Python <=3.9 segmentation fault in code snippet

[NullHypothesis]

Bug report

Bug description:

The following code snippet results in a segmentation fault in Python 3.8 and 3.9, but not 3.10 and newer:

#!/usr/bin/env python3.9

# pip install opentelemetry-sdk
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import (
    BatchSpanProcessor,
    ConsoleSpanExporter,
)

fd = open("foo.txt", "w")
tracer_provider = TracerProvider()
processor = BatchSpanProcessor(ConsoleSpanExporter(out=fd))
tracer_provider.add_span_processor(processor)

The following conditions need to be met for the segmentation fault to happen:

• Python 3.9.18 or older (does not happen with Python 3.10 or newer)
• fd must not be closed (does not happen if fd.close() is run at the end)

I could reproduce this segfault on both macOS Sonoma 14.3.1 and Ubuntu Linux 23.10.

I realize that 3.8 and 3.9 are no longer in the "bugfix" phase but the segfault may have security implications, which is why I decided to open this issue. Please feel free to close if it's not helpful.

CPython versions tested on:

3.8, 3.9, 3.10, 3.13

Operating systems tested on:

Linux, macOS

[gaogaotiantian]

Segfault is sometimes security issue, not always, but the more important thing is - can you repro this without opentelemetry? If it's a CPython issue then it should be reproducible without any 3rd party code.

[NullHypothesis]

I'm unable to spend more cycles on this but I'll create an OpenTelemetry issue. If that results in something useful, I'll re-open.

[methane]

I can reproduce it without otel. ref

import os, time

f = open("foo.txt", "w")

class C:
    def __init__(self):
        self.f = f
        os.register_at_fork(after_in_child=self.atfork)

    def atfork(self):
        print("atfork")
c=C()
del c, f

This issue caused by this code is called during late interpreter shutdown.

cpython/Python/_warnings.c

Line 916 in 1f8b24e

assert(globals != NULL);

This issue is fixed by this commit in Python 3.10. register_at_fork is cleared before sysdict.

#21605

But I am not sure about similar SEGV is not possible anymore.
ResourceWarning for unclosed files are happened during interpreter shutdown often.
So warning code would be more defensive. Instead of assert(global != NULL), it can be if (global == NULL) { goto handle_error; }

[methane]

I checked finalization code and I can not find any way to emit ResourceWarning after Py_CLEAR(interp->sysdict); in interpreter_clean().
So I think this issue is really fixed in Python 3.10+ by #21605 .