bpo-46070: Revert "bpo-36854: Move _PyRuntimeState.gc to PyInterpreterState (GH-17287)"

[vstinner]

This reverts commit 7247407.

Fix a random crash involving subinterpreters on Windows. Revert the
change which made the gc module state per interpreter: the gc module
state is shared again by all interpreters.

https://bugs.python.org/issue36854

[ericsnowcurrently]

LGTM

😢

[ericsnowcurrently]

FTR, this is to resolve https://bugs.python.org/issue46070.

[ericsnowcurrently]

BTW, do we have any sort of reproducer for which we can add a test? That way we don't cause the same problem when we move GC back to PyInterpreterState later.

[vstinner]

BTW, do we have any sort of reproducer for which we can add a test?

So far, I was only able to reproduce the crash on Windows with Python 3.9 using win_py399_crash_reproducer.py script of https://bugs.python.org/issue46070 I failed to write a reliable reproducer which works on any platform.

I didn't fully understand the root issue. My theory is that an object is shared by two interprereter. The first interpreter runs a GC collection which traverses this shared object, an object finalizer releases the GIL. Now the second interpreter gets the GIL and also runs a GC collection which traverses the same shared object.

The problem is that the PyGC_Head header cannot be modified by two interpreters in parallel: data races are likely and Python does crash somehow.

Since the bug is a crash and so far nobody managed to fully understand it, IMO it's safer to revert the change for now.

[ericsnowcurrently]

Your theory sounds plausible.

[vstinner]

I wrote a simpler fix: GH-30577.