Re: ip_conntrack table full after upgrade from RHEL3 (2.4/1.2.8) to RHEL4U4 (2.6.9/1.2.11)

[Pascal Hambourg <pascal.mail@plouf.fr.eu.org>]

Hello,

Alex Tang a écrit :
> 
> I've been having a problem on a machine that does a high volume of 
> sendmail traffic.  The machine gets approx 50,000 connections per hour 
> to port 25.
> The machine was upgraded from a RHEL3 based system (kernel rpm 
> 2.4.21-47.EL and iptables rpm 1.2.8-12.3) to a RHEL4 based system 
> (kernel rpm 2.6.9-55.EL and iptables rpm 1.2.11-3.1.RHEL4).

You should really consider upgrading to something more recent, because 
kernel 2.6.9 and iptables 1.2.11 are *badly* outdated.

> Since the upgrade has occurred, the conntrack table fills up relatively 
> fast (within one day).  The max size is 65536 (as per 
> /proc/sys/net/ipv4/netfilter/ip_conntrack_max).
> 
> I've been searching through the archives, faq, etc and have found the 
> usual standard answer is to increase the ip_conntrack_max.  However, I'm 
> concerned for a couple of reasons that this may not be the proper answer.
> 
> In particular, i have another machine which is still running the RHEL3 
> (kernel 2.4.21-47.EL/iptables 1.2.8-12.3), that gets more connections 
> per hour (80,000 vs. 50,000), and there are only about 9000 entries in 
> the ip_conntrack table on that machine.
> 
> The problem with the conntrack table filling up fast started as soon as 
> we did the upgrade.
> 
> Also, on the machine that is currently experiencing problems, most (98%) 
> of the connections are in the ESTABLISHED state, however the majority of 
> these connections are not seen when doing a "netstat".
> 
> I admit that I do not fully understand the details of the iptables 
> implementation, but it seems that the connection close is not being 
> "seen" by the conntrack code and connections that have already gone away 
> are still in the ip_conntrack table, and we have to wait for these 
> connections to "timeout" before they are expired from the conntrack table.

And the default timeout for established TCP connections is 5 days, so it 
takes a looong time.

A noticeable difference between kernel 2.6.9 and earlier versions is the 
TCP window tracking, which was added in 2.6.9. It adds TCP sequence 
number checking to the connection tracking, so any TCP packet with an 
out-of-window sequence number is tagged INVALID. If the TCP window 
tracking considers that the FIN packet sequence numbers are 
out-of-window for whatever reason, this may be the reason why the 2.6.9 
kernel connection tracking keeps old connexions. You may add LOG 
iptables rules to track TCP FIN packets states.

> I could of course, increase the max size of the table, or decrease the 
> /proc/sys/net/ipv4/netfilter/ip_conntrack_tcp_timeout_established, but 
> it seems that would only mask the problem, not actually fix it.

I agree. You could also decrease the value of 
/proc/sys/net/ipv4/netfilter/ip_conntrack_tcp_timeout_established to 
something much shorter than 5 days but again this would only mask the 
problem.

Try to set /proc/sys/net/ipv4/netfilter/ip_conntrack_tcp_be_liberal to 
1. This setting makes TCP window tracking more liberal, so only 
out-of-window TCP RST packets are tagged INVALID.

Note that later kernel versions provided some bugfixes for the TCP 
window tracking that may fix this issue.