containment, memory / CPU caps, kill-on-close teardown) for hosts without KVM, Firecracker, or gVisor.
Documentation

cellos-host-windows

A Windows CellBackend for CellOS. It isolates a cell inside a Windows Job Object — the closest native analogue to a Linux cgroup-v2 leaf — for hosts that cannot run the Linux-only backends (Firecracker, gVisor).

Select it at runtime with:

$env:CELLOS_CELL_BACKEND = "windows"

On a non-Windows host this value fails closed at supervisor startup.

What it does

For each cell the backend creates a Job Object and uses it to:

Concern Mechanism
Process-tree containment the workload + every descendant are job members
Memory cap spec.run.limits.memoryMaxBytes → JOB_OBJECT_LIMIT_JOB_MEMORY
CPU cap spec.run.limits.cpuMax → hard JOBOBJECT_CPU_RATE_CONTROL_INFORMATION
Residue-free teardown JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE + TerminateJobObject

The backend owns execution: when the spec declares run.argv it spawns the workload itself (cleared environment, fixed PATH), assigns it to the job, and surfaces the exit code through wait_for_in_vm_exit — the same posture as the gVisor backend. A cell with no run.argv still gets a job (containment + teardown) and defers execution to the supervisor.

Honest scope (v1)

A Job Object is process-level containment, not a virtualization boundary. It has no separate kernel, no syscall emulation, no filesystem mount namespace, and no network-egress enforcement. Its isolation tier matches the Linux proprietary (cgroup-only) backend — weaker than gVisor, far weaker than a Firecracker microVM. For a hardware/syscall boundary, run the cell on Linux with CELLOS_CELL_BACKEND=firecracker (or gvisor).

Two gaps are enforced fail-closed rather than ignored:

  • Network egress is not enforced. A spec that declares authority.egressRules gets a per-cell tracing::warn. nft_rules_applied is always None.
  • secretDelivery: env + secretRefs is refused. Because the backend owns execution, the supervisor's env secret-injection path is bypassed; create returns InvalidSpec instead of running the workload without its secrets. Use a broker-socket delivery mode (the default runtimeLeasedBroker), where the child fetches secrets over a socket.

Known limitation: spawn→assign race

The workload is assigned to its job immediately after spawn, leaving a sub-millisecond window in which the child could spawn a grandchild outside the job. TerminateJobObject plus kill-on-close still bound the tree, and the assigned child's own descendants are always contained. A future revision can close the window with CREATE_SUSPENDED + resume once the supervisor exposes a thread handle.

Platform gating

The real implementation (backend.rs, job.rs) compiles only on Windows; the windows crate dependency is pinned to cfg(windows). On other targets the crate exposes a stub WindowsCellBackend whose methods return a Host error, so cargo check --workspace stays green on Linux/macOS CI.