- Load snapshot with ResumeVM: false so MMDS data can be written while VM is paused
- Call ResumeVM explicitly after configureMmds succeeds
- Skip PUT /mmds/config on restored VMs (Firecracker rejects it with 400)
- Strip JSON quotes from MMDS values with tr -d '"' in net-init script
- Add 169.254.169.2/32 link-local addr and flush eth0 before applying new IP
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Introduces optional per-clone IP assignment using the Firecracker Microvm
Metadata Service (MMDS). A background daemon (fc-net-init) is baked into
the rootfs during init and captured in the golden snapshot — on clone
resume it polls 169.254.169.254 and applies the IP/GW/DNS config injected
by the orchestrator immediately after snapshot restore.
- config.go: add AutoNetConfig bool (FC_AUTO_NET_CONFIG=1)
- orchestrator.go: embed fc-net-init daemon + MMDS link-local route in
init script; set AllowMMDS: true on golden NIC; spawnOne/SpawnSingle
accept net bool and propagate it via FC_AUTO_NET_CONFIG in proxy env
- console.go: set AllowMMDS: true on clone NIC; call configureMmds()
after m.Start() when AutoNetConfig is enabled
- network.go: add configureMmds() — PUT /mmds with ip/gw/dns over the
clone's Firecracker Unix socket
- serve.go: POST /clones accepts optional {"net": bool} body to override
the global AutoNetConfig default per-request
- web/terminal.html: spawn button always sends {"net": true}
- docs/commands.md: document manual config + MMDS autoconfiguration
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
A "poor man's" Firecracker VM orchestrator that boots a single golden VM,
snapshots it, then restores N clone VMs from that snapshot with minimal
per-clone overhead.
How it works:
- `init` — downloads a Linux 6.1 kernel and builds a minimal Alpine 3.20
rootfs (512 MiB ext4) with a basic init script
- `golden` — boots the golden VM, lets it settle, then pauses and snapshots
it (vmstate + memory file); the golden VMM is then terminated
since only the artifacts are needed
- `spawn N` — restores N clone VMs concurrently from the golden snapshot:
* rootfs: filesystem-level COW copy via `cp --reflink` (falls
back to a plain copy if reflinks are not supported)
* memory: shared golden `mem` file; Firecracker's MAP_PRIVATE
lets the kernel handle COW page-by-page at no up-front cost
* vmstate: small file, cheap regular copy per clone
* networking: per-clone TAP device (fctapN) bridged to fcbr0
with iptables MASQUERADE NAT on the default route interface
- `status` — reads PID files and checks /proc to report alive/dead clones
- `kill` — stops in-memory clones, kills any stragglers via PID files,
and tears down all fctap* devices
- `cleanup` — kill + remove all state dirs and the bridge
All tunables (binary path, base dir, kernel/rootfs paths, vCPUs, memory,
bridge name/CIDR) are configurable via environment variables.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
