RESUME POINT — cubefs FUSE <-> cube-server daemon durability wiring

Saved: 2026-08-13 (user paused session; pick up from here).

GOAL (from task list)

Make cubefs a real FUSE view of the running cube-server daemon's durable store, and prove a FUSE write lands in the daemon WAL and survives a daemon restart. (Cube-server is a pure command server — it does NOT mount FUSE itself; the FUSE view is a separate cubefs-mount --socket process, per the PDF/spec architecture.)

WHAT IS DONE (compiles + tests green)

Raw-coordinate command family added to cubesys/src/commands.rs dispatch: rawget <c> <z> <y> <x>, rawput <c> <z> <y> <x> <hex>, rawdel <c> <z> <y> <x>, rawkeys, rawscan <c> [z] [y]. Helpers added: parse_u8, hex_encode, hex_decode. (rawput/rawdel go through the durable ConcurrentStore so they hit the WAL.)
cubefs/src/backend.rs — new DaemonBackend implementing CubeBackend over the daemon Unix socket (lazy length-prefixed framing, duplicated from cubesys::net to keep cubefs free of a cubesys dep). Registered in cubefs/src/lib.rs (pub mod backend; pub use backend::DaemonBackend;). DaemonBackend is Clone (clones share socket path, reconnect lazily).
cubefs/src/bin/cubefs_mount.rs — --socket PATH mounts against a live daemon; without --socket it falls back to an in-memory HashBackend.
cargo build --workspace GREEN. cargo test --workspace GREEN. (Note: DaemonBackend is NOT yet exercised by any unit/integration test — only manual host verification below.)

WHAT IS PROVEN (host-level, same kernel/code path as the VM)

A. Daemon durability + WAL replay: launched daemon with --store, rawput, killed it with kill -9 (crash), relaunched same --store → rawget returned the written value; recovery log showed wal_recovery: applied 1. So the DAEMON half of the durability story works. B. Direct rawput from a Python client checkpoints to store.json (4-byte record appeared after the checkpoint interval). Confirmed daemon is alive, reachable, and durable on the host.

WHAT IS BROKEN (the open bug) — RESOLVED 2026-08-13 (commit `0300def`)

Two separate bugs hid behind the "FUSE write didn't reach the daemon" symptom:

--features mount did not compile (E0308: CubeFs<DaemonBackend> vs CubeFs<HashBackend> are distinct types in the match &socket). The build failed silently, so the running binary was the IN-MEMORY build and --socket was ignored. FIXED in 8f9e7e0 by type-erasing the backend (Box<dyn CubeBackend + Send + Sync> + a forwarding impl).
Even after (1), DaemonBackend round-tripped the WRONG layer: put stripped the record envelope to a bare body before rawput, but get returned that bare body as if it were a full record — so the FUSE read-back path (get_record) could not decode it, and the socket backend diverged from the in-memory backend. FIXED in 0300def: DaemonBackend::put/get now round-trip the FULL envelope (header+body) verbatim through rawput/rawget, matching the in-memory HashBackend exactly. This is now an enforced ./check daemon integration test (spins up a live cube-server, runs the two #[ignore]d daemon-backed tests via --ignored).

GATE STATUS (2026-08-13, end of session)

./check (fmt+test+clippy -D warnings) GREEN.
./check daemon GREEN — proves the socket-backed record round-trip against a real daemon (3 ignored tests via --ignored all pass).
./check mount (live FUSE, root) GREEN — 27/27 pass. This is the regression guard exercising the real kernel VFS path; it now runs cleanly with the fixed durable-mount code. Run before any future filesystem work.

NEXT STEPS (current)

~~Repeat the single-daemon + single-mount e2e INSIDE the VM...~~ DONE (see below).
Add a --features mount build to the VM image's build step / ./check so the durable mount can't silently regress to in-memory again. (DONE on host ./check; VM image build step not yet updated.)
Optionally surface DaemonBackend::put failures as FUSE EIO instead of eprintln-only (best-effort today, acceptable).
~~Confirm the desktop "CUBE Shell" launcher execs cubec...~~ DONE 2026-08-13: /usr/share/applications/cube.desktop -> Exec /opt/cube/bin/cube-term, which exports CUBE_SOCKET=/run/cube/cube.sock and exec /opt/cube/bin/cubec. The skill's #1 gotcha (must use cubec, not cube repl) is satisfied at the file level (not visually launched, per user instruction to not block on viewability).

COMPLETED THIS SESSION

Fixed DaemonBackend envelope round-trip (commit 0300def) — was returning the bare body instead of the full record envelope; the --socket mount could not read back FUSE writes. Fixed by round-tripping the full envelope verbatim.
Fixed --features mount compile break (commit 8f9e7e0) — Czyx field access.
Enforced BOTH bugs via ./check daemon (the two #[ignore]d integration tests now run via --ignored and would fail if either regressed).
Host: ./check + ./check daemon + ./check mount all GREEN.
VM (/var/lib/cubevm/cubelinux-vm.qcow2, was pre-fix): rebuilt cubefs-mount from the fixed source, rewired cubefs.service to --socket /run/cube/cube.sock with Requires=cube-server.service, and proved a FUSE write survives a daemon restart (WAL persistence). Injected SSH key for cube+root to drive the VM. ./check also passes inside the VM (rustfmt+clippy components added).

KEY FILES

/home/CUBELinux/CUBELinux-2/cubesys/src/commands.rs (raw* commands + helpers)
/home/CUBELinux/CUBELinux-2/cubefs/src/backend.rs (DaemonBackend — FIX HERE)
/home/CUBELinux/CUBELinux-2/cubefs/src/lib.rs (module export)
/home/CUBELinux/CUBELinux-2/cubefs/src/bin/cubefs_mount.rs (--socket wiring)
/home/CUBELinux/CUBELinux-2/cubestore/src/lib.rs (CubeStore::put_raw L358)
/home/CUBELinux/CUBELinux-2/cubefs/src/vfs.rs (CubeFs::write L392)

TEST HARNESS NOTES

Launch daemon/mount as terminal(background=true) so Hermes tracks them.
A python3 one-shot client for rawget/rawput: socket connect, sendall(struct.pack('<I',len(payload))+payload), read 4-byte len then body.
Daemon flags used: --socket S --store STORE.json --recovery-log REC.ndjson --allow-anonymous --checkpoint-ms 400 --wal-fsync-ms 30

6.3 KiB Raw Blame History