Author dmalcolm
Recipients dmalcolm
Date 2010-08-25.22:48:50
SpamBayes Score 1.2663e-07
Marked as misclassified No
Message-id <1282776532.58.0.602772290765.issue9688@psf.upfronthosting.co.za>
In-reply-to
Content
On 64-bit bigendian machines (ppc64 and s390x), I'm seeing:
>>> print object.__basicsize__
0

(Discovered via a segfault in Jinja2 tries to use ctypes to manipulate ob_refcnt of variables, and gets the wrong location, corrupting the objects instead; see https://bugzilla.redhat.com/show_bug.cgi?id=627347 )

struct _typeobject declares tp_basicsize and tp_itemsize as Py_ssize_t:
{
...
  Py_ssize_t tp_basicsize, tp_itemsize; /* For allocation */
...
}

but type_members defines them as T_INT:
 {"__basicsize__", T_INT, offsetof(PyTypeObject,tp_basicsize),READONLY},
 {"__itemsize__", T_INT, offsetof(PyTypeObject, tp_itemsize), READONLY},

Hence when accessing "object.__basicsize__", PyMember_GetOne reads it as a
T_INT, which gets it as 0 (incorrect).  Accessing it as Py_ssize_t reads it as
16 (correct)
(gdb) p *(Py_ssize_t*)addr
$9 = 16
(gdb) p *(int*)addr
$10 = 0

I'm attaching a patch which changes them to use T_PYSSIZE_T and adds a selftest.
History
Date User Action Args
2010-08-25 22:48:52dmalcolmsetrecipients: + dmalcolm
2010-08-25 22:48:52dmalcolmsetmessageid: <1282776532.58.0.602772290765.issue9688@psf.upfronthosting.co.za>
2010-08-25 22:48:51dmalcolmlinkissue9688 messages
2010-08-25 22:48:51dmalcolmcreate