dangernoodle-io / breadboard / 29299811138

95%

push

github

fix: synchronize bb_event test teardown with dispatcher drain (#854)

- bb_event_port_reset_for_test() joined the async dispatcher thread
  without waiting for it to drain a pending entry first; if the
  dispatcher hadn't yet been scheduled to pop the queued entry when
  thread_running flipped false, it exited without dispatching,
  silently dropping the entry
- since test_main.c's global setUp() calls this teardown before
  every one of ~5,135 host tests, whether dispatcher_loop's
  pop/dispatch path executed was pure OS-scheduling luck, making
  platform/host/bb_event/bb_event_host.c's coverage flaky and
  intermittently failing the coverage-baseline gate on unrelated PRs
  (#852, #853)
- add a dedicated drained_cond the dispatcher broadcasts on whenever
  its queue hits zero; teardown now waits on it (mutex correctly
  dropped across the wait via pthread_cond_timedwait, bounded by
  BB_EVENT_DRAIN_WAIT_TIMEOUT_MS=500ms so a genuinely stuck
  dispatcher cannot hang CI) before signalling shutdown
- add regression tests that deterministically force the race via a
  test-only dispatcher startup-delay hook and assert on actual
  delivery through a counting subscriber (not just no-crash/no-hang);
  verified by reverting only the drain-wait and confirming the
  drain-before-teardown test fails (Expected 1 Was 0), then
  restoring the fix and confirming it passes again
- add a test-only bb_event_port_test_get_last_drain_wait_ms() hook
  that times just the pthread_cond_timedwait segment (not the
  surrounding pthread_join, whose duration is dominated by the
  dispatcher's startup delay regardless of the drain-wait) and
  strengthen the timeout-fallback regression test to assert on it;
  the prior version of that test passed both with and without the
  fix (masked-by-construction -- its startup delay outlasted the
  join in both cases), confirmed empirically by reverting only the
  drain-wait block and re-running before adding this assertion
- add a comment... (continued)

8093 of 8860 branches covered (91.34%)

Branch coverage included in aggregate %.

44 of 45 new or added lines in 1 file covered. (97.78%)

13836 of 14231 relevant lines covered (97.22%)

2224.62 hits per line