When Python cannot report an error and its is not possible to recover from the error, Py_FatalError() displays an error message and exit immediately Python.

It's common that Python crashes are coming from third party extension modules, rather than Python itself. Recent example: bpo-42891 crash reported in Python, but it was a bug in lsm-db third party extension module.

I propose to enhance Py_FatalError() to attempt to dump the list of extension modules, to ease debugging.

I'm working on a PR.

Example using PR 24207, the new part the "Extension modules:" list at the end:
---------------

$ ./python -X faulthandler
Python 3.10.0a4+ (heads/master-dirty:2396614b89, Jan 13 2021, 12:09:15) 
>>> import ctypes
>>> ctypes.string_at(0)
Fatal Python error: Segmentation fault

Current thread 0x00007f0aabd09740 (most recent call first):
File "/home/vstinner/python/master/Lib/ctypes/init.py", line 517 in string_at
File "<stdin>", line 1 in <module>

Extension modules:

• sys
• builtins
• _imp
• _thread
• _warnings
• _weakref
• _io
• marshal
• posix
• time
• faulthandler
• _codecs
• _signal
• _abc
• _stat
• readline
• atexit
• itertools
• _operator
• _collections
• _functools
• _opcode
• _sre
• _locale
• _ctypes
• _struct
  ---------------

Here you can see that the _ctypes module is loaded.

It's common that Python crashes are coming from third party extension modules, rather than Python itself. Recent example: bpo-42891 crash (...)

Another recent example: bpo-42879 "python: Objects/abstract.c:155: PyObject_GetItem: Assertion `(item != NULL) ^ (PyErr_Occurred() != NULL)' failed." was a bug in pygame.

Example of a Pillow crash in Python 3.10:
https://mail.python.org/archives/list/python-dev@python.org/message/5DCR2R6POCOHTPFUOY4F5G7GTBYX6OPO/

New changeset e232025 by Victor Stinner in branch 'master':
bpo-42923: Add Py_FatalError() test in test_capi (GH-24240)
e232025

New changeset 314b878 by Victor Stinner in branch 'master':
bpo-42923: Py_FatalError() avoids fprintf() (GH-24242)
314b878

New changeset 250035d by Victor Stinner in branch 'master':
bpo-42923: Dump extension modules on fatal error (GH-24207)
250035d

I created https://bugs.python.org/issue42955 to add sys.modules_names tuple: names of stdlib modules. Once it will be merged, I will updated this PR to filter the list of modules (ignore stdlib modules).

New changeset 66f77ca by Victor Stinner in branch 'master':
bpo-42923: _Py_DumpExtensionModules() ignores stdlib ext (GH-24254)
66f77ca

More "real world" example using cinder which imports 220 modules (ctypes is used to simulate a crash):
---------------

$ ./python -m venv env
$ env/bin/python -m pip install wheel
$ env/bin/python -m pip install cinder
$ env/bin/python -X dev -c 'import cinder, ctypes, sys; print(f"{len(sys.modules)=}"); print(); ctypes.string_at(0)'
(...)
len(sys.modules)=220

Fatal Python error: Segmentation fault

Current thread 0x00007fc4a88e0740 (most recent call first):
File "/home/vstinner/python/master/Lib/ctypes/init.py", line 517 in string_at
File "<string>", line 1 in <module>

Extension modules: greenlet._greenlet, __original_module__thread, __original_module_select, __original_module_time, _cffi_backend (total: 5)
Erreur de segmentation (core dumped)
---------------

cinder only uses 2 third party extension modules (on a total of 220 modules): greenlet and cffi.

Note: __original_xxx modules are aliases of stdlib modules created by eventlet monkey patching.

So if cinder does crash, I suggest to look at Python, but *also* look at these 2 extensions ;-)

See also bpo-42955 "Add sys.module_names: list of stdlib module names (Python and extension modules)".

New changeset 3d55aa9 by Victor Stinner in branch 'master':
bpo-42923: Fix _Py_DumpExtensionModules() for NULL sysdict (GH-25262)
3d55aa9