Debug build assertion failure with native threads attempting to acquire GIL on termination

[arielb1]

Bug description:

I'm writing PyO3/pyo3#4874, which tries to avoid Rust crashing on Python interpreter termination when there are native threads attempting to acquire the GIL.

To test it, I created a test that constantly hammers the GIL on a daemon thread, and on debug builds, I get this assertion failure fairly reliably on Python 3.13:

Python/pystate.c:345: void unbind_gilstate_tstate(PyThreadState *): Assertion `tstate == tstate_tss_get(&(tstate->interp->runtime)->autoTSSkey)' failed

It looks like zapthreads is attempting to zap a native thread that does not currently hold the GIL.

Would it help if I'll reproduce this in a C example?

CPython versions tested on:

3.13

Operating systems tested on:

Linux

[arielb1]

So I think the GIL-hammering thread is adding itself to the list of threads after the call to _PyThreadState_RemoveExcept. I'll stand up an Ubuntu with rr to confirm.

[the test is not really simulating a production workload I have, it's just hammering the GIL to make sure that Python + Rust is stable in that case].

[eendebakpt]

@arielb1 A minimal example reproducing the issue (in python or C) would be helpful. And would it be possible for you to test on the current main branch?

[ZeroIntensity]

How are you hammering the GIL? If it's via PyGILState_Ensure, then that's a known issue, but there's nothing we can do to fix it. There's a new API being discussed for safely acquiring the GIL (or really, attaching a thread state) at finalization.

[arielb1]

Yea, basically calling PyGILState_Ensure in a loop. The Rust code has an exception handler that turns the pthread_exit into a hang, but the tstate issue is separate AFAICT.

[ZeroIntensity]

The Rust code has an exception handler that turns the pthread_exit into a hang, but the tstate issue is separate

There wouldn't be any hang or exit (that would happen at somewhere like a Py_END_ALLOW_THREADS when a thread state tries to get attached). The problem is that PyGILState_Ensure has no handling at all for when Python is shutting down; it just crashes, like you're seeing here.

[arielb1]

It’s not crashing, it calls pthread_exit which my handler converts to a hang. It’s the main thread that’s crashing.

[ZeroIntensity]

It's probably because PyGILState_Ensure is setting some value to tell the interpreter that it exists (but sending your code would be helpful), then the thread gets hung, and then the main thread crashes while trying to destroy it because it's only partially initialized.

Bottom line is that you can't call PyGILState_Ensure while finalizing, in any way shape or form :(

[arielb1]

This sounds like a fixable race even with the current API - have finalization prevent threads from inserting themselves into the tstate list before emptying the tstate list.

[ZeroIntensity]

Yeah, and it will probably be fixed with the new API. But PyGILState_Ensure has several other issues anyway. I'm in favor of completely deprecating it once we get something better in place.

[arielb1]

But existing extensions will keep using it for a long time, I personally believe it should try to be safe, and safely hanging should be safe here.

If it's not safe, then every Python program using native code in a daemon thread has a chance of crashing on exit.

[arielb1]

So I think (but have no confirmation) that at this code section

cpython/Python/pystate.c

Lines 1593 to 1605 in 98fa4a4

	/* We serialize concurrent creation to protect global state. */
	HEAD_LOCK(interp->runtime);
	// Initialize the new thread state.
	interp->threads.next_unique_id += 1;
	uint64_t id = interp->threads.next_unique_id;
	init_threadstate(tstate, interp, id, whence);
	// Add the new thread state to the interpreter.
	PyThreadState *old_head = interp->threads.head;
	add_threadstate(interp, (PyThreadState *)tstate, old_head);
	HEAD_UNLOCK(interp->runtime);

It needs to detect that the interpreter is shutting down, release runtime_lock, and do something that triggers an hang/pthread_exit,

Of course, this won't help existing Python versions