bpo-46417: Finalize structseq types at exit

[vstinner]

Add _PyStructSequence_FiniType() and _PyStaticType_Dealloc()
functions to finalize a structseq static type in Py_Finalize().
Currrently, these functions do nothing if Python is built in release
mode.

Clear static types:

• AsyncGenHooksType: sys.set_asyncgen_hooks()
• FlagsType: sys.flags
• FloatInfoType: sys.float_info
• Hash_InfoType: sys.hash_info
• Int_InfoType: sys.int_info
• ThreadInfoType: sys.thread_info
• UnraisableHookArgsType: sys.unraisablehook
• VersionInfoType: sys.version
• WindowsVersionType: sys.getwindowsversion()

https://bugs.python.org/issue46417

[vstinner]

There are a few more structseq types:

• Modules/_cursesmodule.c: NcursesVersionType
• Modules/_threadmodule.c: ExceptHookArgsType
• Modules/signalmodule.c: SiginfoType
• Modules/timemodule.c: StructTimeType
• Objects/longobject.c: Int_InfoType <= DONE
• Python/errors.c: UnraisableHookArgsType <= DONE
• Python/sysmodule.c: FlagsType <= DONE
• Python/sysmodule.c: VersionInfoType <= DONE
• Python/sysmodule.c: WindowsVersionType <= DONE
• Python/thread.c: ThreadInfoType <= DONE

[vstinner]

cc @erlend-aasland @corona10

[vstinner]

I changed my mind and also enabled the code on Python release build. Previously, I only enabled the code if the Py_DEBUG macro was defined. For example, the Windows CI job on pull requests builds Python in release mode. If something is wrong, I prefer to catch the bug early and fix it.

[vstinner]

The PR is now ready for a review :-)

[shihai1991]

OK. This PR looks good to me :)
But I think another question: Can we add a global cache to maintain the StructSequenceType and don't call the dealloc function in finalize_interp_types() manually?

[vstinner]

Can we add a global cache to maintain the StructSequenceType and don't call the dealloc function in finalize_interp_types() manually?

Do you mean to create a list of structseq types when _PyStructSequence_InitType() or _PyStructSequence_InitType2() is called? Yeah, I had this idea, but I am not sure about it.

A problem is to control exactly when these types are deleted and types them in a specific order. The Python finalization is quite complex, see my notes about past issues: https://pythondev.readthedocs.io/finalization.html

assert(Py_REFCNT(type) == 1); of _PyStructSequence_FiniType() fails if something still keeps the a reference to the type somewhere: if the type is deallocated "too early".

[shihai1991]

Can we add a global cache to maintain the StructSequenceType and don't call the dealloc function in finalize_interp_types() manually?

Do you mean to create a list of structseq types when _PyStructSequence_InitType() or _PyStructSequence_InitType2() is called? Yeah, I had this idea, but I am not sure about it.

Yes. That's what I mean : )

A problem is to control exactly when these types are deleted and types them in a specific order. The Python finalization is quite complex, see my notes about past issues: https://pythondev.readthedocs.io/finalization.html

assert(Py_REFCNT(type) == 1); of _PyStructSequence_FiniType() fails if something still keeps the a reference to the type somewhere: if the type is deallocated "too early".

OK, Thanks for your information. I will study your notes.

[corona10]

I am +1 on this change :)

but same as @encukou, we may need tests to sure if the types are correctly reinitialized if there are multiple finalize/init cycles

[vstinner]

@corona10:

but same as @encukou, we may need tests to sure if the types are correctly reinitialized if there are multiple finalize/init cycles

I added tests. Would you mind to review these tests as well?

[vstinner]

I enhanced the tests to check more PyTypeObject members.