On 07/12/12 08:20, Dave Taht wrote:
> Also, some new unaligned exception traps have surfaced with ipv6. They
> are not bad - only about 1200 on a 60 second rrul test with 24.5/5.5
> ceroshaper, but enough to dramatically affect ipv6 latencies, and
> affect fq_codel. (see attached)

I have managed to track down possible causes of these new traps now.  
I've noticed that in the IPv6 stack in particular, there's a tendency to 
access the flow label and traffic class fields by casting the header 
structure to __be32*.  Needless to say, this ends up producing a lot of 
unaligned access traps.

There's a possible issue with setting the IPv6 ECN bits, but that should 
also affect the Sugarland build too.  There's also a few *(__be32*) 
dereferences in net/ipv6/ip6_route.c and net/ipv6/ip6_tunnel.c that are 
new to the 3.6 kernel.  The attached patches should hopefully fix all of 
the new traps.

-- 
Robert Bradley