Re: NAT failure with TCP, too

[Bernardo Innocenti <bernie@develer.com>]

Cristian Grigoriu wrote:
> Hi,
> 
> I can confirm the same bug you reported here 
> http://www.ussg.iu.edu/hypermail/linux/kernel/0509.2/0279.html
> This time it happens with TCP connections originating from the same TCP 
> port (1234) from multiple machines. The SNAT simply doesn't take place 
> and the normal routing occurs.
> 
> Kernel is Debian stock 2.6.18-1.
> 
> Please let me know if you have find a workaround.

It turned out that the real thing that was triggering the bug
for me was unloading and reloading the ip_nat module without
also reloading ip_conntrack.

The connection tracking tuple would remain in the kernel, visible
in /proc/net/ip_conntrack, but no longer linked to the SNAT rule.
I'd consider this a bug, but very few users will ever be affected.

The workaround for me was to remove my hand-cracted iptables
rules from ppp's ip-up.local and move them to the distro-supplied
iptables firewall instead.  The only downside is that I must now
hardcode the destination ip of the SNAT rule because it's too early
to read the interface address of ppp0.

-- 
   // Bernardo Innocenti - Develer S.r.l., R&D dept.
 \X/  http://www.develer.com/