gh-150157: fix: critical section for PyDict_Next in _pickle.c

[KowalskiThomas]

Fixes #150157.

• Issue: Missing critical section for PyDict_Next in _pickle.c #150157

[serhiy-storchaka]

whichmodule is an internal detail not visible to users. Describe the cause in general words, like it looks from user's point.

[KowalskiThomas]

How does the following sound? (I pushed it for now.)

Fix a potential crash occurring when pickling objects
concurrently in free-threaded builds.

Tried to make it clear from and end-user perspective while not sounding too scary (since it's a rare race condition), hence the potential.

[serhiy-storchaka]

It's not just a matter of luck. There have to be several conditions for this to happen. First, we must pickle by name an object without the __module__ attribute. Second, sys.modules should be concurrently modified (in a specific way, but we will leave this).

[KowalskiThomas]

Okay, I feared this might be a little bit too detail-y. I'll rephrase it again 😅

[KowalskiThomas]

Updated to this.

Fix a crash in free-threaded builds that occurs when pickling
an object without a ``__module__`` attribute while :data:`sys.modules`
is concurrently being modified.

[serhiy-storchaka]

Technically, this is not correct, most objects do not have such issue even without having __module__, only objects pickled by name (some singletons, named constants or enum-like objects). But if you think that the current wording is enough, we can leave it.

[KowalskiThomas]

@serhiy-storchaka Right... Thanks for bearing with me; I've updated it again and I think it should be clear and accurate now 😅

Latest version is:

Fix a crash in free-threaded builds that occurs when pickling by name
objects without a ``__module__`` attribute while :data:`sys.modules`
is concurrently being modified.