Data race in lwt_unix_start_job

[OlivierNicole]

Something strange showed up in ThreadSanitizer logs: in Lwt_unix, more precisely in lwt_unix_start_job, the thread_waiting_count variable is written with the pool mutex taken:

lwt/src/unix/lwt_unix_stubs.c

Lines 1002 to 1010 in 68cf601

	lwt_unix_mutex_lock(&pool_mutex);
	DEBUG("waiting for something to do");
	/* Wait for something to do. */
	while (pool_queue == NULL) {
	++thread_waiting_count;
	lwt_unix_condition_wait(&pool_condition, &pool_mutex);
	}

This happens on the worker thread.

However, it can be read by the main thread without locking any mutexes:

lwt/src/unix/lwt_unix_stubs.c

Lines 1061 to 1069 in 68cf601

	CAMLprim value lwt_unix_start_job(value val_job, value val_async_method) {
	lwt_unix_job job = Job_val(val_job);
	lwt_unix_async_method async_method = Int_val(val_async_method);
	int done = 0;
	/* Fallback to synchronous call if there is no worker available and
	we can not launch more threads. */
	if (async_method != LWT_UNIX_ASYNC_METHOD_NONE && thread_waiting_count == 0 &&
	thread_count >= pool_size)

This looks like a genuine data race.

[OlivierNicole]

Similarly TSan reports unsynchronised read and write between:

lwt/src/unix/lwt_unix_stubs.c

Lines 912 to 918 in 68cf601

	static void execute_job(lwt_unix_job job) {
	DEBUG("executing the job");
	lwt_unix_mutex_lock(&job->mutex);
	/* Mark the job as running. */
	job->state = LWT_UNIX_JOB_STATE_RUNNING;

and

lwt/src/unix/lwt_unix_stubs.c

Line 1126 in 68cf601

done = job->state == LWT_UNIX_JOB_STATE_DONE;