> Adam, did you notice the change on revision 59231 ? the
> PyGILState_Ensure stuff should now remain valid during the
> PyInterpreterState_Clear() call.

That doesn't matter.  PyGILState_Ensure needs to remain valid *forever*.
 Only once the process is completely gone can we be sure it won't be called.

Note that PyGILState_Ensure has two behaviours: it can be called when
your thread is already running python code and has a tstate, or it can
be called when you have none.  revision 59231 fixed the former, but only
when Py_Finalize hasn't finished.  It doesn't fix it for any daemon
threads, or for dummy threads (created outside of python's control).

...

You're right, I did forget the 3 other places that acquire the
interpreter_lock.  The more I think about it the more I think our two
approaches are equivalent, but conditionalizing the release means we
don't need to touch the broken PyThread_exit_thread functions.

I think the banning should be as early as possible, right after
call_sys_exitfunc() has finished.  You can't do anything sane once the
interpreter is half-gone (despite the precedent of trying anyway.)