The following patch seems to fix the issue explained in http://mail.python.org/pipermail/python-dev/2012-June/120659.html :


diff --git a/Objects/typeobject.c b/Objects/typeobject.c
--- a/Objects/typeobject.c
+++ b/Objects/typeobject.c
@@ -2417,6 +2417,10 @@ PyType_FromSpec(PyType_Spec *spec)
     if (res->ht_type.tp_dictoffset) {
         res->ht_cached_keys = _PyDict_NewKeysForClass();
     }
+    if (res->ht_type.tp_dealloc == NULL) {
+        /* It's a heap type, so needs the heap types' dealloc */
+        res->ht_type.tp_dealloc = subtype_dealloc;
+    }
 
     if (PyType_Ready(&res->ht_type) < 0)
         goto fail;

That does look like it will fix the leak, but now I'm actually thinking there's more code from type_new that should also be executed in the PyType_FromSpec case.

I mean things like:
- ensuring __new__ is a static method
- ensuring the standard attribute lookup machinery is configured
- hooking up tp_as_number, tp_as_mapping, etc
- ensuring GC support is configured correctly

If that's all happening somehow, it could use a comment, because I certainly can't see it.

If not, we probably need to factor out some helper functions that type_new and PyType_FromSpec can both call to make sure everything is fully configured.

> - ensuring __new__ is a static method

This shouldn't be necessary. __new__ won't be a method at all,
and not even exist. Instead, a type may or may not fill the tp_new
slot.

> - ensuring the standard attribute lookup machinery is configured

This is what PyType_Ready does, no?

> - hooking up tp_as_number, tp_as_mapping, etc

This is indeed missing. Robin Schreiber is working on a patch.

> - ensuring GC support is configured correctly

This is the responsibility of the caller, as always with C types.

In any case, one issue at a time, please. This issues is about a reference leak.

You're right, I was confusing what happens automatically for classes defined in Python (i.e. the full treatment in type_new) vs those defined statically (i.e. just the parts in PyType_Ready).

Given that PyType_FromSpec doesn't currently support inheritance, providing a default tp_dealloc before the inherit_slots() call in PyType_Ready would work OK in the near term.

However, once inheritance support is added by #15146 then it would be wrong - the default slot entry would override an inherited one.

So, I think this adjustment actually needs to be handled in PyType_Ready, at some point after the inherit_slots() call.

Something like:

    /* Sanity check for tp_dealloc. */
    if ((type->tp_flags & Py_TPFLAGS_HEAPTYPE) &&
        (type->tp_dealloc == type_dealloc)) {
        /* Type has been declared as a heap type, but has inherited the
           default allocator. This can happen when using the limited API
           to dynamically create types.
         */
        type->tp_dealloc = subtype_dealloc;
    }

> However, once inheritance support is added by #15146 then it would be
> wrong - the default slot entry would override an inherited one.

It would not be wrong. subtype_dealloc will properly call a base class'
tp_dealloc, if necessary.

True, I didn't follow the bouncing ball far enough. In that, case I think all that is needed is a comment like:

"subtype_dealloc walks the MRO to call the base dealloc function, so it is OK to block inheritance of the slot"

New changeset 1794308c1ea7 by Antoine Pitrou in branch '3.2':
Issue #15142: Fix reference leak when deallocating instances of types created using PyType_FromSpec().
http://hg.python.org/cpython/rev/1794308c1ea7

New changeset 9945d7dfa72c by Antoine Pitrou in branch 'default':
Issue #15142: Fix reference leak when deallocating instances of types created using PyType_FromSpec().
http://hg.python.org/cpython/rev/9945d7dfa72c

Ok, fixed, thanks.