State Mismatch and tcp.closed

[Martin Turgeon]

Hi list!

I had a problem with state mismatch on my DB server that I solved by 
lowering the tcp.closed timeout. I setted it to 2 instead of 90.

I now have what looks like the same problem on the front-end web server. 
However, when I tried to apply the same fix, I got connection problem 
with the back-end DB, but the state mismatch disappearred.

On the front-end web server, the state mismatch occurs on the external 
interface, only on port 80.

I enabled misc debugging and got this in /var/log/messages on the 
front-end web server:

May 28 05:02:19 francis kernel: pf: BAD state: TCP 127.0.0.25:80 
206.125.166.65:80 98.207.239.10:54737 [lo=820536733 high=820603340 
win=65535 modulator=0 wscale=0] [lo=2871317100 high=2871375106 win=8326 
modulator=0 wscale=3] 7:4 R seq=820536733 (820536732) ack=2871317100 
len=0 ackskew=0 pkts=43:69 dir=in,fwd
May 28 05:02:19 francis kernel: pf: State failure on:         |
May 28 05:02:19 francis kernel: pf: BAD state: TCP 127.0.0.25:80 
206.125.166.65:80 98.207.239.10:54733 [lo=374985971 high=375052578 
win=65535 modulator=0 wscale=0] [lo=2999164748 high=2999229169 win=8326 
modulator=0 wscale=3] 7:4 R seq=374985971 (374985970) ack=2999164748 
len=0 ackskew=0 pkts=40:54 dir=in,fwd
May 28 05:02:19 francis kernel: pf: State failure on:         |
May 28 05:03:06 francis kernel: pf: BAD state: TCP 127.0.0.20:80 
206.125.166.80:80 123.116.84.41:59776 [lo=3407758259 high=3407823796 
win=4096 modulator=0 wscale=2] [lo=374200006 high=374216390 win=8192 
modulator=0 wscale=3] 4:2 A seq=3407758259 (3407758260) ack=2320196160 
len=0 ackskew=-1945996154 pkts=1:1 dir=in,fwd
May 28 05:03:06 francis kernel: pf: State failure on:     3   |
May 28 05:03:06 francis kernel: pf: BAD state: TCP 127.0.0.20:80 
206.125.166.80:80 123.116.84.41:59776 [lo=3407758259 high=3407823796 
win=4096 modulator=0 wscale=2] [lo=374200006 high=374216390 win=8192 
modulator=0 wscale=3] 4:2 RA seq=3407758259 (3407758260) ack=2320196160 
len=0 ackskew=-1945996154 pkts=1:1 dir=in,fwd

This server has been up for 12 days and already got almost 600000 state 
mismatch!

I tried to lower tcp.finwait, no result. I tried to set optimization to 
aggressive, no result. I tried to disable port randomization via sysctl, 
no result either.

I tcpdumped and there is only a few RST so I don't understand why 
tcp.closed would solve my problem. If it's a problem with source port 
reuse, tcp.finwait should be the timeout that would help, not 
tcp.closed, right?

How can a lower tcp.closed on the front-end cause mysql connection 
problem with the back-end? I tcpdumped while there is a connection 
problem with the DB and there is nothing that seems wrong, no RST at 
all! The front-end web server tries to connect to the DB, wait 3 sec and 
if it fails to establish a connection, it then tries to connect to a 
read-only backup DB, on another server, which never fails to connect.

The only thing I'm sure is that it's the tcp.closed that cause the DB 
connection problem. As soon as I remove it, the state mismatch comes 
back on the external interface but there's no DB connection problem anymore.

What am I missing?

Martin