gh-123961: move PyCursesError in _cursesmodule.c to a global state variable

[picnixz]

UPDATE: managed to make it smaller.

---

Previous discussion (now resolved)

Ideally, I wanted to make multiple PRs for the curses module but simply adding a module state with the minimal logic would not be helpful at all. So I decided to do the following in one go:

• Add a module's state that holds the curses exception type and the curses window object type.
• Make the window object type a heap type.
• Use multi-phase initialization for the module.

The reason I couldn't find a way to split those changes is due to the exposed C API. The C API uses parameter-less functions and thus does not have access to the module's state. More precisely, we need to access the exception type in three different contexts without relying on a global variable (the purpose is to remove the global variables):

• In a parameter-less function: we need to re-import the curses module.
• In a window object's method: if we use a static type, we still can't access the module's state so it does not makes sense to me to make a change where I would first import the exception and in the next PR I would access it directly from the module's state.
• In a module's method: we can avoid using the state since the exception is stored in the module's dict, but again, this does not make sense to keep a temporary workaround for a few days.

---

There is a way to split the PR into two, namely by getting PyCursesError via an import. Informally, we would consider the module's dict as the state (which we must do for parameter-less functions) though I honestly think it's not worth that effort for window and module's methods.

Alternatively, we could first make the module a multi-phase initialization module, without any state and keep the global variables. However, I'm not sure that's a good idea. Note that I cleanly separated the commits so that I can pick them up if needed.

Finally, I did not apply any cosmetic changes even though I was very tempted. The only cosmetic change I did was for the macros so that they readability is improved and if I'm anyway editing the lines. If a cosmetic change slipped through, please tell me.

• Issue: Refactorization of _cursesmodule.c #123961

[picnixz]

I don't know whether I should include a NEWS entry or not. One entry could be that I actually fixed the reference leaks in the curses module. However, I want to know that what I did would not cause an issue with the exported C API (I don't want a segfault for this one and I would like to be sure that what needs to be process-wide is indeed process-wide and not stored in the module's state).

[bedevere-bot]

🤖 New build scheduled with the buildbot fleet by @picnixz for commit 228348c 🤖

If you want to schedule another build, you need to add the 🔨 test-with-refleak-buildbots label again.

[picnixz]

Ok, so test_class leaks (#124722).

[picnixz]

Note: for now, I only used the get_cursesmodule_state function in the module's methods but not in the methods for window objects. Windows objects directly access the global state (this is just another global variable) but in the future, when the type will be a heap type, I'll be able to access the underlying module (and convert them to get_cursesmodule_state_by_cls calls).

[picnixz]

If I were to also change PyCursesCheckERR so that it accepts a module / state argument, the PR will be larger. Do you want me to that change?

Ideally, PyCursesCheckERR(code, name) should be changed into PyCursesCheckERR(module, code, name) and PyCursesCheckERR_ForWin(window, code, name) so that we only retrieve the state when needed.

This will also require an additional helper:

static inline _cursesmodule_state *
get_cursesmodule_state_by_cls(PyObject *Py_UNUSED(cls))
{
    return &curses_global_state;
}

[picnixz]

I don't think I can use more the state than what is being currently used.

[vstinner]

You may rename st to state.

[vstinner]

LGTM