[NumPy-Tickets] [NumPy] #2151: flatiter does not work correctly with non-writeable arrays

NumPy Trac numpy-tickets@scipy....
Tue Jun 5 21:45:42 CDT 2012

#2151: flatiter does not work correctly with non-writeable arrays
 Reporter:  tux         |       Owner:  somebody   
     Type:  defect      |      Status:  new        
 Priority:  normal      |   Milestone:  Unscheduled
Component:  numpy.core  |     Version:  1.6.1      
 Keywords:              |  
 I found out that numpy.flatiter objects do not retain the base object's
 writeable flag. While this only leads to unexpected behavior for normal
 non-writeable arrays, this causes a segfault if the array's buffer is a
 read-only memory area (e.g. mmap).

 crashme1.py shows one part of the problem. The array b is created with the
 mmaped array a as its base. The current code creates a copy and sets the
 UPDATEIFCOPY flag if a is not C_CONTIGUOUS (I used Fortran order here).
 Now when b is destroyed it tries to write to its base (UPDATEIFCOPY) which
 causes a segfault because the buffer is a read-only mmap.

 I tried to fix this problem by only setting updateifcopy only if the
 original array is writeable. If not, the flatiter should be non-writeable
 as well in my opinion to prevent people from trying to change the original
 array with it which would not work. I'm not perfectly sure that I got the
 semantics of iter_array(iterators.c) right. Somebody with more insight
 into numpy should judge. At least the comment of iter_array should also be
 updated to reflect the new situation.

 After I implemented this, stumbled on another bug. Even if the flatiter is
 non-writeable, the function array_flat_set(getset.c) does not actually
 check the flag before writing to the array. This is demonstrated in the
 example write_the_unwriteable.py. If the non-writeable array is a read-
 only memory area, we get a segfault again (see crashme2.py).

 To fix this I added a check for the writeable flag at the beginning of
 array_flat_set which throws an appropriate exception (the same that direct
 assignment would cause).

 I tested all of the above on Debian Sid (Python 2.7.3 and numpy 1.6.2),
 Ubuntu Precise (Python 2.7.3 and numpy 1.6.1) and a self-compiled version
 of Python 2.7.1 and numpy 1.6.1. I could also reproduce it with python

 I attached my patches against the current git HEAD to this report.

Ticket URL: <http://projects.scipy.org/numpy/ticket/2151>
NumPy <http://projects.scipy.org/numpy>
My example project

More information about the NumPy-Tickets mailing list