gh-128400: Only show the current thread in faulthandler if the GIL is disabled

[ZeroIntensity]

A few notes here:

• Displaying other threads is useful for debugging, so some cases such as where the GIL is re-enabled or during a garbage collection (and therefore threads are stopped) still show all threads.
• Py_FatalError still displays all threads, but that's a seperate bug--is it safe to stop-the-world during a fatal error?
• Should this get backported?

• Issue: Crash when using _Py_DumpTracebackThreads #128400

---

📚 Documentation preview 📚: https://cpython-previews--128425.org.readthedocs.build/

[colesbury]

The gc.collecting field doesn't necessarily imply that all threads are paused:

• gc.collecting is set before the stop-the-world pause
• The GC resumes threads while calling finalizers

You could check tstate->interp->stoptheworld.world_stopped and maybe also that tstate matches stoptheworld.requester. (The faulting thread may not be attached to the interpreter when it crashes.)

[colesbury]

I'm not fully convinced that this is worth the extra complexity or the extra chance of crashing during the faulthandler:

• Most crashes probably won't occur during a stop-the-world pause and for those that do, the stack traces of other threads are even less likely to be relevant (because they're paused).
• Even during a stop-the-world pause, looping over the linked list of thread states is not safe without a HEAD_LOCK(), which you can't do during a signal handler.

[ZeroIntensity]

Yeah, I was on the fence about it too. I threw it in there because it seemed simple enough at first, but I think it's better to remove it too.

[colesbury]

Py_FatalError still displays all threads, but that's a seperate bug--is it safe to stop-the-world during a fatal error?

I think it's best not to stop-the-world. Deadlocks that hide the fatal error message would not be fun.

[ZeroIntensity]

I'll deal with Py_FatalError in a follow-up (after this gets merged to prevent conflicts).

[colesbury]

Looks good. Some minor formatting suggestions below

[ZeroIntensity]

ASan failure looks unrelated.

[colesbury]

Thanks!

[colesbury]

Should this get backported?

I'd lean towards "no" for now