gh-123961: convert curses to a multi-phase initialization module (PEP-489)

[picnixz]

I think this is the very last PR for this task, conditioned to some final cleanups just for macro readability and comments.

• Issue: Refactorization of _cursesmodule.c to fix reference leaks #123961

[picnixz]

Following our discussion on the previous PR, I also changed the _cursesmodule prefix to cursesmodule so that we have at least a bit more of consistency. I won't touch the AC-generated functions.

[picnixz]

I'm really un-sure of how free-threaded builds with multi-phase initialization would interact with capsule objects. So I'd happy if any free-threaded expert could help me here. Here are some questions (and we may perhaps address them in separate PRs):

• Should the global flags (simple int) be atomically set?
• Should we use PyRawMem_* instead of PyMem_* functions when allocating the memory for capsule objects?
• Why am I only seeing the assertion failure on my pointer not NULL when I converted it into a multi-phase initialization module and not before? (see 9e0dd70 for the commit who actually "fixed" it).

cc @colesbury (I don't know other free-threaded experts but feel free to delegate the question to anyone who knows about the topic if you don't have enough time!)

[vstinner]

Can you try to add a global variable to prevent having more than 1 instance at the same time?

[picnixz]

Here's the output, do you think it's good enough?

>>> import sys
...
>>> import _curses as curses1
>>> del sys.modules['_curses']
...
>>> import _curses as curses2
...
Traceback (most recent call last):
  File "<python-input-3>", line 1, in <module>
    import _curses as curses2
ImportError: module 'curses' can only be loaded once per process

[vstinner]

cursesmodule_free() should set curses_module_loaded to 0. So you can load-unload the extension multiple times:

import sys
import gc

print("load-unload")
import _curses
del _curses
del sys.modules['_curses']
gc.collect()

print("load again")
import _curses
del _curses
del sys.modules['_curses']

[picnixz]

cursesmodule_free() should set curses_module_loaded to 0. So you can load-unload the extension multiple times:

I'm not sure it works. Even if I do this it still tells me that the curses module cannot be loaded more than once per process. The cursesmodule_free() does not seem to be called even if gc.collect() is called (at least not in the REPL). What should I do?

[vstinner]

The cursesmodule_free() does not seem to be called even if gc.collect() is called (at least not in the REPL).

It works if you run my script:

load-unload
load again

[picnixz]

After some tests, it appears that the REPL only calls free() upon exiting the process (the REPL imports curses in Lib/_pyrepl/curses.py). So, reloading curses inside the REPL won't work. In a normal script, however:

$ read -r -d '' code << EOM
import sys

print("load-unload")
import _curses
del _curses
del sys.modules['_curses']

print("load again")
import _curses
del _curses
del sys.modules['_curses']
EOM
$ ./python -c "$code"
load-unload
load again
Traceback (most recent call last):
  File "<string>", line 9, in <module>
    import _curses
ImportError: module 'curses' can only be loaded once per process

When free is called via gc.collect():

$ read -r -d '' code << EOM
import sys
import gc

print("load-unload")
import _curses
del _curses
del sys.modules['_curses']
gc.collect()

print("load again")
import _curses
del _curses
del sys.modules['_curses']
EOM
$ ./python -c "$code"
load-unload
load again

[vstinner]

Merged, thanks. That's a nice step forward for the _curses extension :-)

[vstinner]

Python no longer leaks at exit:

$ ./python -X showrefcount -c 'import _curses'
[0 refs, 0 blocks]

Python 3.13 for comparison:

$ ./python -X showrefcount -c 'import _curses'
[234 refs, 159 blocks]