Py_FatalError calls PyErr_Occurred() which requires a current thread
state.  This mean that if the original error was a thread state error
Py_FatalError is triggered again, ad infinitum.

Do you have a simple code sample to showcase that?

Py_BEGIN_ALLOW_THREADS
	PyErr_SetString(PyExc_SystemError, "bork bork bork");
	Py_END_ALLOW_THREADS

> Roger Upole <rupole@hotmail.com> added the comment:
> 
> Py_BEGIN_ALLOW_THREADS
> 	PyErr_SetString(PyExc_SystemError, "bork bork bork");
> 	Py_END_ALLOW_THREADS

Well, if you start using all kinds of internal APIs without holding the
GIL, you should expect much worse than just a Py_FatalError ;-)

Of course, it would still be better if the infinite loop could be
avoided.

Well, you asked for a simple case.  In practice, I'm dealing with
much more complicated code that contains callbacks which may need
to acquire or release the theadlock.  This problem makes it very 
difficult to debug such code.

Also, the PyErr_Occurred is not present in 2.5.2, probably
for this very same reason.

I have a fix at http://codereview.appspot.com/1184043, which makes PyErr_Occurred() return NULL when there is no thread.  I'll commit it tomorrow unless there are comments.

Fixed in r81142.

I fixed the new test for Windows: r81183 (py3k). The new test failed on Windows because of the different newline characters:

Traceback (most recent call last):
  File "...\test_capi.py", line 49, in test_no_FatalError_infinite_loop
    b'Fatal Python error:'
AssertionError: b'Fatal Python error: PyThreadState_Get: no current thread\r\n' != b'Fatal Python error: PyThreadState_Get: no current thread\n'

I know this has been closed, but I disagree with the fix. Py_FatalError is designed to give a clean exit with a readable error message when the state of the interpreter is known to be compromised.  Calling *any* python functions or accessing *any* python internals is only likely to lead to more issues that obscure the original problems.  IMO the PyErr_Occurred needs to be removed entirely.

Would it be possible for this issue's fix (PyErr_Occurred() returning null when the thread state is null) to be applied to the 2.7 branch?

New changeset d12d0e7c0fe2b49c40ac4d66365147c619d6c475 by Victor Stinner in branch 'master':
bpo-38733: PyErr_Occurred() caller must hold the GIL (GH-17080)
https://github.com/python/cpython/commit/d12d0e7c0fe2b49c40ac4d66365147c619d6c475