[Python-Dev] are CObjects inherently unsafe?

[Michael Hudson]

It seems to me that any time there are more than two CObjects around we 
have a good chance of causing mischief:

 >>> import cStringIO, _curses
 >>> cStringIO.cStringIO_CAPI = _curses._C_API
 >>> import cPickle
 >>> cPickle.dumps(range(10))
Illegal instruction

(if you're trying this at home, starting python with -E might be 
necessary to make sure cPickle hasn't been imported already).

Question 1: do we care?  I think we do: "no core dumps" is a pretty 
strong Pythonic principle, even though this is certainly a "doctor it 
hurts when I do this" situation.  OTOH, Armin's gc.get_referrers hack 
was deemed allowable.

It seems to me that it would be more sensible to have a dict mapping 
names (maybe just module names...) to cobjects in the interpreter 
state.  I guess there might be a way of getting ahold of the dict with 
the gc module (though I can't think of one off hand).  This wouldn't be 
a difficult change, and if the CObjects are left in the module dicts, 
it shouldn't even do that much damage to binary compatibility.

Thoughts?

Cheers,
mwh
(who also notices that the cobject section in Doc/ext/ could do with 
being updated to use PyCObject_Import...)