Ammar Askar added the comment:
The proposed patch fixes this, not sure if a regression test is appropriate
here.
Here's a more minimal example that demonstrates the exact problem:
```
class Index():
def __index__(self):
global a
a.append("2")
return 999
a = bytearray(b"1")
buf = buffer(a)
s = buf[:1:Index()]
# buf[Index():x:x] or buf[x:x:Index()] will also crash
```
The problem doesn't show up when doing buffer[x:Index()] or [Index():x] because
this syntax calls the sq_slice method implemented by buffer object which is
passed the indexes as numbers.
However when using slice notation with three arguments, the equivilant of these
lines of code is executed:
```
slice_object = slice(x, Index(), x)
buffer[slice_object]
```
During the `buffer[slice_object]`, a call is made in the slice object to find
the indexes of the slice. This calls into the __index__ method of the Index
class which mutates the underlying storage behind the buffer. However, buffer's
subscript method stores the underyling storage in a local variable before
calling the GetIndices method (assuming the object won't be mutated) which
means that when it returns, it returns a pointer to an older portion of memory.
I took a quick look at listobject, stringobject, unicodeobject, tupleobject and
bytearrayobject's subscript methods and it seems they all only access their
members after the call to PySlice_GetIndices, so I think they should be fine.
memoryview objects cause a `BufferError: Existing exports of data: object
cannot be re-sized` error so Py3 should be fine.
----------
keywords: +needs review, patch
nosy: +ammar2
stage: -> patch review
Added file: http://bugs.python.org/file46046/buffer-use-after-free-fix.patch
_______________________________________
Python tracker <[email protected]>
<http://bugs.python.org/issue29028>
_______________________________________
_______________________________________________
Python-bugs-list mailing list
Unsubscribe:
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com
