The example object in the xxlimited module can be part of a reference loop (it can contain itself), so it needs GC and tp_traverse.The tp_dealloc hook was incorrect, and a correct version would be difficult to generalize for something more complicated than a example class (#16690; [0]). It's better to avoid dealloc and show off tp_finalize.Same for the class in _testmultiphase (PEP 489 tests), which is based on xxlimited. The incorrect dealloc is causing the reference leak mentioned in#24268.The Xxo object also wasn't actually added to the module.Here is a patch to fix these.[0]https://mail.python.org/pipermail/python-dev/2015-June/140422.html

Would it also be worth making at docs update to tp_dealloc, suggesting the use of tp_traverse/finalize?:https://docs.python.org/3/c-api/typeobj.html#c.PyTypeObject.tp_deallocAnd perhaps from PyType_FromSpec?https://docs.python.org/3/c-api/type.html?highlight=pytype_fromspec#c.PyType_FromSpec

tp_traverse is completely orthogonal to tp_dealloc, it's needed to detect (and then break) reference cycles like:    obj = xxlimited.Xxo()    obj.foo = objAs for tp_finalize: yes, mentioning it in tp_dealloc docs would be good, but I'll need a bit more studying to understand the problem correctly. The cases fixed here are relatively simple; Antoine gives more complex ones in [0]. When I feel qualified to give advice, I'll change the docs. (And most likely, write a PEP to make things easier; some changes to classes will be needed anyway to makePEP 489 multi-phase init work nicely in all cases).But, I plan to focus my CPython time on documentingPEP 489 before diving in here. I thinkissue 16690 is a good place to track tp_dealloc docs changes.[0]https://mail.python.org/pipermail/python-dev/2015-June/140423.html

New changeset265eeb60443a by Nick Coghlan in branch '3.5':Issue#24373: EliminatePEP 489 test refleakshttps://hg.python.org/cpython/rev/265eeb60443aNew changesetf24cd8bc5250 by Nick Coghlan in branch 'default':Merge fix for issue#24373 from 3.5https://hg.python.org/cpython/rev/f24cd8bc5250

tp_finalize handlers should return "void". Seeissue33644.