_multiprocessing leaks references when initialised multiple times

[kumaraditya303]

By converting _multiprocessing to heap types, the module can be used in subinterpreters and does not leak memory when Python finalizes or when it is initialized multiple times. The converted heap types are not performance critical.

Current Memory leak at exit:

[170 refs, 71 blocks]

See PR #94336

[encukou]

It seems the leak is with SEM_VALUE_MAX, a static integer?
That could be solved by a PyGetSetDef getter.

[erlend-aasland]

Repeating #94336 (comment): AFAICS, the multiprocessing types do not access global module state, so there should be no reason to convert them to heap types.

[kumaraditya303]

It seems the leak is with SEM_VALUE_MAX, a static integer?

Where is that declared static? I don't see how it is considered a static integer. I am happy to be wrong though.

[erlend-aasland]

I changed the issue title. It seems to me that the issue is a ref leak. The proposed solution is converting types to heap types. IMO, the GitHub Issue title should reflect the problem, not necessary the solution.

[kumaraditya303]

It seems to me that the issue is a ref leak.

Do you have a fix that does not use heap types or private APIs but fixes the issue?

[encukou]

Ah, terminology. It's not C static, but it's always the same, and it's currently meant to be initialized only once (and it's not decref'd, so it leaks). You can't change the value from Python.

Do you have a fix that does not use heap types or private APIs but fixes the issue?

I can send one later today, if nothing comes up.

[kumaraditya303]

I can send one later today, if nothing comes up.

Sure, please add me as a reviewer, I'll learn something new.

[encukou]

OK, I see now that the leak is not caused by SEM_VALUE_MAX, but (mostly?) by the method/getter/setter wrappers set by PyType_Ready.
It seems to me that these are harmless: they won't be created a second time when _multiprocessing is imported again, and they can be reused in other interpreters. OTOH, they are flagged by -X showrefcount, and it would be nice to get those numbers to zero so -X showrefcount is useful in flagging new issues.

@vstinner, are you aware of this issue with -X showrefcount and static types with methods/getters/setters?