Type object's ob_type does not get set when tp_bases is set before PyType_Ready

[lysnikolaou]

Bug report

This came up while testing numpy on 3.12. A bug report has been filed on numpy/numpy#23766. It's happening due to #103912, which introduced an additional check for tp_bases not being NULL in type_ready_set_bases, which is called from PyType_Ready.

numpy sets tp_bases manually before calling PyType_Ready, which means that the afore-mentioned check succeeds, and so, the line that sets ob_type does not get executed (it used to before #103912), which leads to a segmentation fault later on, when trying to set mro.

This looks like a bug, but I'm not sure whether that's expected and numpy should be adjusted. If the latter is true, should a note be added in the What's new document?

Linked PRs

• gh-104614: Make Sure ob_type is Always Set Correctly by PyType_Ready() #105122
• [3.12] gh-104614: Make Sure ob_type is Always Set Correctly by PyType_Ready() (gh-105122) #105211
• gh-104614: Fix potential ref. leak in _testcapimodule/get_basic_static_type() #105225
• [3.12] gh-104614: Fix potential ref. leak in _testcapimodule/get_basic_static_type() (GH-105225) #105248

[lysnikolaou]

cc @ericsnowcurrently as the author of #103912.

[lysnikolaou]

Verified that the following diff solves this issue, and the test suite passes. (Still not sure whether it has any unintended side-effects though)

diff --git a/Objects/typeobject.c b/Objects/typeobject.c
index 2fbcafe91a..495e123f8f 100644
--- a/Objects/typeobject.c
+++ b/Objects/typeobject.c
@@ -6990,10 +6990,6 @@ type_ready_pre_checks(PyTypeObject *type)
 static int
 type_ready_set_bases(PyTypeObject *type)
 {
-    if (lookup_tp_bases(type) != NULL) {
-        return 0;
-    }
-
     /* Initialize tp_base (defaults to BaseObject unless that's us) */
     PyTypeObject *base = type->tp_base;
     if (base == NULL && type != &PyBaseObject_Type) {

[ericsnowcurrently]

I should have a fix up shortly.

[ericsnowcurrently]

@lysnikolaou, please verify that gh-105122 fixes the problem. If not, what would be a good way to reproduce the problem?

[ericsnowcurrently]

main and 3.12 should be good now.