[tproxy] Transparent proxying - different approach - NAT on demand.

[KOVACS Krisztian]

Hi,

On sze, jan 21, 2009 at 01:42:38 +0100, NTPT wrote:
> Hi all.
> 
> I have some proposition about squid and tproxy. I call it "NAT ON
> DEMAND". I am not really a good C programmer or kernel hacker, but...
> 
> AFAIK in early tproxy days a dynamically created SNAT and DNAT rules,
> was considered but this idea was abandon as unfeasible. 
> 
> My idea is to allow "setting nat on application explicit demand". To
> create a infrastructure in with userspace application (squid for
> example) can send a "demand"  for source address to witch their outgoing
> connections should be natted to - on per socket basis.Usage in
> transparent proxying of all kind. Especially useful for SQUID and realy
> transparent proxy scenario.
>  
> 
> The idea is roughly inspired and is analogic to  "setting FWMARK from
> userspace", an old kernel patch that implement ioctl for setting fwmark
> on local originated packets on per socket basis. See
> http://oss.sgi.com/archives/netdev/2001-12/msg00053.html
> 
> 
> 1:applications should be able  to explicitly set on per socket basis, on
> wich IP address they need outgoin connection to be SNATed. Probably add
> one 32bit item to skbuff and appropriate mechanizm for setting it from
> userspace
> 
> 2: change netfilter / iptables to allow "-j SNAT --to-demand" option to
> witch set a source of the connection to address that application
> demanded
> 
> 3: add "-m natdemand" iptables match  to  find if local originator of
> this connection demand NAT for it (ie if the additional field in skbuff
> is not 0x0 ).
> 
> 4: patch squid to request a nat for its outgoing connections. Only thing
> that it should to take a client ip address and "demand it"
> 
> Usage like this
> 
> in squid conf
> 
> acl src 192.168.1.0/24      transparent_clients # set ACL for clients
> network that we need handled transparently nat_on_demand
> transparent_clients          # demand specific SNAT for all outgoing
> connections (ie connections that fetching content) it made on proxying
> clients from this network. ie take IP of the squid client and demand
> SNAT to this IP for outgoing sockets that fetching data for that client
> on cache miss (sorry my bad english)  
>  
> in iptables:
> 
> iptables -t nat -I OUTPUT (or POSTROUTING ?) -m natdemand -j SNAT
> --to-demand # to tell netfilter to SNAT source address of that
> connection to what application demanded. ie copy address stored in
> additional skbbuff field 

You've just described how tproxy 2 worked, with a slightly different
interface.

> From that point, all the hard work for us is done  automatically by
> netfilter.No need to worry about fancy routing setup etc. No need
> additional overhead for socket match. Does not interfer with connection
> tracking as tproxy afaik does. Should work for related (icmp etc)
> packets as well. Should instantly work for every protocol that have nat
> helper too. Should be easy to implement  and less intrusive (afaik most
> of tproxy patch is to circumvent some kernel safety measurements and
> control checks on various places)

I'd expect the overhead of Netfilter NAT *much* larger than the overhead
of the socket match. Also, the need for the 'fancy' routing setup is just
because we would have liked to avoid intrusive changes to the kernel.

We had a patchset with some changes in the core ipv4 routing code so that
absolutely no modifications were necessary to the routing config. We
transitioned to the new approach because

  * it required no changes to routing code, and thus had a better
    possibility of merging upstream,

  * it's slightly more flexible (ie. now that the kernel can do policy
    routing for IPv6 implementing tproxy support for that protocol should
    be easier.

As far as I know the current upstream implementation does not have any
problems with (related) ICMP and does not usually interfere with
connection tracking. (Unfortunately, because of the way conntrack works
you cannot have full connection tracking and do transparent proxying in
certain corner cases.)

Most of the modifications of the tproxy patch weren't necessary to
'circumvent kernel safety measurements and control checks', but to make it
possible to have a TCP endpoint with a non-local IP address in the Linux
stack. It *almost* worked out-of-the-box, it's just that the code
contained a few checks which interfered with tproxy. This does not mean in
any way that the security of the network stack is breached, however: you
need to be highly privileged to set the IP_TRANSPARENT flag to make the
kernel omit those checks in the IPv4 output path.

On the contrary, there were all kinds of complications with tproxy2, all
coming from the fact that the local socket table and the connection
tracking database of the kernel were out-of-sync (by design). Having been
there and done that, I don't think we should go back to that approach...

-- 
KOVACS Krisztian