If you saw my restore-mark query last week, but thought "I'll answer if nobody else does", please note that nobody else did. :-) Here's my query again, in case you've lost the original. Ciao, Sheldon. -----Forwarded Message----- From: Sheldon Hearn <sheldonh@xxxxxxxxxxxxx> To: netfilter@xxxxxxxxxxxxxxxxxxx Subject: CONNMARK restore-mark creates conntrack entry? Date: Fri, 28 May 2004 15:55:08 +0200 Hi folks, It seems that, in linux-2.6.5 w/ CONNMARK from patch-o-matic 20040302, the CONNMARK --restore-mark modifier causes a conntrack entry to be created if it doesn't already exist. I'm using CONNMARK to remember customer-initiated connections and route them differently from "wild world"-initiated connections. I don't want to mark "wild world"-initiated connections, because a SYN flood then becomes a trivial way to DOS the conntrack table, and thus the firewall. Any idea how to stop this happening? Here are the PREROUTING rules in question: *mangle :PREROUTING ACCEPT :setup-mark - :restore-mark - // // IFcic is an interface used to short-circuit the load balanced group // completely. // -A PREROUTING -i IFext -d HOSTself -j RETURN -A PREROUTING -i IFcic -p tcp --syn -j setup-mark -A PREROUTING -i IFext -p tcp ! --syn -j restore-mark // // How UDP is handled isn't pertinent. // ... // -A PREROUTING -i IFext -p tcp -j MARK --set-mark MARKlvs // // The setup-mark chain is responsible for keeping customer-initiated // connections out of LVS. The restore-mark chain is responsible for // routing return traffic for those connections inward. // -A setup-mark -j MARK --set-mark MARKint -A setup-mark -j CONNMARK --save-mark -A setup-mark -j MARK --set-mark 0 -A setup-mark -j ACCEPT -A restore-mark -j CONNMARK --restore-mark -A restore-mark -m mark --mark MARKint -j ACCEPT So I should end up with packets marked in one of 3 ways: No mark: route out MARKint: route in via IFcic MARKlvs: route in via the load-balanced cluster This works well. I just don't want conntrack entries created by the SYNACK from clients in "wild world"-initiated connections, when the restore-mark chain is traversed. Any ideas? Thanks, Sheldon.
