memory_item() function creates a new memoryview object if ndim is different than 1. Example with numpy:

   from numpy import array
   y=array([ 0,  1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11])
   y.shape = 3,4
   view=memoryview(y)
   view2 = view[0]
   print type(view2)

Result: <type 'memoryview'>. (Without the shape, ndim equals 1, and view2 is a standard str object.)

The problem is that view attribute of the view2 (PyMemoryViewObject) is not initialized. Extract of memory_item():

        /* Return a new memory-view object */
        Py_buffer newview;
        memset(&newview, 0, sizeof(newview));
        /* XXX:  This needs to be fixed so it actually returns a sub-view */
        return PyMemoryView_FromBuffer(&newview);

"This needs to be fixed" :-/

--

view2 is not initialized and so view2.tolist() does crash.

Full example (using numpy) crashing Python: memoryview_crash.py

According to #2394, the implementation of the new buffer API is not complete.

I'll leave this to our numpy experts. I don't even know how to unit test this without installing numpy.

The crash is fixed in the features/pep-3118 repo:

>>> from numpy import array
>>> y=array([ 0,  1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11])
>>> y.shape = 3,4
>>> view=memoryview(y)
>>> view2 = view[0]
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
NotImplementedError: multi-dimensional sub-views are not implemented
[182251 refs]

Fixed in #10181.

New changeset 3f9b3b6f7ff0 by Stefan Krah in branch 'default':
- Issue #10181: New memoryview implementation fixes multiple ownership
http://hg.python.org/cpython/rev/3f9b3b6f7ff0

Since this issue targeted 2.7 and 3.2:

In a brief discussion on python-dev it was decided that the 3.3 fixes
from #10181 won't be backported for a number of reasons, see:

http://mail.python.org/pipermail/python-dev/2012-February/116872.html