Attached is a very short patch against r59568 which asserts tp_traverse
on (the types of) objects allocated in PyType_GenericAlloc(). As far as
I'm aware, tp_traverse should always be set at this point. Catching that
error early, even if only in debug builds, would help to prevent bugs
like http://bugzilla.gnome.org/show_bug.cgi?id=504337 .

Bryan: can you provide a test case without external dependencies? If
not, confirming this is still valid in 2.6 would also help.

Well, there's no Python bug per se, hence no test case; this patch just
adds a single additional assert that might catch a particular extension
implementation mistake. It was prompted by tracking down the bug in
pygtk mentioned above.

I've attached an updated patch against r72055. It's a trivial change,
but I would suggest that someone more familiar with the Python core sign
off on it regardless.

It would be probably be better to put this check in PyType_Ready() instead.

Are you sure this assertion is correct? tp_traverse is optional.

I submitted this patch 14 years ago and am sure of nothing. :)

tp_traverse is optional, so we should not add this assertion.

This issue was fixed differently in bpo-44263, by adding a test in PyType_Ready():

    // bpo-44263: tp_traverse is required if Py_TPFLAGS_HAVE_GC is set.
    // Note: tp_clear is optional.
    if (type->tp_flags & Py_TPFLAGS_HAVE_GC
        && type->tp_traverse == NULL)
    {
        PyErr_Format(PyExc_SystemError,
                     "type %s has the Py_TPFLAGS_HAVE_GC flag "
                     "but has no traverse function",
                     type->tp_name);
        return -1;
    }

commit ee7637596d8de25f54261bbeabc602d31e74f482
Author: Victor Stinner <vstinner@python.org>
Date:   Tue Jun 1 23:37:12 2021 +0200

    bpo-44263: Py_TPFLAGS_HAVE_GC requires tp_traverse (GH-26463)
    
    The PyType_Ready() function now raises an error if a type is defined
    with the Py_TPFLAGS_HAVE_GC flag set but has no traverse function
    (PyTypeObject.tp_traverse).

> tp_traverse is optional, so we should not add this assertion.

It is required by types implementing the GC protocol. This assumption is non-obvious and was not documented. The documentation was completed in bpo-44263 by commit 8b55bc3f93a655bc803bff79725d5fe3f124e2f0.