[BUG]: PyCapsule_SetName segfault

[tylerjereddy]

Required prerequisites

• Make sure you've read the documentation. Your issue may be addressed there.
• Search the issue tracker and Discussions to verify that this hasn't already been reported. +1 or comment there if it has.
• Consider asking first in the Gitter chat room or in a Discussion.

What version (or hash if on master) of pybind11 are you using?

2.10.2

Problem description

A detailed analysis of the situation that caused a hold-up of the SciPy release process on Windows is in this issue: scipy/scipy#17644 (comment)

Our testsuite can't even start with pybind11 2.10.2 -- it literally segfaults during the pytest collection phase, and more specifically import scipy.stats segfaults, and scipy.fft import will fail as desribed here: scipy/scipy#17644 (comment)

Reproducible example code

Unfortunately, I probably don't have the bandwidth to excise an isolated reproducer, but I bet the pybind11 folks will benefit from my DLL analysis over here: scipy/scipy#17644 (comment)

Is this a regression? Put the last known working version here if it is.

2.10.1

[matthew-brett]

For convenience, here's some information ported over from scipy/scipy#17644 (comment) and similar.

Scipy has a compiled module scipy.fft.pypocketfft. This module compiles and runs fine on all platforms with Pybind 2.10.1.

With Pybind 2.10.2, on Windows only, the compiled pypocketfft module segfaults on import during test collection.

Tyler's investigation revealed this was due a new missing dependency. Direct import generates the following:

>>> import scipy.fft
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "C:\Users\treddy\python310_venv_windows\lib\site-packages\scipy\fft\__init__.py", line 92, in <module>
    from ._helper import next_fast_len
  File "C:\Users\treddy\python310_venv_windows\lib\site-packages\scipy\fft\_helper.py", line 3, in <module>
    from ._pocketfft import helper as _helper
  File "C:\Users\treddy\python310_venv_windows\lib\site-packages\scipy\fft\_pocketfft\__init__.py", line 3, in <module>
    from .basic import *
  File "C:\Users\treddy\python310_venv_windows\lib\site-packages\scipy\fft\_pocketfft\basic.py", line 6, in <module>
    from . import pypocketfft as pfft
ImportError: DLL load failed while importing pypocketfft: A dynamic link library (DLL) initialization routine failed.

Further investigation revealed that there was a difference in declared imports between Pybind 2.10.1 and 2.10.2 - the 2.10.2 binary no longer imported Function PyCapsule_SetName (with Import from module python310.dll).

@rgommers wondered whether this might be due to #4254 .

[rwgk]

I left a comment here: scipy/scipy#17644 (comment)

[tylerjereddy]

I excised the reproducer from SciPy so that you now have a standalone project with a single C++ source file, 1 C++ header file, and a build file: https://github.com/tylerjereddy/pybind_repro

The README there has a short description of reproducing on 2.10.2 vs. success with 2.10.1. It could probably be cut down to an even simpler reproducer that is test-suite appropriate, but that probably gets into the weeds of C++ that are better dealt with by pybind11 team.

At the very least, this should completely uncouple your debugging cycles from SciPy, which is going to save a ton of time most likely.

[EthanSteinberg]

This turned out to not be a pybind11 issue as it was a bug within scipy. See scipy/scipy#17662.

We should fix have better error logging within pybind11 though. See #4426

[rgommers]

Thanks a lot for getting to the bottom of this @lalaland!

This turned out to not be a pybind11 issue

Given that this went into a bug fix rather than a minor release, and caused issues in SciPy and PyTorch already, it may still be wise to yank this release? You can argue that going from "working in practice but UB" to "hard to debug segfault" is not a pybind11 bug technically, but it's still pretty impactful. It even defeated SciPy's fairly careful pinning for the last already released version ("pybind11>=2.4.3,<2.11.0", here). Yanking is a cheap safety measure here.

[EthanSteinberg]

@rgommers Yeah, I think I agree with you. This is not acceptable for a bug fix change and it should probably be reverted and added back with better error reporting in the next minor release. Super sorry for all the work this caused on your part (especially during this time of the year).

[rwgk]

I'm sorry, too, that the failure reporting is so unobvious on some platforms.

But what is "yanking"?

I was thinking we make a 2.10.3 release with the better error reporting asap. Should we do something else?

[rwgk]

I forgot to add: only debug builds are affected.

[matthew-brett]

Yanking is a process you can do on Pypi to remove the release so it does not appear as a candidate for installation.

[rwgk]

Thanks! Sounds great to me. @henryiii do you know how to do that?

…

On Wed, Dec 28, 2022 at 22:03 Matthew Brett ***@***.***> wrote: Yanking is a process you can do on Pypi to remove the release so it does not appear as a candidate for installation. — Reply to this email directly, view it on GitHub <#4420 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAFUZACOV4ZBGAOIMMXIYZ3WPRJFNANCNFSM6AAAAAATISH6VI> . You are receiving this because you commented.Message ID: ***@***.***>

[henryiii]

Okay, I've yanked it. Let's get a 2.10.3 out ASAP since anyone relying on the Windows ARM support added in 2.10.2 is now broken (unless they request >=2.10.2, which is not likely just to add a single modern platform).

[henryiii]

From what I understand, we revert the UB -> error behavior for 2.10.3, and make it an error with better error reporting in 2.11?