Faulthandler segfaults when called from threads

[devdanzin]

Crash report

What happened?

In a GILfull build of main, it's rarely possible to get faulthandler to segfault when calling it from threads. Since it's a rare crash and I'm on mobile, no reduced test case is available at the moment.

Here's the traceback:

#0  _PyFrame_GetCode (f=0xfffff7fea1f0) at ./Include/internal/pycore_frame.h:94
#1  dump_frame (fd=fd@entry=2, frame=frame@entry=0xfffff7fea1f0)
    at Python/traceback.c:914
#2  0x000000000076254c in dump_traceback (fd=fd@entry=2,
    tstate=tstate@entry=0xaeb528 <_PyRuntime+329832>,
    write_header=write_header@entry=0) at Python/traceback.c:1007
#3  0x0000000000762710 in _Py_DumpTracebackThreads (fd=2,
    interp=<optimized out>, current_tstate=current_tstate@entry=0x0)
    at Python/traceback.c:1114
#4  0x00000000007773b4 in faulthandler_thread (unused=unused@entry=0x0)
    at ./Modules/faulthandler.c:631
#5  0x000000000075e77c in pythread_wrapper (arg=<optimized out>)
    at Python/thread_pthread.h:242
#6  0x0000fffff7def188 in start_thread () from /lib64/libc.so.6
#7  0x0000fffff7e5971c in thread_start () from /lib64/libc.so.6

Crash happened on ARM, if that's important.

Hmm, might be a duplicate of #128421.

Edit: reduced test case:

import faulthandler
from threading import Thread

for x in range(20):
    Thread(target=faulthandler.dump_traceback_later, args=(1e-10,)).start()

Found using fusil by @vstinner.

CPython versions tested on:

CPython main branch

Operating systems tested on:

Linux

Output from running 'python -VV' on the command line:

Python 3.14.0a6+ (heads/main:55815a6474c, Mar 15 2025, 14:05:03) [GCC 11.5.0 20240719 (Red Hat 11.5.0-5.0.1)]

[ZeroIntensity]

Hmm, might be a duplicate of #128421.

Probably not, that seems to be FT only. From what I know, faulthandler is notoriously buggy in multithreaded programs, and there's not much we can do to fix it.

[dgrisby]

I think this is a duplicate of #116008

[tom-pytel]

I think this is a duplicate of #116008

Not EXACTLY same, but arises for the same reason yes, using python function _Py_DumpTracebackThreads() in multithreaded scenario without holding the GIL.

[devdanzin]

Closing in favor of #116008.