gh-124715: Fix method_dealloc(): use PyObject_GC_UnTrack()

[vstinner]

Replace _PyObject_GC_UNTRACK() with PyObject_GC_UnTrack() to not fail if the method was already untracked.

• Issue: Not all types have "trashcan" protection and tp_dealloc can overflow stack #124715

[vstinner]

This change fix test_functools on s390x: #132280 (comment)

But I'm not sure why "Move trashcan mechanism into Py_Dealloc" requires this change.

cc @iritkatriel@markshannon@nascheme

[markshannon]

But I'm not sure why "Move trashcan mechanism into Py_Dealloc" requires this change.

I don't see any reason either.
Is it just s390x showing this failure?

[vstinner]

Is it just s390x showing this failure?

s390x buildbot is fast. Other buildbots are still running. Maybe test_functools will fail on other buildbots, I don't know so far.

[vstinner]

Oh, I cannot reproduce the issue on s390x RHEL9 (GCC 11.5.0) nor on s390x Fedora Stable (GCC 14.2.1), only on s390x RHEL8 (GCC 8.5.0).

[vstinner]

s390x buildbot is fast. Other buildbots are still running. Maybe test_functools will fail on other buildbots, I don't know so far.

test_functools does also crash on:

• aarch64 Fedora Stable LTO 3.x: https://buildbot.python.org/#/builders/336/builds/6843
• aarch64 Fedora Stable Clang 3.x: https://buildbot.python.org/#/builders/234/builds/7548
• aarch64 Fedora Stable Clang Installed 3.x: https://buildbot.python.org/#/builders/14/builds/7693
• aarch64 RHEL8 LTO + PGO 3.x: https://buildbot.python.org/#/builders/78/builds/9032
• AMD64 Fedora Stable Clang 3.x: https://buildbot.python.org/#/builders/441/builds/7851
• s390x Fedora Stable LTO 3.x: https://buildbot.python.org/#/builders/1654/builds/316

[vstinner]

AMD64 Fedora Stable Clang 3.x: https://buildbot.python.org/#/builders/441/builds/7851

I confirm that test_functools does crash randomly when I build Python using clang on my laptop:

$ ./configure --with-pydebug CC=clang LD=clang $ make $ ./python -m test -v test_functools (...) test_recursive_pickle (test.test_functools.TestPartialPy.test_recursive_pickle) ... Objects/classobject.c:247: _PyObject_GC_UNTRACK: Assertion "_PyObject_GC_IS_TRACKED(((PyObject*)(op)))" failed: object not tracked by the garbage collector 

[vstinner]

It seems like methods are special and can be untracked by something that I failed to identify when they enter the trashcan mechanism. I hope that only methods are treated in a special way and that other objects don't have to be modified for the recent trashcan change.

[markshannon]

Does this fix all the buildbots?
If so, let's merge this but keep the issue open so we don't forget to investigate further.

[vstinner]

See also #133261 : test_frame crash randomly on RHEL8, but this time on x86-64.

[vstinner]

Does this fix all the buildbots?

Yes. Buildbots which are back to green:

• s390x RHEL8 3.x
• aarch64 Fedora Stable LTO 3.x
• aarch64 Fedora Stable Clang 3.x
• aarch64 Fedora Stable Clang Installed 3.x
• aarch64 RHEL8 LTO + PGO 3.x
• AMD64 Fedora Stable Clang 3.x

If so, let's merge this but keep the issue open so we don't forget to investigate further.

#124715 is still open.