| msg26375 - (view) |
Author: Armin Rigo (arigo) *  |
Date: 2005-09-24 23:40 |
| 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. |
|
|
| msg26376 - (view) |
Author: Armin Rigo (arigo) * |
Date: 2005-09-25 08:55 |
| 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. |
|
|
| msg26377 - (view) |
Author: Armin Rigo (arigo) * |
Date: 2005-12-14 14:49 |
| 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. |
|
|
| msg26378 - (view) |
Author: Armin Rigo (arigo) * |
Date: 2005-12-14 15:32 |
| 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. |
|
|
| msg26379 - (view) |
Author: Michael Hudson (mwh) |
Date: 2005-12-14 15:36 |
| Logged In: YES user_id=6656 Yikes! |
|
|
| msg26380 - (view) |
Author: Brett Cannon (brett.cannon) * |
Date: 2006-06-29 17:41 |
| 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. |
|
|
| msg26381 - (view) |
Author: Brett Cannon (brett.cannon) * |
Date: 2006-06-29 17:45 |
| 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? |
|
|
| msg26382 - (view) |
Author: Armin Rigo (arigo) * |
Date: 2006-06-29 21:19 |
| 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. |
|
|
| msg26383 - (view) |
Author: Brett Cannon (brett.cannon) * |
Date: 2006-06-30 01:14 |
| 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. |
|
|
| msg26384 - (view) |
Author: Armin Rigo (arigo) * |
Date: 2006-06-30 07:05 |
| 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. |
|
|
| msg26385 - (view) |
Author: Jeremy Hylton (jhylton)  |
Date: 2007-02-25 22:23 |
| 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? |
|
|
| msg26386 - (view) |
Author: Jeremy Hylton (jhylton) |
Date: 2007-02-25 22:36 |
| 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? |
|
|
| msg26387 - (view) |
Author: Armin Rigo (arigo) * |
Date: 2007-02-26 08:23 |
| 456? Now that's interesting. Not 579? I'm not sure if I ever thought about what the expected answer should be, but I'd say that you are correct: 'x.__dict__' should be empty in the end. Actually I don't really see how it manages *not* to be empty in IronPython; looks like either a very minor detail or a deeper bug... |
|
|
| msg26388 - (view) |
Author: Ziga Seilnacht (zseil) * |
Date: 2007-04-17 00:42 |
| Here is a patch that shold fix the deldict bug. It also removes the new condition for metaclasses, because it isn't needed anymore, and fixes bug #1174712. These two changes were needed so that the patch could be properly tested. The patch does what Armin suggested; the __dict__ descriptor looks for a builtin base with tp_dictoffset != 0, and if one is found, it delegates to that base's __dict__ descriptor. File Added: modify_dict_bug.diff |
|
|
| msg26389 - (view) |
Author: Armin Rigo (arigo) * |
Date: 2007-04-18 09:22 |
| I'm skeptical about the whole revision 53997 which introduced not only the unneeded metaclass condition, but also the strange check when assigning to __bases__. I don't think that this check about __bases__ is correct or relevant or really fixes any crash. The link between the bases and the metaclass of a class is tenuous anyway: the bases are just used to compute the metaclass if none is specified explicitly. As I said on python-dev (with no answer) I am thinking about reverting r53997 completely. I'm going to check what I said above a bit more in-depth first. |
|
|
| msg26390 - (view) |
Author: Armin Rigo (arigo) * |
Date: 2007-04-19 14:46 |
| Reverted r53997 in revision 54875. zseil, could you update your patch for the new svn head? Thanks! It should mostly be a matter of simplifying it. |
|
|
| msg26391 - (view) |
Author: Ziga Seilnacht (zseil) * |
Date: 2007-04-19 16:40 |
| Here is the updated patch. I also added a few more tests, and removed the news entry for revision 53997. File Added: modify_dict_bug2.diff |
|
|
| msg26392 - (view) |
Author: Armin Rigo (arigo) * |
Date: 2007-05-02 19:23 |
| Thanks zseil for the patch, which looks really good to me. For reference, it also contains the fix for #1174712. Checked in as r55080. |
|
|
| msg61603 - (view) |
Author: Guido van Rossum (gvanrossum) * |
Date: 2008-01-23 20:49 |
| Reopening, since test67.py still causes a segfault in the trunk. It is represented as Lib/test/crashers/loosing_dict_ref.py [sic]. |
|
|
| msg61608 - (view) |
Author: Guido van Rossum (gvanrossum) * |
Date: 2008-01-23 21:33 |
| Armin, what do you think of the attached patch (deldict.diff)? |
|
|
| msg61610 - (view) |
Author: Guido van Rossum (gvanrossum) * |
Date: 2008-01-23 21:45 |
| Here's a better fix, which also fixes the similar code path in GenericSetAttr(). As a bonus, I reviewed all uses of _PyObject_GetDictPtr() and found one questionable place where the dict at *dictptr was DECREF'ed before that location was set to NULL. I think we're better off setting it to NULL and *then* DECREF'ing the dict. |
|
|
| msg61616 - (view) |
Author: Guido van Rossum (gvanrossum) * |
Date: 2008-01-23 23:02 |
| The more i think about it, I don't think the issue in typeobject.c can ever occur, so I'm skipping that part of the fix. |
|
|
| msg61620 - (view) |
Author: Neal Norwitz (nnorwitz) * |
Date: 2008-01-24 07:12 |
| I looked at Guido's latest deldict.diff patch--the one to Objects/object.c only. It seems good. I can't convince myself either way about the change to Objects/typeobject.c. I can't think of a way to cause a problem. It seems safer to use Py_CLEAR in this case though. There are several other uses of _PyObject_GetDictPtr in Objects/typeobject.c. It was pretty much the same--I can't convince myself either way. Can Py_VISIT cause any Python code to execute that might lead to a problem? The other uses of _PyObject_GetDictPtr in Objects/typeobject.c seemed safer. Not a very useful review. |
|
|
| msg61640 - (view) |
Author: Guido van Rossum (gvanrossum) * |
Date: 2008-01-24 17:05 |
| On Jan 23, 2008 11:12 PM, Neal Norwitz <report@bugs.python.org> wrote: > > Neal Norwitz added the comment: > > I looked at Guido's latest deldict.diff patch--the one to > Objects/object.c only. It seems good. I can't convince myself either > way about the change to Objects/typeobject.c. I can't think of a way > to cause a problem. It seems safer to use Py_CLEAR in this case > though. Here's my reasoning: the object whose dict is being cleared itself has a refcount of zero at this point. So it is truly unreachable from Python code. So I'm not going to submit that part of the change. > There are several other uses of _PyObject_GetDictPtr in > Objects/typeobject.c. It was pretty much the same--I can't convince > myself either way. Can Py_VISIT cause any Python code to execute that > might lead to a problem? The answer lies in the gc module which does all the visiting. A quick scan of all the traverse() calls there indicates that none of them call back into Python -- not a DECREF in sight. > The other uses of _PyObject_GetDictPtr in > Objects/typeobject.c seemed safer. Not a very useful review. I reviewed those too and found them safe. I'll submit the change now. |
|
|
| msg61649 - (view) |
Author: Guido van Rossum (gvanrossum) * |
Date: 2008-01-24 18:57 |
| r60247 (2.5.2 branch), r60246 (2.6 trunk). |
|
|