On my system (Ubuntu Karmic, Python 2.6.4 (r264:75706, Dec  7 2009, 18:45:15), Kernel 2.6.31-17-generic, libc6 2.10.1-0ubuntu16) the attached test script produces the following output:

Traceback (most recent call last):
  File "test1.py", line 27, in <module>
    raise OSError(err, os.strerror(err), path)
OSError: [Errno 0] Success: '/does/not/exist'

So the function call failed, but the errno provided by ctypes is zero.

(There are two variants of construction the getxattr() foreign function in the script and both produce the same result).

In the libc case you shouldn't give absolute path in CDLL: CDLL('libc.so.6') is better. You use "/lib/libc.so.6" path, but Python (and ctypes.so) actually uses something like "/lib/tls/i686/cmov/libc.so.6" - these are two separate libraries with separate "errno" variables, mapped to different memory regions. Your call to getxattr modifies errno in the former library, but ctypes get_errno() fetches value from the latter library.

If I fix CDLL() call, the test works ok for me.

I can confirm that without the path it works for me too.

But I have to admit that I don't really understand your explanation. Should I generally not use full paths with CDLL? Or just in the case of libc?

In either case, I think the ctypes dokumentation could be more explicit about this.