Bugs item #1303614, was opened at 2005-09-24 23:40 Message generated for change (Comment added) made by jhylton You can respond by visiting: https://sourceforge.net/tracker/?func=detail&atid=105470&aid=1303614&group_id=5470
Please note that this message will contain a full copy of the comment thread, including the initial issue submission, for this request, not just the latest update. Category: Python Interpreter Core Group: None Status: Open Resolution: None Priority: 5 Private: No Submitted By: Armin Rigo (arigo) Assigned to: Nobody/Anonymous (nobody) Summary: Bypassing __dict__ readonlyness Initial Comment: The __dict__ attribute of some objects is read-only, e.g. for type objects. However, there is a generic way to still access and modify it (without hacking with gc.get_referrents(), that is). This can lead to obscure crashes. Attached is an example that shows a potential "problem" involving putting strange keys in the __dict__ of a type. This is probably very minor anyway. If we wanted to fix this, we would need a __dict__ descriptor that looks more cleverly at the object to which it is applied. BTW the first person who understand why the attached program crashes gets a free coffee. ---------------------------------------------------------------------- >Comment By: Jeremy Hylton (jhylton) Date: 2007-02-25 22:36 Message: Logged In: YES user_id=31392 Originator: NO I tried test67.py in IronPython. It reports that x.hello has the value 456. Is that the correct behavior? It seems incorrect to me. If the __eq__ method is called, then the object should no longer have either the Strange() or hello attributes. They are cleared by reseting __dict__. Is that right? ---------------------------------------------------------------------- Comment By: Jeremy Hylton (jhylton) Date: 2007-02-25 22:23 Message: Logged In: YES user_id=31392 Originator: NO I tried test67.py in IronPython. It reports that x.hello has the value 456. Is that the correct behavior? It seems incorrect to me. If the __eq__ method is called, then the object should no longer have either the Strange() or hello attributes. They are cleared by reseting __dict__. Is that right? ---------------------------------------------------------------------- Comment By: Armin Rigo (arigo) Date: 2006-06-30 07:05 Message: Logged In: YES user_id=4771 Well, try making an "empty" class Foo(object): pass, and see what magically shows up in Foo.__dict__.keys(). Here is the __dict__ descriptor used for instances of Foo. This descriptor is connected to subtype_dict() and subtype_setdict() in typeobject.c. INCREF/DECREFs are in theory missing around every use of the dictionary returned by _PyObject_GetDictPtr(), because more or less any such use could remove the dict from the object from which _PyObject_GetDictPtr() returned from, and so decref the dict - while it's used. This might require some timing, to check the performance impact. ---------------------------------------------------------------------- Comment By: Brett Cannon (bcannon) Date: 2006-06-30 01:14 Message: Logged In: YES user_id=357491 OK, then I am going to need a little guidance to dive into this more since this is going into some murky waters for me. =) For the normal, non-evil case of an empty Python class (``class Foo(object): pass``), I would think that accessing __dict__ would fall through to the tp_getset for object, and then fall to the same slot for type. Now that is obviously not happening since you get a straight dict back for that attribute access and not a dict proxy as would be suggested based on my logic listed above. So, my question is, where is the code that handles fetching Foo().__dict__? And if you have an inkling of where I should be looking in terms of possible refcounting additions that would be great as well. ---------------------------------------------------------------------- Comment By: Armin Rigo (arigo) Date: 2006-06-29 21:19 Message: Logged In: YES user_id=4771 Brett: I think you're approaching the problem from the wrong angle. The problem is being allowed to freely tamper with the dict stored in objects. Getting NULL errors here and there is only a symptom. As I explain, the '__dict__' descriptor object needs to do some more checks, and to be fully safe some Py_INCREF/Py_DECREF are needed in some critical places. ---------------------------------------------------------------------- Comment By: Brett Cannon (bcannon) Date: 2006-06-29 17:45 Message: Logged In: YES user_id=357491 For the deldict.py crasher, if you look at the traceback there is no good place to do a check that tp_dict is sane except in type_module() or PyDict_GetItem(). Now adding the NULL check in type_module() will fix this specific problem, but that seems like an ad-hoc patch. Why don't we check for NULL in PyDict_GetItem() and return NULL just like the PyDict_Check() test? I am sure the answer is "performance", but this is not PyDict_GETITEM()and thus already is not the performance-critical version anyway. So why shouldn't we check for NULL there and possibly catch other errors? ---------------------------------------------------------------------- Comment By: Brett Cannon (bcannon) Date: 2006-06-29 17:41 Message: Logged In: YES user_id=357491 Simple patch for the loosing_dict_ref.py crasher is attached. Just checked for possible NULL values in what is needed by _Py_ForgetReference(). Let me know what you think, Armin. ---------------------------------------------------------------------- Comment By: Michael Hudson (mwh) Date: 2005-12-14 15:36 Message: Logged In: YES user_id=6656 Yikes! ---------------------------------------------------------------------- Comment By: Armin Rigo (arigo) Date: 2005-12-14 15:32 Message: Logged In: YES user_id=4771 A proposed fix for the later crash: the __dict__ descriptor of user-defined classes should verify if there is another __dict__ descriptor along the solid path of the type (i.e. following "tp_base" pointers). If so, it should delegate to it. ---------------------------------------------------------------------- Comment By: Armin Rigo (arigo) Date: 2005-12-14 14:49 Message: Logged In: YES user_id=4771 Uh oh. There is a much simpler crash. The technique outlined in deldict.py can be used to rebind or even delete the __dict__ attribute of instances where it should normally not be allowed. ---------------------------------------------------------------------- Comment By: Armin Rigo (arigo) Date: 2005-09-25 08:55 Message: Logged In: YES user_id=4771 The bug is related to code like PyObject_GenericGetAttribute and _PyType_Lookup which are not careful about the reference counters of the objects they operate on. This allows a much simpler crash (test67.py): the __dict__ of an object can be decrefed away while lookdict() is looking at it. This has a simple fix -- drop some amount of Py_INCREF/Py_DECREF in core routines like PyObject_GenericGetAttr. We probably need to measure if it has a performance impact, though. ---------------------------------------------------------------------- You can respond by visiting: https://sourceforge.net/tracker/?func=detail&atid=105470&aid=1303614&group_id=5470 _______________________________________________ Python-bugs-list mailing list Unsubscribe: http://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com
