Separates what a VM IS (durable intent + identity + deterministic
derived paths — `VMRuntime`) from what is CURRENTLY TRUE about it
(firecracker PID, tap device, loop devices, dm-snapshot target — new
`VMHandles`). The durable state lives in the SQLite `vms` row; the
transient state lives in an in-memory cache on the daemon plus a
per-VM `handles.json` scratch file inside VMDir, rebuilt at startup
from OS inspection. Nothing kernel-level rides the SQLite schema
anymore.
Why:
Persisting ephemeral process handles to SQLite forced reconcile to
treat "running with a stale PID" as a first-class case and mix it
with real state transitions. The schema described what we last
observed, not what the VM is. Every time the observation model
shifted (tap pool, DM naming, pgrep fallback) the reconcile logic
grew a new branch. Splitting lets each layer own what it's good at:
durable records describe intent, in-memory cache + scratch file
describe momentary reality.
Shape:
- `model.VMHandles` = PID, TapDevice, BaseLoop, COWLoop, DMName,
DMDev. Never in SQLite.
- `VMRuntime` keeps: State, GuestIP, APISockPath, VSockPath,
VSockCID, LogPath, MetricsPath, DNSName, VMDir, SystemOverlay,
WorkDiskPath, LastError. All durable or deterministic.
- `handleCache` on `*Daemon` — mutex-guarded map + scratch-file
plumbing (`writeHandlesFile` / `readHandlesFile` /
`rediscoverHandles`). See `internal/daemon/vm_handles.go`.
- `d.vmAlive(vm)` replaces the 20+ inline
`vm.State==Running && ProcessRunning(vm.Runtime.PID, apiSock)`
spreads. Single source of truth for liveness.
- Startup reconcile: per running VM, load the scratch file, pgrep
the api sock, either keep (cache seeded from scratch) or demote
to stopped (scratch handles passed to cleanupRuntime first so DM
/ loops / tap actually get torn down).
Verification:
- `go test ./...` green.
- Live: `banger vm run --name handles-test -- cat /etc/hostname`
starts; `handles.json` appears in VMDir with the expected PID,
tap, loops, DM.
- `kill -9 $(pgrep bangerd)` while the VM is running, re-invoke the
CLI, daemon auto-starts, reconcile recognises the VM as alive,
`banger vm ssh` still connects, `banger vm delete` cleans up.
Tests added:
- vm_handles_test.go: scratch-file roundtrip, missing/corrupt file
behaviour, cache concurrency, rediscoverHandles prefers pgrep
over scratch, returns scratch contents even when process is
dead (so cleanup can tear down kernel state).
- vm_test.go: reconcile test rewritten to exercise the new flow
(write scratch → reconcile reads it → verifies process is gone →
issues dmsetup/losetup teardown).
ARCHITECTURE.md updated; `handles` added to Daemon field docs.
Previously withVMLockByRef held the per-VM mutex across InspectRepo,
waitForGuestSSH, dialGuest, ImportRepoToGuest (the tar stream!), and
the readonly chmod. A large repo could block `vm stop` / `vm delete`
/ `vm restart` on the same VM for however long the import took.
Split into two phases:
1. VM mutex held briefly to validate state (running + PID alive)
and snapshot the fields needed for SSH (guest IP, api sock).
2. VM mutex released. Acquire workspaceLocks[id] — a separate
per-VM mutex scoped to workspace.prepare / workspace.export —
for the guest I/O phase.
Lifecycle ops (stop/delete/restart/set) only take vmLocks, so they
no longer queue behind a slow import. Two concurrent prepares on the
same VM still serialise via workspaceLocks so tar streams don't
interleave. ExportVMWorkspace also acquires workspaceLocks to avoid
snapshotting a half-streamed import.
Two regression tests (sequential — they swap package-level seams):
ReleasesVMLockDuringGuestIO: stall the import fake, assert the VM
mutex is acquirable from another goroutine during the stall.
SerialisesConcurrentPreparesOnSameVM: 3 concurrent prepares, assert
Import is only ever invoked 1-at-a-time per VM.
ARCHITECTURE.md documents the split + updated lock ordering.
The `image build` flow spun up a transient Firecracker VM, SSHed in,
and ran a large bash provisioning script to derive a new managed
image from an existing one. It overlapped heavily with the golden-
image Dockerfile flow (same mise/docker/tmux/opencode install logic
duplicated in Go as `imagemgr.BuildProvisionScript`) and had far more
machinery: async op state, RPC begin/status/cancel, webui form +
operation page, preflight checks, API types, tests. For custom
images, writing a Dockerfile is simpler and more reproducible.
Removed end-to-end:
- CLI `image build` subcommand + `absolutizeImageBuildPaths`.
- Daemon: BuildImage method, imagebuild.go (transient-VM orchestration),
image_build_ops.go (async begin/status/cancel), imagemgr/build.go
(the 247-line provisioning script generator and all its append*
helpers), validateImageBuildPrereqs + addImageBuildPrereqs.
- RPC dispatches for image.build / .begin / .status / .cancel.
- opstate registry `imageBuildOps`, daemon seam `imageBuild`,
background pruner call.
- API types: ImageBuildParams, ImageBuildOperation, ImageBuildBeginResult,
ImageBuildStatusParams, ImageBuildStatusResult; model type
ImageBuildRequest.
- Web UI: Backend interface methods, handlers, form, routes, template
branches (images.html build form, operation.html build branch,
dashboard.html Build button).
- Tests that directly exercised BuildImage.
Doctor polish (task C):
- Drop the "image build" preflight section entirely (its raison d'être
is gone).
- Default-image check now accepts "not local but in imagecat" as OK:
vm create auto-pulls on first use. Only flag when the image is
neither locally registered nor in the catalog.
Net: 24 files touched, 1,373 lines deleted, 25 added.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
internal/daemon/doc.go and ARCHITECTURE.md were written before the
subpackage extractions and still referenced old structure (in-progress
phrasing, missing opstate/dmsnap/fcproc/imagemgr/session/workspace,
mentions of opRegistry by its old name). Both now describe the current
shape: composition root + six leaf subpackages, lock ordering rooted
at vmLocks[id], and the one intra-package dependency (workspace →
session for ShellQuote + FormatStepError).
README.md and AGENTS.md mark the local web UI as experimental. It is
still enabled by default at 127.0.0.1:7777, but the docs now state
plainly that its surface is not stable or hardened and not intended for
anything beyond single-user localhost use. AGENTS.md also points at
ARCHITECTURE.md for the subpackage layout.
No code changes; tests still green.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Daemon no longer owns a coarse mu shared across unrelated concerns.
Each subsystem now carries its own state and lock:
- tapPool: entries, next, and mu move onto a new tapPool struct.
- sessionRegistry: sessionControllers + its mutex move off Daemon.
- opRegistry[T asyncOp]: generic registry collapses the two ad-hoc
vm-create and image-build operation maps (and their mutexes) into one
shared type; the Begin/Status/Cancel/Prune methods simplify.
- vmLockSet: the sync.Map of per-VM mutexes moves into its own type;
lockVMID forwards.
- Daemon.mu splits into imageOpsMu (image-registry mutations) and
createVMMu (CreateVM serialisation) so image ops and VM creates no
longer block each other.
Lock ordering collapses to vmLocks[id] -> {createVMMu, imageOpsMu} ->
subsystem-local leaves. doc.go and ARCHITECTURE.md updated.
No behavior change; tests green.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
vm.go (1529 LOC) splits into vm_create, vm_lifecycle, vm_set, vm_stats,
vm_disk, vm_authsync; firecracker/DNS/helpers stay in vm.go.
guest_sessions.go (1266 LOC) splits into session_controller,
session_lifecycle, session_attach, session_stream; scripts and helpers
stay in guest_sessions.go.
Mechanical move only. No behavior change. Adds doc.go and
ARCHITECTURE.md capturing subsystem map and current lock ordering as
the baseline for the upcoming subsystem extraction.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
