`ThreadPoolExecutorTest.test_no_stale_references` may hang with GIL disabled

[colesbury]

In #114824, I modified test_no_stale_references so that it passes in the --disable-gil build. Unfortunately, that change was not sufficient and the test may still hang once the GIL is actually disabled.

Relevant code:

cpython/Lib/test/test_concurrent_futures/executor.py

Lines 84 to 87 in a25c02e

	my_object = MyObject()
	my_object_collected = threading.Event()
	my_object_callback = weakref.ref(
	my_object, lambda obj: my_object_collected.set())

cpython/Lib/test/test_concurrent_futures/executor.py

Line 103 in a25c02e

collected = my_object_collected.wait(timeout=support.SHORT_TIMEOUT)

The problem is due to the combination of two issues:

• Due to biased reference counting, the destructor for my_object is usually called on the main thread asynchronously (by the eval breaker logic)
• The destructor may be called somewhere in the implementation of my_object_collected.wait(). The my_object_collected.wait() implementation holds some of the same locks that my_object_collected.set() also needs. This can lead to deadlock if the timing is unlucky: the my_object_collected.set() call from the weakref callback tries to acquire locks already held by the current thread and deadlocks.

Linked PRs

• gh-117688: Fix deadlock in test_no_stale_references with GIL disabled #117720