In the C API, we have some functions that need to access state that is deleted by finalization. For example, PyGILState_Ensure accesses the gilstate, and that can be concurrently deallocated. In that specific case, I proposed #132510 as a fix, but really, we should have a better way to protect state against a finalizing interpreter. I'm not sure whether to categorize this as a bug or a feature.

I think the best way to do this is through a finalizing counter similar to how PEP-788 proposes preventing interpreter shutdown. So, for functions that don't require an attached thread state, we need something like:

int
PySomething_Whatever(PyInterpreterState *interp)
{
    if (_PyInterpreterState_CriticalSection(interp) < 0) {
        // Interpreter has finalized
        return -1;
    }

    /* Do critical operation */

    _PyInterpreterState_EndCriticalSection(interp);
    return 0;
}

Does this approach seem reasonable to everyone?

cc @ericsnowcurrently