Re: [PATCH net] ip: make IP identifiers less predictable

[Hannes Frederic Sowa <hannes@stressinduktion.org>]

Hi,

On Fr, 2014-07-25 at 20:38 +0200, Eric Dumazet wrote:
> On Fri, 2014-07-25 at 11:35 -0700, Linus Torvalds wrote:
> > On Fri, Jul 25, 2014 at 11:09 AM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
> > >
> > > We might change the hash to use both daddr & saddr to increase
> > > protection.
> > 
> > .. and maybe protocol too, so that you can't easily use icmp echo
> > packets to do it for udp packets etc. The underlying jhash is
> > jhash_3words(), so that would actually be fairly natural for at least
> > ipv4 (the ipv6 case I didn't look at).
> 
> Right, in fact saddr is probably not worth it.
> 
> Its not like servers have dozen of IPv4 addresses anyway...

Another idea, maybe worth looking at:

Since commit 703133de331a7a ("ip: generate unique IP identificator if
local fragmentation is allowed") we started to generate fragmentation
ids in the output path for every packet that has ignore_df set, which
nearly is every packet.

We could try to push that down the stack and only insert the
fragmentation id in ip_fragment. We still need to generate the frag_id
directly from the socket layer, but we can reuse the ip6_frag_id field
in skb_shinfo for IPv4, too.

Then we actually only need to generate fragmentation ids for the VJ
compression workaround, generated from the socket->inet_id. Do we still
need this (I guess, yes)?

Does this sound worth a try or are there any unseen protocol specific
consequences I am not yet aware of? We would stop leaking too many
fragmentation ids with this change.

Thanks,
Hannes