The makesockaddr() function in the socket module assumes that
AF_UNIX addresses have a null-terminated sun_path, but Linux
actually allows unterminated addresses using all 108 bytes of
sun_path (for normal filesystem sockets, that is, not just
abstract addresses).

When receiving such an address (e.g. in accept() from a
connecting peer), makesockaddr() will run past the end and return
extraneous bytes from the stack, or fail because they can't be
decoded, or perhaps segfault in extreme cases.

This can't currently be tested from within Python as Python also
refuses to accept address arguments which would fill the whole of
sun_path, but the attached linux-pass-unterminated.diff (for 2.x
and 3.x) enables them for Linux.  With the patch applied:

Python 2.7a4+ (trunk, Apr  8 2010, 18:20:28) 
[GCC 4.2.4 (Ubuntu 4.2.4-1ubuntu4)] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> import socket
>>> s = socket.socket(socket.AF_UNIX)
>>> s.bind("a" * 108)
>>> s.getsockname()
'aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa\xfa\xbf\xa8)\xfa\xbf\xec\x15\n\x08l\xaaY\xb7\xb8CZ\xb7'
>>> len(_)
126

Also attached are some unit tests for use with the above patch, a
couple of C programs for checking OS behaviour (you can also see
the bug by doing accept() in Python and using the bindconn
program), and patches aimed at fixing the problem.

Firstly, the return-unterminated-* patches make makesockaddr()
scan sun_path for the first null byte as before (if it's not a
Linux abstract address), but now stop at the end of the structure
as indicated by the addrlen argument.

However, there's one more catch before this will work on Linux,
which is that Linux system calls return the length of the address
they *would* have stored in the structure had there been room for
it, which in this case is one byte longer than the official size
of a sockaddr_un structure, due to the missing null terminator.

The addrlen-* patches handle this by always calling
makesockaddr() with the actual buffer size if it is less than the
returned length.  This silently ignores any truncation, but I'm
not sure how to do anything sensible about that, and some
operating systems (e.g. FreeBSD) just silently truncate the
address anyway and don't return the original length (POSIX
doesn't make clear which, if either, behaviour is required).
Once these patches are applied, the tests pass.

There is one other issue: the patches for 3.x retain the
assumption that socket paths are in UTF-8, but they should
actually be handled according to PEP 383.  I've got a patch for
that, but I'll open a separate issue for it since the handling of
the Linux abstract namespace isn't documented and there's some
slightly unobvious behaviour that people might be depending on.