[WIP] gh-151722: Defer GC tracking of frozendict.fromkeys() as possible

[corona10]

• Issue: Data race on frozendict reads during construction in the free-threading build #151722

[corona10]

cc @tonghuaroot (I add you as the co-author), too many things need to be addressed from #151967

[corona10]

If there is a good way to obtain an untracked object here, that would be great, but I think that is a separate topic.

[corona10]

cc @vstinner

[methane]

Is these cases considered?

• subclass of frozendict implements __init__ instead of __new__, passes its self to another thread
• another thread obtains a reference to the instance via GC while the __init__ is being executed

[corona10]

@tonghuaroot
By this, we don't need additional test code.

[tonghuaroot]

Agreed, the assert covers it. Thanks for taking this over.

[tonghuaroot]

Thanks for driving this. I believe the test_fromkeys CI failure originates here: for an empty exact frozendict, cls() already returns an untracked object (thanks to your gh-151740 defer-track), so this unconditional _PyObject_GC_UNTRACK trips its internal IS_TRACKED assertion. Guarding it with if (_PyObject_GC_IS_TRACKED(d)) should resolve it. Please feel free to adjust as you see fit.

[vstinner]

tl; dr I wrote PR gh-152067 which is a similar fix, but with a safer design.

I don't think that calling _PyObject_GC_UNTRACK(d) on any type is a good approach. We need to track again the dictionary in many code paths which is very tricky. For example, if dict_dict_fromkeys() fails (return NULL), we need to track again the dictionary... but we no longer have a (safe) reference to d since dict_dict_fromkeys() called Py_DECREF(d). Having to track again the dictionary on error is really complicated to implement.

Moreover, we cannot use Py_INCREF(d) at the beginning (to keep a safe reference to the dictionary) since assert(can_modify_dict(mp)); fails in this case on frozendict.

IMO we should only be worried about the dictionary being not tracked by the GC if the dictionary is a frozendict. And it's a frozendict, we can copy it using frozendict_new_untracked() to create a copy which is not tracked by the GC.