PyMutex failure in parking_lot.c on Windows during interpreter shutdown

[colesbury]

Bug report

Bug description:

Fatal Python error: _PySemaphore_PlatformWait: unexpected error from semaphore: 4294967295 (error: 6)
Python runtime state: finalizing (tstate=0x6aa6c460)

Thread 0x00001360 (most recent call first):
  <no Python frame>
Windows fatal exception: code 0x80000003

Current thread 0x00001360 (most recent call first):
  <no Python frame>

Current thread's C stack trace (most recent call first):
  <cannot get C stack on this system>

4294967295 is the -1 return code from WaitForMultipleObjects. Error code 6 is ERROR_INVALID_HANDLE.

I'm pretty sure that the _PyOS_SigintEvent() handle is getting closed concurrently with the mutex wait.

cpython/Python/parking_lot.c

Lines 118 to 136 in 1ffe913

	HANDLE sigint_event = _PyOS_SigintEvent();
	HANDLE handles[2] = { sema->platform_sem, sigint_event };
	DWORD count = sigint_event != NULL ? 2 : 1;
	wait = WaitForMultipleObjects(count, handles, FALSE, millis);
	if (wait == WAIT_OBJECT_0) {
	res = Py_PARK_OK;
	}
	else if (wait == WAIT_OBJECT_0 + 1) {
	ResetEvent(sigint_event);
	res = Py_PARK_INTR;
	}
	else if (wait == WAIT_TIMEOUT) {
	res = Py_PARK_TIMEOUT;
	}
	else {
	_Py_FatalErrorFormat(__func__,
	"unexpected error from semaphore: %u (error: %u)",
	wait, GetLastError());
	}

Seen in https://github.com/python/cpython/actions/runs/15423887404/job/43405853684 at the end of running test_decorators.

CPython versions tested on:

CPython main branch

Operating systems tested on:

Windows

Linked PRs

• gh-135099: Only wait on _PyOS_SigintEvent() in main thread #135100
• [3.14] gh-135099: Only wait on _PyOS_SigintEvent() in main thread (GH-135100) #135116

[encukou]

Since #134747 ("Change PyThread_allocate_lock() implementation to PyMutex") was merged, Windows tests got really flaky, with several tests failing in each buildbot run with the same error:

Fatal Python error: _PySemaphore_PlatformWait: unexpected error from semaphore: 4294967295 (error: 6)
Python runtime state: finalizing (tstate=0x00007ff9a3c4aac8)
Thread 0x000023c8 (most recent call first):
  <no Python frame>
Windows fatal exception: code 0x80000003

For example, here: https://buildbot.python.org/#/builders/146/builds/11597/steps/5/logs/stdio

Currently at the release status page, on several Windows buildbots you can see a point where they "switched" from mostly-green to consistent "warnings" result; on the few I explored the first "yellow" build was for #134747.

Do you think that's related?

[colesbury]

Yes, I think it's related