Files
OwnCord/Server/service
J3vbandClaude Opus 5 d6c768cb90 feat(invariants): add server invariant rules and close five deadlock blind spots (#1383)
* feat(invariants): add the invariant-rule harness and the syncutil-locks rule

* fix(ws,service): route the last five raw mutexes through syncutil

The -tags deadlock CI pass only observes locks declared via syncutil, whose
Mutex/RWMutex are build-tag aliases. These five were declared as raw sync
types and were invisible to it, including the hub voice key-holder lock and
the permission and role caches.

TestServerInvariants now gates the tree against regressions.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* fix(invariants): walk the tree through os.Root to close a symlink TOCTOU

gosec G122: reading a filepath.WalkDir-supplied path is race-prone, since a
symlink swapped between the walk and the read escapes the intended tree.
os.Root confines every read to the root and cannot be traversed out of.

Walking the root's fs.FS also yields slash-separated paths already relative
to it, so the filepath.Rel and ToSlash conversion is no longer needed.

* fix(invariants): close syncutil-locks evasions, isolate per-rule tests, harden the gate

- I1: TestServerInvariants now asserts every registered Rule.Scope
  directory exists and holds at least one non-test .go file, so the
  gate cannot pass by scanning nothing.
- I2: split CheckSource into a thin wrapper over an unexported
  checkSourceWith(rules, ...), so TestSyncutilLocks tests the
  syncutil-locks rule in isolation instead of the whole registry.
- I3: broaden checkSyncutilLocks to a single SelectorExpr match (any
  sync.Mutex/sync.RWMutex reference bound via f.Imports, aliases
  included) instead of only *ast.Field/*ast.ValueSpec. Catches :=
  composite literals, untyped var specs, type aliases, and
  []sync.Mutex/map[K]sync.Mutex, none of which the old rule saw. A
  dot-import of "sync" is now its own violation, since it would
  otherwise let a bare Mutex evade the selector match entirely.
- M2: suppression now keys off the violation's own Rule id
  (allowed[v.Line][v.Rule]) rather than the running rule's ID, so a
  rule that ever emits a sub-id isn't silently unsuppressible.
- M3: Run sorts with sort.SliceStable, since an unreasoned allow
  comment and the violation it fails to suppress can share a
  file:line.
- M4/M5/M1-partial: add a build-tag-gated fixture test, document that
  allow comments must be same-line, and correct the skipDirs comment
  to describe both the generated-code and gitignored-runtime-dir
  cases it actually covers.

All ten original TestSyncutilLocks subtests pass unchanged; six new
subtests cover the evasions above.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* docs(server): point at the syncutil-locks invariant gate

Server/CLAUDE.md told developers not to hand-roll around syncutil but
never said it's enforced. Note that Server/invariants/ checks it at
go test time and that exceptions are greppable via
grep -rn "invariant:allow" Server/.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-08-16 17:05:06 +02:00
..