sm_coretest_post_during_exit_1

Core test that reproduces, without a simulator, the race between an event posted from the executor thread and a state exit on the state machine thread (Next Level Study NL-2). ISmaccStateMachine::postEvent reads currentState_.back() and walks that state’s reactors on the posting thread, while the state machine thread pushes and pops currentState_ and destroys the state under m_mutex_, which postEvent does not hold. The nav2 machines showed it under ThreadSanitizer on every action-driven transition; this test shows it in a few seconds with no simulator.

Design

• CpEventPulse (on ClPulse in OrPulse) posts EvPulse at pulse_rate_hz (default 200) from a node wall timer, that is from the executor thread, the posting thread of every subscription, timer and action callback.
• StA and StB alternate on the natural end of a 50 ms CbSleepFor in OrSleep, about twenty transitions a second. No state reacts to EvPulse.
• StA creates a SrEventCountdown on EvPulse with a count it never reaches, so every post walks into a reactor object owned by StA, whether StA is running or being destroyed.

What it shows

Under ThreadSanitizer on the current core: data races pairing propagateEventToStateReactors (the executor thread, inside postEvent<EvPulse>) with StA’s destruction on the state machine thread (the vptr write in ~SmaccState, the deallocation of the state, and the reactor’s eventCount_). After a fix that moves the reactor walk onto the state machine thread, none. The plain run is the regression check: the StA/StB alternation and the pulse count, with no TRAP-style failure, since no transition depends on EvPulse.

Running

ros2 launch sm_coretest_post_during_exit_1 sm_coretest_post_during_exit_1.launch.py

The launch file tees the node output to /tmp/sm_coretest_post_during_exit_1_latest.log. For the sanitizer run, build in the sanitizer base and run the node binary directly with the sanitizer environment of the CI gate script (.github/scripts/sanitizer_core_tests.sh).