Python 3.3.2 (default, Dec  4 2013, 20:19:27) 
[GCC 4.7.3] on linux (gentoo)

Suppose we create a context manager that will fail during exit because of a simple coding error:

@contextmanager
def f():
 try:
  yield 6
 finally:
  throdbog() # woops, forgot to define this

Now let's stack up a few of these:

with ExitStack() as stack:
 i = stack.enter_context(f())
 i = stack.enter_context(f())
 i = stack.enter_context(f())
 print(i)

... prints 6, then hangs, won't respond to ^C, can be killed.

Three levels on the stack seems to be the magic number. With one or two it works as expected ("During handling of the above exception, another exception occurred", etc.).

Is the ExitStack code somehow creating a circularly-linked exception structure when there are three levels?

New changeset b3c2472c12a1 by Nick Coghlan in branch '3.3':
Issue #20317: Don't create a reference loop in ExitStack
http://hg.python.org/cpython/rev/b3c2472c12a1

Oh, that's quite neat. Because of the way _GeneratorContextManager works, the context is actually already set correctly on the thrown exceptions, but ExitStack is assuming it will be wrong.

This means _fix_exception (part of ExitStack.__exit__) ends up setting the first exception's context to itself while handling the second exception, which then creates an infinite loop when attempting to handle the third exception. Hence why three is the magic number :)

The existing tests didn't pick this up, because they just used callbacks for simplicity, and hence the context was wrong as expected.

Commit incoming.

New changeset 46c3ea358784 by Nick Coghlan in branch 'default':
Merge #20317 from 3.3
http://hg.python.org/cpython/rev/46c3ea358784

New changeset b3eaeb4bdf84 by Nick Coghlan in branch '3.3':
Issue 20317: Remove debugging code from contextlib
http://hg.python.org/cpython/rev/b3eaeb4bdf84

New changeset a0bf53afedfa by Nick Coghlan in branch 'default':
Merge removal of issue 20317 debugging code from 3.3
http://hg.python.org/cpython/rev/a0bf53afedfa