SHARD-213: drive Virtualization.framework through a per-VM shim the daemon re-adopts by socket by presmihaylov · Pull Request #177 · presmihaylov/shard

presmihaylov · 2026-09-19T13:35:24Z

SHARD-213. Fourth of the M10 stack, stacked on #176.

Why. M10 needs a Virtualization.framework driver that survives a daemon restart and stays under the framework's two-VMs-per-process ceiling.

How.

Look first at cmd/shard-vz-shim/main.go and pkg/vz/server.go: one detached shim per VM, one length-prefixed JSON request per unix connection; connect splices a vsock stream onto it.
pkg/vz/client.go keeps no connection between verbs, so re-adopting a VM is only dialing its socket.
machine_darwin.go follows hypeman's device assembly (NOTICE). Save and restore are darwin arm64 only; elsewhere ErrUnsupported.
Zero cpus keeps the framework default; zero memory is shard's 512 MiB DefaultMemory, not the 4 MiB floor. Out of range is refused, never clamped.
The restore test skips while the login session is locked: a locked screen withholds the Secure Enclave key (Code=12, permission denied, spike item 9).

What. pkg/vz, cmd/shard-vz-shim, make build-shard-vz-shim. Darwin tests passed on nairiclaw.

…aemon re-adopts by socket pkg/vz is the thin driver: a Config, the bounds and host probes on every platform, and the darwin VM assembly behind a build tag. cmd/shard-vz-shim holds one VM and answers the verbs over a length-prefixed unix socket, so a daemon restart re-adopts a running VM instead of losing it, and the framework's two-VMs-per-process ceiling never binds. The darwin tests boot a fixture PID 1 from the shard kernel, prove it answers over vsock, save and restore under the same identifier, and re-adopt a VM whose starter was killed. The restore test skips when the login session is locked: the helper unwraps the saved state with a Secure Enclave key that a locked screen withholds (Code=12, permission denied).

…s, a claimed socket, and bounded socket calls - A zero memory request boots on 512 MiB, as the docs promise, not the framework's 4 MiB minimum. - The VM retains every *os.File behind a device and closes them after the VM ends, so a collection in the shim cannot close a live descriptor. - The shim claims its socket before the boot: a live owner is refused, and Start refuses an answer from a pid it did not start. - Every client call is bounded, dial to reply, and a connect stream clears the deadline once the handshake is in. - Serve bounds the handshake and closes the connections still in it when the listener closes.

…every handshake, and say what a zero memory means The guest fixture now writes to /dev/hvc0: on macOS 26 a write to /dev/console never reaches the host console file, and the GC regression read it.

presmihaylov mentioned this pull request Sep 19, 2026

SHARD-214: embed the shim in the daemon and ad-hoc sign it at first use #178

Merged