dangernoodle-io / breadboard / 28735275474

100%

push

github

feat: add bb_cache_delete + runtime eviction (B1-592 A1) (#722)

- add bb_cache_delete(key): frees an entry's owned buffer + memoized
  serialized bytes and marks the slot reusable, holding s_reg_lock across
  the whole operation (wider than find_entry_locked) so a concurrent
  register can't reuse the slot mid-delete
- add free_entry_locked() helper mirroring bb_cache_reset_for_test's
  per-entry teardown
- fix bb_cache_foreach to snapshot keys by value (not raw pointers) since
  entries are no longer add-only
- document the known bb_event topic-unregister gap (SSE-flagged keys leak
  their event topic handle on delete) as a Phase-A limitation
- fix reader-vs-delete TOCTOU/UAF: per-slot mutexes are now process-lifetime
  (created once via pthread_once, never destroyed/re-init'd) with a
  generation counter bumped on every free<->in-use transition; every reader
  (update/post/serialize_into/post_serialized/get_serialized/get_raw)
  captures (entry, generation) under s_reg_lock then re-validates both under
  the entry's own lock, closing the destroyed-mutex UB and cross-key
  cache-poisoning windows
- fix bb_cache_key_at to copy the key by value instead of returning a raw
  registry pointer (same UAF class the foreach fix closed)
- dedup bb_cache_reset_for_test's per-entry teardown via free_entry_locked
- extend the coverage gate to platform/{espidf,host}/bb_cache/ and close
  every newly-visible branch gap to 100%, including a host-only
  bb_json_host_force_serialize_fail_after() fault-injection hook
- add multi-threaded delete-vs-reader host tests proving the race is closed
  (verified against the pre-fix logic, which fails deterministically and
  crashes)
- document the bb_cache_foreach stack-budget note on the WS-connect path

2624 of 2624 branches covered (100.0%)

Branch coverage included in aggregate %.

128 of 128 new or added lines in 1 file covered. (100.0%)

4390 of 4390 relevant lines covered (100.0%)

10727.4 hits per line