High Performance Firewall
Imagine this, you have more than two networks separated by Virtual Lans protocols (IEEE 802.1q) or VLANs, carried to you by an intelligent/manageable switch on one troncal line 10/100/1000 MB HD/FD (naturally the best is 1000 MB FD).
You have to share internet to a really BIG numbers of hosts, and maintain a good performance. The first choice is to separate the networks into a equal numbers of ports and maybe a more numbers of firewalls machines. This is not really cost effective, but works.
The second one is what I did. The history of how this begin is related to a some emergency/burn/crash/out of a group of Cisco PIXs. I won't go too deeply into that.
- I have about 4 networks mask 21 i.e a 8 class C each!!! This is a lots of MACs addresses and even more dangerous a lot of BROADCAST. This is insane but this is the way that my company works.
- They gave me 30 publics IPs addresses in 3 groups.
- And a machine, a little one, not a big deal.... just for testing. Later (about 8 months) we changed the PC. We put in a real monster.
- Later we added a class C network with a routed lot of increasing subnets.... PLUS!!!!
The first thing we have to do is give the kernel the capacity to work with Jumbo Frames. This is done by adding the 8021q module to the kernel
# modprobe 8021q
and/or put in modules in /etc/rc.conf
Next we have to create the virtuals NICs with this command vconfig. Let's suppose we have vlans 20,30,40 and 50 working in our core network.
# vconfig add ethX 20 # vconfig add ethX 30 ... # vconfig add ethX 50
Where ethX is the troncal NIC Now, if we want to see the interfaces just put ifconfig -a and we will get a list.
So .... next just set the private address as you want. For example (using eth1 like a troncal)
# ifconfig eth1.20 192.168.0.1 netmask 255.255.248.0 # ifconfig eth1.30 192.168.8.1 netmask 255.255.248.0 ... etc.
I will not explain the number of host neither the mask.... I do a kind of daemon with this, a vlan.conf file in /etc and a vland in /etc/rc.d... I can share this if you want?
It is really easy to make a firewall/nat with iptables there is a lot of information around there...
Just take in mind that you will work with a lot of traffic, this means a lot of CPU usage, so keep your rules tiny, just those necessary. Accept all by default just want to make NAT, no more.... Maybe it is good to accelerate some ports, (80,443,25,110,21,20,53 etc) remember, every packet running through our firewall will past through every rule until it matches to one or fall in the pool (default policy).
The round robin NAT
Let's suppose we have a one ip: 200.aaa.bbb.6 and our gateway is 200.aaa.bbb.1. we can safely put these parameters by default in our configuration. It will not get participation at all in our firewall.
I say I have 3 groups of 10 IPs each to play...... we'll define the NEXT in our firewall script:
Gr1='200.AAA.CCC.10-200.AAA.CCC.20' Gr2='200.AAA.DDD.10-200.AAA.DDD.20' Gr3='200.AAA.EEE.10-200.AAA.EEE.20'
And the next important line is:
iptables -t nat -A POSTROUTING -s 192.168.0.0/21 -j SNAT --to $Gr1 #ACCESS VLAN 10 iptables -t nat -A POSTROUTING -s 192.168.8.0/21 -j SNAT --to $Gr2 #ACCESS VLAN 20 iptables -t nat -A POSTROUTING -s 192.168.15.0/21 -j SNAT --to $Gr1 #ACCESS VLAN 30 .... etc
You can repeat the groups for access, subdivide the networks ETC, iptables make a round robin over the Gr1, Gr2 and Gr3 by default, no modification is needed.
It's not necessary to create a virtual card (alias) to every IP in the group.
It's important that every real router knows every group and publishes its via BGP (or similar) to the neighbours.
To accelerate some ports you can put this in the top of FORWARD chain
iptables -A FORWARD -m state --state ESTABLISHED,RELATED -j ACCEPT iptables -A FORWARD -p icmp -o eth0 -j ACCEPT iptables -A FORWARD -p tcp -m multiport --dports 80,443,110,53 -j ACCEPT # FAST FAST FAST iptables -A FORWARD -p udp --dport 53 -j ACCEPT
- the packets incoming will pass only 1 rule if it is an establish connection
- the packet incoming will pass 2 rules if is a ping or similar
- the packet will pass 3 rules if is http, mail or similar
- and the DNS request will pass 3 o 4 rules until go out
The outgoing virus will KILL our machine, and we not need to share "windows" conversations so, kill them!!!!
#VIRUS iptables -A FORWARD -p tcp --dport 135:139 -j DROP iptables -A FORWARD -p tcp --dport 445 -j DROP iptables -A FORWARD -p udp --dport 135:139 -j DROP iptables -A FORWARD -p udp --dport 445 -j DROP
If you can, before they reach our machine.
The High Performance
We get to the real important part of this howto.
In our run to get a really big number of hosts running through our machine we miss some things
- We forget that is just one NICs to potentially more than 8000 Mac Addresses. The card shared memory is not prepare for this!!!!!
- By default iptables is not prepared to make this number of connections simultaneously !!!!!!
To the first issue... I get some error messages in the logs relative to this, I'm really sorry, I lost these logs and don't remember what they said. But the answer is this, increase the threshold memory to the neighbours. Type this and read:
# cat /proc/sys/net/ipv4/neigh/default/gc_thresh1 128 # cat /proc/sys/net/ipv4/neigh/default/gc_thresh2 512 # cat /proc/sys/net/ipv4/neigh/default/gc_thresh3 1024
Next you can put this in the /etc/sysctrl.conf
net.ipv4.neigh.default.gc_thresh1 = 512 net.ipv4.neigh.default.gc_thresh2 = 1024 net.ipv4.neigh.default.gc_thresh3 = 2048
and make sysctl -p to increase to the double!!! (no reboot needed) with this I get no errors!!!!!
The next part will need some comprehension about buckets and conntracks and hashsize (the way how iptables manage the nat connections). There is a very good document about this at here. Read it!!!! Some thing are change since IPtables is know as Netfiler.
In resume!!! Put this in your modules section:
MODULES=(8021q 'nf_conntrack hashsize=1048576' nf_conntrack_ftp ...and other nf_stuff .......)
The last ones is just to avoid some problems that we have with ftp connections (I thing this is not necessary anymore). The 'nf_conntrack hashsize=1048576' increase the numbers of the hashsize (increase the kernel memory designated to NAT connections) (need reboot or reload module :-) see with dmesg | grep conntrack)
And the next is put some similar to the /etc/sysctrl.conf file
... net.netfilter.nf_conntrack_max = 1048576 ...
And do the sysctl -p command
In my case is the same number, that means that I have 1 connection for bucket!!!! I don't need more!!!! by default NetFilter put rate of 1:8. I.E. 8 conections per bucket!! (I think, not remember well)..
In our case we get about 600.000 simultaneous connections in 2 1Giga NICs cards, You can see this with the next command
# cat /proc/sys/net/netfilter/nf_conntrack_count
And put this in a snmpd agent to get and graph it in a MRTG/cacti server ..... uuuuuuu homework
A sample here. (Not actualized, not ever here, but I'll try.)
We have 3 big access to Internet!!! This is because we manage 3 class C groups of IPs (some restrictions of BGP) in this firewall. So, we have 3 incoming traffics that we can manage, but only one outgoing!!! Our default gateway. This can easily fill our outgoing quote, so we have to spare it.
First we have to put some new tables to /etc/iproute2/rt_tables file
# echo 200 PRO_1 >> /etc/iproute2/rt_tables # echo 205 PRO_2 >> /etc/iproute2/rt_tables # echo 210 PRO_3 >> /etc/iproute2/rt_tables
Can be more, can be less, depends on traffic
Second we have to give a default gateway to this tables
# ip route add default via 200.aaa.bbb.2 table PRO_1 # ip route add default via 200.aaa.bbb.3 table PRO_2 # ip route add default via 200.aaa.bbb.4 table PRO_3
It's recommended but not necessary put the local interfaces to each table. If you don't put the next few lines you will get not answer of ping in the local network, but you will be able to pass trough.
# ip route add 192.168.0.0/21 via 192.168.0.1 table PRO_1 # ip route add 192.168.8.0/21 via 192.168.8.1 table PRO_1 # ip route add 192.168.15.0/21 via 192.168.15.1 table PRO_1 ..... same PRO_2, same PRO_3
The last thing is to give the order to the incoming packages
# ip rule add from 192.168.0.0/21 table PRO_1 .... ....
Again, you can play with the PRO_X and even you can play with the mask and submask For example we want to give only a one class C to outgoing to PRO_3
# ip rule add from 192.168.1.0/24 table PRO_3
Put this before the <NET>/21
and then TEST IT!!!!
pick a WinPC in one of the private networks and make a tracert to somewhere!!!
Before this, you can browse to a some MYIPsite like www.whatismyip.com and get your "RightNow" address, test it later and get some other address ETC...
Actually we are using a PC with 2G ram , Intel Pentium 4 2.8GHz Dual and 2 NICs 1Gb (one integrated one PCI-E), the top CPU usage don't get the 80% pic...
- I got a really big surprise when upgrade net-snmp to 5.4.1 (to get statistics)on the Firewall, the CPU usage goes down a lot... from 80% to almost 3%... YES I mean it... I got a confirmation on the other machine (the bridge one, I mention it below) so, we have a much more machine to keep increasing the traffic... this is very good news.
Out traffics is about 180 MB OUT and maybe the half IN.
We get some lost in the cards, about 1 packets per second, it's really low compared with the traffic, so we just don't care.
- Now we have no errors and no packets lost.
Nevertheless we get a big problem, the MAIL/SPAM outgoing, provoked by VIRUS/MAD_GUYs, can block all the publics IP in a really short time. (in the public Black Lists).
The answer to this may not be ethical, but we have to do it.
We put a similar machine before this one with a bridge configuration blocking every port 25 traffic and sending to another anti SPAM/VIRUS/MAIL machine.
To do this is necessary to give to the bridge the capacity of VLANs and one virtual card (in bridge mode i.e.'br0.30') in every net to answer to the sender...
It works and the customer doesn't notice the difference. Just if you do a telnet anywhere.net 25 you get a Hi I'm a scanner hello answer!!!.. It's really good!!! and works, the mail outgoing is sender from another IP, so we get our 3 groups safe!!!.