docs: the volume-manager plan — V0 unmount ownership through V5 close-out

2026-08-09 16:09:20 +01:00
parent 9da63e81e2
commit 451abba000
1 changed files with 115 additions and 0 deletions
@@ -0,0 +1,115 @@
 # The volume manager: the plan
 *2026-08-09. Executes the settled design in
 [storage-architecture.md](file-system-development/storage-architecture.md) and
 [storage-design-rationale.md](file-system-development/storage-design-rationale.md)
 (decisions 1–8). Track discipline as always: one commit per coherent step, suite
 green at phase boundaries, every new test shown to fail against the old
 behavior, one QEMU suite at a time, work on main.*
 **One refinement of decision 4, flagged for sign-off rather than silently
 applied.** The decision said endpoint-per-volume. The service harness serves one
 endpoint per process, and kernel-ipc has no wait-on-many; a driver serving N
 range endpoints would need threads or a multi-endpoint harness — real machinery,
 none of it needed for the security property. The property ("a channel carries
 exactly the authority it grants") is delivered instead by **per-sender range
 confinement**: every packet already arrives with the kernel-stamped,
 unforgeable badge; the storage driver keeps a per-badge range (set by the
 volume manager, which spawned the filesystem process and knows its id) and
 clamps-and-translates every transfer by the sender's range. A filesystem
 process addresses volume-relative LBAs from 0; the provider adds the base —
 offset translation at the provider, exactly the Fuchsia session-mapping shape,
 and the FAT engine's `base_lba` code is deleted rather than moved.
 Endpoint-per-volume can still arrive later with a multi-endpoint harness; the
 wire contract does not change either way. **If this refinement is wrong, say so
 before V2.**
 ## V0 — `fs_unmount` ownership (the defect fix)
 The kernel records the mounting process on each mount slot; `fs_unmount` is
 refused for any caller but the owner (the `/protocol` special case stays).
 Death cleanup is unaffected (the lazy sweep is not an unmount). Discrimination:
 a fixture unmounting a prefix it does not own must be refused — fails against
 today's kernel, which lets any process unmount anything.
 ## V1 — the filesystem harness (extraction, no behavior change)
 fat's 434-line shell becomes `library/file-system/harness` (name per
 convention): establishment, the badge-scoped open-node table, the nine
 vfs-protocol handlers, mount registration, bring-up/teardown. fat becomes
 engine + on-disk + a thin main wiring the harness. Suite green is the gate;
 nothing observable changes. This lands FIRST so every later phase touches the
 harness once, not fat and the harness both.
 ## V2 — the driver mechanism: ranges, confinement, presence
 - Block protocol additions (appended; numbering holds): `define_range`
  (volume-manager-only in practice — see grants), and the pushed
  `medium_changed` event (present/absent + change counter).
 - usb-storage: per-badge range table (clamp + translate per sender), range
  definitions from the volume manager, and presence: a slow idle-time
  TEST UNIT READY poll plus sense-key inspection on failed transfers, emitting
  `medium_changed` on transitions. QEMU test lever: `eject` /
  `blockdev-remove-medium` against a `removable=on` usb-storage device; if
  QEMU's model refuses, the fallback drill is device_del/add of the whole
  stick (the H-matrix already proves that path) and presence gets its real
  test on the bench with a card reader.
 - Discrimination: a fixture transferring outside its assigned range must be
  refused; fails against a driver without the clamp.
 ## V3 — the volume manager service
 New binary `system/services/volume-manager`, spawned by init, serving the
 (genuinely singular) name `volume-manager`. Duties, all moved OUT of fat:
 - subscribe to the device manager; consumer-hello each storage provider for
  its block channel;
 - probe: partition table walk (MBR now, GPT next — the walk LEAVES the FAT
  engine) and content identity (the decision-8 ladder: GPT GUID → fs UUID →
  FAT serial+label → MBR signature+index → anonymous);
 - configuration: `filesystems.csv` (signature → filesystem binary) and
  `volumes.csv` (identity → mount prefix; the fstab). Boot volume identity
  recorded at first sight of `/system/configuration`;
 - define ranges on the driver; spawn one filesystem process per volume
  (argv: volume id); answer each filesystem's startup hello with its volume
  channel (the reply-capability path, same as the device manager's);
 - supervise: hello/mount deadline, crash-loop cap, reap on removal.
 fat sheds `acquireVolume` and its device-manager grant; filesystem binaries
 get `open volume-manager` only — a filesystem cannot acquire, only be given.
 Grants move with the code in the same commits.
 ## V4 — the removal lifecycle, end to end
 Two triggers, one path: storage-channel death and `medium_changed(absent)`
 both drive kill-the-filesystem-process + retire-its-mounts; return (device
 re-report or `medium_changed(present)`) drives re-probe → respawn → remount
 at the identity's prefix. The logger gains resume patience (retry flushes on
 the fat cadence, never abandon) and the ring-wrap gap marker. QEMU cases:
 - `volume-replug`: yank the boot stick mid-run, replug, assert remount at the
  same prefixes and the logger appending to the SAME boot-stamp tree with the
  gap marked. Discrimination: against pre-V4, fat wedges (`mounted` forever)
  and no remount happens.
 - `volume-two-partitions`: a two-partition FAT image → two volumes, two
  filesystem processes, two mounts from one stick; yank once, both die; return
  once, both remount. Proves multi-volume and the reap breadth. (New image
  fixture beside make-fat-image.py.)
 - `volume-clone-policy`: two sticks with identical FAT serials — first keeps
  the mapped name, second mounts suffixed, loudly logged.
 - The **lifecycle conformance drill**, parameterized by filesystem: mount,
  serve, yank mid-write, verify honest loss (dirty flag set, gap said),
  replug, remount. FAT is implementation #1; the drill is the definition of
  "danos supports filesystem X".
 ## V5 — close-out
 Full suite green; the architecture doc's *(planned)* markers flip to built;
 `fs_unmount` ownership, ranges, presence, identity, and the volume manager
 lose their future-tense; memory updated; Ryzen bench note: pull the stick,
 watch the log gap get marked, plug it back anywhere.
 **Not in this track** (recorded so absence is deliberate): GPT parsing beyond
 the identity read (row exists in the prober's ladder; full GPT when a GPT
 medium matters), AHCI/NVMe drivers (NVMe gated on the shm-ring data plane),
 formatting/entropy, per-process namespaces.