Changes proposed in PEP3121 and PEP384 have now been applied to the dbm module!

Robin, why do you increment refcounter for type object on every construction of dbm instance?

Also I would to see macros in uppercase if possible.

PEP384 demands, among other things, that the TypeObjects themselves are transformed to Heaptypes. This means that the Typeobjects, that are created from this new stable ABI, reside within the heap and therefore have to be managed by the Python GC. This is of course done by appropriately incref- and decrefing of the specific Heaptype.

From my perspective type object is referenced by module state.
While module is present it's state hold all types exposed by module.
After module has deleted it's state has cleaned up and types has gone away — thats look good to me.

Instances of python classes (not extensions) hold ownership for it's class. I guess it done because python class can be constructed inside function and instance of this class should work after exit from that function like closure does.

In case of extension modules enough to hold type reference into module state itself.

Thoughts?

Robin, after thinking I would to agree with your decision to hold reference to type into type instance.

Please, can you describe your check like:

if((void *)type->tp_dealloc == (void *)dbm_dealloc) {
    Py_DECREF(type);
}

Why you decref only if type->tp_dealloc points to module's dealloc? What are you protected from? What other values also possible?

Take a look at the discussion in Issue 15653.

Fixed by:

commit bf69a8f99f1b0e19a59509c6c4d7015a31d881a1
Author: Dong-hee Na <donghee.na92@gmail.com>
Date:   Tue Jun 16 01:20:54 2020 +0900

    bpo-1635741: Port _dbm module to multiphase initialization (GH-20848)