STINNER Victor <vstin...@python.org> added the comment:
In _abcmodule_exec(), when _abc_data type is created, it's created with refcnt=3: * 1 strong reference (normal) * +1 ref from tp_dict['__new__'] slot * +1 ref from tp_mro type_traverse() visits tp_dict and tp_mro, so it's fine. In Py_EndInterpreter(), PyInterpreterState_Clear() clears os.register_atfork() callbacks which were one of the last references to _abc_data type. The first call to _PyGC_CollectNoFail() destroys _abc_data *instances* but not the _abc_data type. The following patch works around the issue: diff --git a/Modules/_abc.c b/Modules/_abc.c index 1efc98bf72..410dbcf96a 100644 --- a/Modules/_abc.c +++ b/Modules/_abc.c @@ -54,6 +54,7 @@ typedef struct { static int abc_data_traverse(_abc_data *self, visitproc visit, void *arg) { + Py_VISIT(Py_TYPE(self)); Py_VISIT(self->_abc_registry); Py_VISIT(self->_abc_cache); Py_VISIT(self->_abc_negative_cache); $ ./python -m test -R 3:3 test_threading -m test.test_threading.SubinterpThreadingTests.test_threads_join_2 (...) Tests result: SUCCESS I'm not sure why Py_VISIT(Py_TYPE(self)) is needed. Maybe it's a regression caused by commit 364f0b0f19cc3f0d5e63f571ec9163cf41c62958 of bpo-35810. It sems like the GC doesn't take in account that instances of types allocated on the heap (if type->tp_flags has the Py_TPFLAGS_HEAPTYPE flag) hold a strong refeference to the type (PyObject.ob_type). I created bpo-40217: "The garbage collector doesn't take in account that objects of heap allocated types hold a strong reference to their type". ---------- _______________________________________ Python tracker <rep...@bugs.python.org> <https://bugs.python.org/issue40149> _______________________________________ _______________________________________________ Python-bugs-list mailing list Unsubscribe: https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com