multiprocessing.c currently has code like this:

	temp = PyDict_New();
	if (!temp)
		return;
	if (PyModule_AddObject(module, "flags", temp) < 0)
		return;

PyModule_AddObject consumes the reference to temp, so it could
conceivable be deleted before the rest of this function finishes.

I believe this patch solves the problem. 
I added the line
Py_DECREF(temp)
after the code block shown above. 
I also changed the line
if (PyDict_SetItemString(temp, #name, Py_BuildValue("i", name)) < 0) return
to
if (PyDict_SetItemString(PyObject_GetAttrString(module, "flags"), #name,
Py_BuildValue("i", name)) < 0) return

This doesn't look right.  PyDict_SetItemString doesn't steal the
references passed to it, so your reference to flags will be leaked each
time.  Besides, I think it's a little cleaner to INCREF it before call
PyModule_AddObject, then DECREF it at any point you return.

Additionally, I've just noticed that the result of Py_BuildValue is
getting leaked.  It should be stored to a temporary, added to flags,
then the temporary should be released.

I think the easiest way is to just call AddObject after inserting the
values.

I fixed this, and the leaking BuildValue references, in r64223.

Aww, that's cheating.  (Why didn't I think of that?)