* Herbert Xu <20050124225423.GA15405@xxxxxxxxxxxxxxxxxxx> 2005-01-25 09:54
> On Mon, Jan 24, 2005 at 04:15:10PM +0100, Thomas Graf wrote:
> > After inspecting your iptables rule set I think it is a general UDP DNAT
> > problem under some circumstances. Some defragmentation weirdness in
> > prerouting might be invovled. It would definitely help to have a dump
> > of a complete ip fragments sequence causing this bug but I can't tell
> > what exactly is the cause just now so yes it might be a good idea to
> > limit the dump to the above subnet and hope the dodgy traffic comes
> > from the same subnet again.
> OK, I think I've found the problem. It's a totally innocuous bug
> in ip_fragment/ip6_fragment. When we're in the fast path and use
> the pre-existing frag_list skb's, we forgot to clear ip_summed.
I don't quite understand how this solves the problem. How could
ip_summed be non zero after ip_forward? The earliest possible call
to ip_fragment is in postrouting. Please correct me if I'm wrong.
The bug isn't triggered for every fragment only once in a while so I don't
think it's that simple.