sub-interpreters: datetime segfaults if imported from two sub-interpreters

[costasgambit]

Bug report

Bug description:

I've started putting a branch together that moves all the global state into a module-level state object.

If that looks useful, I'll continue to work on it and can rebase on main.

It passes tests and "works on my machine".

I haven't looked at python internals in some time, so apologies if I'm a little rusty!

CPython versions tested on:

3.12, CPython main branch

Operating systems tested on:

Linux

Linked PRs

• gh-110415: move datetime global state into module state #110420

[gaogaotiantian]

Sorry but this does not look like a bug report. The content of the issue is a bit unrelated to the title.

Could you paste your code in a minimum reproducible example so we can track it down?

[costasgambit]

Sorry @gaogaotiantian, here is a bit of code that should demonstrate the segfault on 3.12+:

import _xxsubinterpreters

interp1 = _xxsubinterpreters.create()
interp2 = _xxsubinterpreters.create()

_xxsubinterpreters.run_string(interp1, "import datetime")
_xxsubinterpreters.run_string(interp2, "import datetime")   # crash

[gaogaotiantian]

I'm not able to repro it with 3.12.0 + WSL(Ubuntu 20.04). Are you using the official release? What's your python -V?

[gaogaotiantian]

I am able to repro it on main with --with-pydebug.

[costasgambit]

Python 3.12.0+, but main is the same.

If you scan over _datetime_exec() you'll see there's global state everywhere.

[erlend-aasland]

The datetime module is not fully isolated. I started working on it last year, but I took a break from module isolation work around the 3.12 feature freeze, and I haven't found the energy needed to restart that work. You can take a look at the existing issue/PR:

• gh-71587: Isolate _datetime #102995
• Datetime NoneType after calling Py_Finalize and Py_Initialize #71587

[erlend-aasland]

Duplicate of #71587

[erlend-aasland]

The major pain point of isolating _datetime is that it has an exposed C API (though datetime.h, which must be explicitly included). I discussed this with @pganssle at the Language Summit and it was not immediately obvious if the datetime.h C API is bound by PEP-387 in the same way API exposed by Python.h is. This point must be resolved before the _datetime module can be completely isolated.