homepage

The recvmsg tests in test_socket.py check that the address returned by recvmsg matches the original address to which the socket was bound. For IPv6, sockaddr includes sin6_scope_id, in addition to the address and port.
The test connects to host "::1", which is loopback, but is an under-specified address because the link scope is left ambiguous.
The scope_id in the original bind call defaults to "0", which represents an ambiguous scoped address and allows the IPV6 protocol and implementation to choose the interface or site identifier. The recvmsg call returns the actual scope_id.
The test incorrectly checks that the full sockaddr matches. sin6_scope_id may not match for IPv6 addresses. This generates bogus failures on AIX.
(Microsoft has a good description about scope_id:
http://msdn.microsoft.com/en-us/library/windows/desktop/ms739166%28v=vs.85%29.aspx)

The current arguments to checkRecvmsgAddress() are the sockaddrs, so it does not know the protocol family. One potential patch to infer the family and apply the correct test is:
diff -r 035d8fed07ad Lib/test/test_socket.py
--- a/Lib/test/test_socket.py Tue Jun 25 22:54:35 2013 +0200
+++ b/Lib/test/test_socket.py Wed Jun 26 15:16:31 2013 -0700
@@ -1809,7 +1809,10 @@
 def checkRecvmsgAddress(self, addr1, addr2):
 # Called to compare the received address with the address of
 # the peer.
- self.assertEqual(addr1, addr2)
+ if len(addr1) > 3 or len(addr2) > 3:
+ self.assertEqual(addr1[:-1], addr2[:-1])
+ else:
+ self.assertEqual(addr1, addr2)
 
 # Flags that are normally unset in msg_flags
 msg_flags_common_unset = 0

"The scope_id in the original bind call defaults to "0", which represents an ambiguous scoped address and allows the IPV6 protocol and implementation to choose the interface or site identifier."
Ok, so here is a patch using scope_id=1 to get a reliable IPv6 address. Can you please try it on AIX?

The patch in msg192259 probably will work on the particular AIX system runnnig the buildbot, but I do not believe that the patch is the correct solution for the problem.
scope_id 1 is not necessarily the correct link for IPv6 address ::1. The scope_id values are system dependent and the result could be any value depending on the interfaces available on the particular system running the test. scope_id 1 might not exist on the system, and forcing that value would generate an error.
I think the safest solution is not to compare scope_id when comparing addresses.

> I think the safest solution is not to compare scope_id when comparing
> addresses.
Agreed.
However, it might be simpler to special-case the IPv6 addresses
comparison by overriding it in the IPv6 sendmsg base test.
Could you try the patch attached?

"However, it might be simpler to special-case the IPv6 addresses
comparison by overriding it in the IPv6 sendmsg base test."
Yes, this is why I proposed another approach. (But David convinced me
that it is not the right approach.)
Victor

The patch in msg192405 works and fixes that error on AIX. That is exactly what I had in mind, but I incorrectly had been looking higher up the class hierarchy to override the method. Thanks!

New changeset 330c7aa2922b by Charles-François Natali in branch '3.3':
Issue #18308: don't take the scope ID into account when comparing IPv6
http://hg.python.org/cpython/rev/330c7aa2922b
New changeset b44749cee660 by Charles-François Natali in branch 'default':
Issue #18308: don't take the scope ID into account when comparing IPv6
http://hg.python.org/cpython/rev/b44749cee660 

Fixed, thanks!