[subinterpreters] Refleaks on Windows Under Specific Conditions

[ericsnowcurrently]

Bug report

Bug description:

One of my recent PRs, gh-117662, led to one of the Windows refleak buildbots failing¹. I'm fairly confident the PR revealed an existing source of refleaks (albeit an unlikely one), rather that introducing new leaks. I have resolved the failures with gh-117913, but the underlying potential source of refleaks remains.

I've been able to reproduce the leak conditions, but only on Windows (for now?). The following conditions are necessary:

• interpreter created using a config with check_multi_interp_extensions=False (e.g. "legacy")
• interpreter imports test.support.os_helper

try:
    import _interpreters
except ModuleNotFoundError:
    import _xxsubinterpreters as _interpreters
config = _interpreters.new_config('legacy')  # critically, check_multi_interp_extensions=False
interpid = _interpreters.create(config)
_interpreters.exec(interpid, 'import test.support.os_helper')
_interpreters.destroy(interpid)
# leaks a bunch of objects

When the bug is triggered, around 190 objects are leaked. At first I thought this was interpreter finalization failing silently, but now I think it is something else.

I suspect the underlying problem relates to legacy extension modules. In the specific example above, I'm pretty sure the _ctypes module is leaking and ~190 is how many objects it holds, directly or indirectly. The main clue there is that, when I run ./python -v ..., the _ctypes module is never noted as destroyed, whereas all other modules are. As to test.support.os_helper, _ctypes is one of the modules that gets indirectly imported. All that said, I haven't been able to reproduce the leak if the subinterpreter above imports _ctypes instead.

Why did we only see this on Windows? The only clue I can think of is that on Windows stdlib extension modules are all turned into builtin modules, IIRC. The logic of the check_multi_interp_extensions check (via _PyImport_CheckSubinterpIncompatibleExtensionAllowed() in Python/import.c) is specific to extension modules, not builtin modules.

CPython versions tested on:

CPython main branch

Operating systems tested on:

Windows

Footnotes

1. https://buildbot.python.org/all/#/builders/920/builds/733 ↩

[ericsnowcurrently]

FYI, gh-117487 (or something like it) may help.

[neonene]

The _ctypes module has switched to the multi-phase initialization, and I can confirm the leaks above disappear on Windows with a patch like #117181 (comment). I think we are on #117181 (comment) at this point.