A NATural Progression

Editor's Note: Due to a printer error, David Bandel's article on iptables building was not complete in the magazine. We present it here in it's entirety.

One of the best tools at our disposal within the Netfilter framework is NAT. NAT allows us to obfuscate (but not hide) our true network, forcing would-be black hats to work harder (and possibly go after easier targets). It also permits us to make the best use of limited IPs. Last month [see “Netfilter 2: in the POM of Your Hands” in the May 2002 issue of LJ] we looked at extending iptables to include experimental or beta matches and targets. This month we look at doing NAT the correct way, take a closer look at one or two other matches, then see what some of the more common errors are when constructing iptables rules and how to avoid them.

Given the shortage of usable IPv4 addresses left today, it's likely your ISP didn't provide you enough IPs to run all your systems. If you're lucky, you got more than half a dozen you could actually use. But if you don't use them, you'll lose them. And you don't want that, or there's no room for expansion tomorrow. So you're going to make your firewall look like as many systems as you have IPs to use. How to do that? The easiest way is to assign all your IPs to one NIC (the one connected to your ISP) and SNAT connections so they look like they come from each IP in turn (this example assumes your internal network is 192.168.0.0/24, which is bound to eth1, and your usable IPs are 209.127.112.26-209.127.112.30, which are bound to eth0):

iptables -t nat -A POSTROUTING
-o eth0 -s
--to-source 209.127.112.26-209.127.112.30

Now iptables will NAT the first connection to .26, the second to .27, the third to .28 and so on, wrapping around to .26 after the connection to .30.

One word of caution: test this before you deploy it. I've had one router that didn't like seeing multiple IPs sourced from the same MAC address. It would pass the first connection, but subsequent connections would time out. The router's built-in firewall (which couldn't be turned off by the client) most likely thought the other packets were spoofed and was silently dropping them.

Let's make sure we accept all outgoing connections but only accept incoming connections that are related to these outgoing connections:

iptables -t filter -A FORWARD -i ! eth0 -m state
--state NEW,ESTABLISHED,RELATED -j ACCEPT
iptables -t filter -A FORWARD -i eth0 -m state
--state ESTABLISHED,RELATED -j ACCEPT
iptables -t filter -A FORWARD -i eth0 -m state
--state NEW,INVALID -j DROP

Now that we've managed outgoing traffic, let's assume we've moved all our services inside this firewall. We'll further assume they are all inside eth1 on the 192.168.0.0/24 network. Each service has two IPs associated with it: an external IP that the world sees and an internal IP that we see. Specifically, we'll assign the following:

Because of our iptables state table rules above, each service (or more specifically, each port that corresponds to that service) will not only have to be forwarded through the firewall to the correct IP inside, but we'll need a rule to accept that NEW traffic. Starting with Apache, which in our case uses both ports 80 and 443 (for SSL), we have:

iptables -t filter -I FORWARD -i eth0
-d
iptables -t nat -A PREROUTING -d
-p tcp --dport 80 -j DNAT --to-destination
iptables -t filter -I FORWARD -i eth0
-d 209.127.112.26 -p tcp --dport 443 -j
ACCEPT
iptables -t nat -A PREROUTING -d
--dport 443 -j DNAT --to-destination 192.168.0.4

Notice that we had to insert a rule in the FORWARD chain. This is because we already have a more general rule that would have dropped NEW connections. We can insert a rule anywhere in a chain, but if we don't specify where, the default is to insert it as the first rule. Normally, this will not be a problem and will put our specific rules ahead of our general rules.

Note that we also specified the IP on which this connection should show. This is not necessary because connections showing up on other IPs will be handled by the state table and dropped, unless we've done something really questionable and started a web server on our firewall. If no ports on our firewall are open, we're okay. We always can protect them just in case by ensuring we also have stateful rules for our INPUT chain:

iptables -t filter -A INPUT -i ! eth0 -m state
--state NEW,RELATED,ESTABLISHED -j ACCEPT
iptables -t filter -A INPUT -i eth0 -m state
--state RELATED,ESTABLISHED -j ACCEPT
iptables -t filter -A INPUT -i eth0 -m state
--state NEW,INVALID -j DROP

The first rule above allows NEW, ESTABLISHED and RELATED connections from lo (the localhost interface) as well as any internal devices, omitting only our external device, which is dealt with in the next two rules.

Next we look at the FTP connection. This is straightforward and exactly the same as the above rules, but for port 21:

iptables -t filter -I FORWARD -i eth0 -p tcp
--dport 21 -j ACCEPT
iptables -t nat -A PREROUTING -i eth0
-d
--to-destination 192.168.0.5

We don't have to worry about the FTP-data channel (port 20) because our FTP server opens it outgoing, and our state rules will pass this new connection out.

Now it gets a little more difficult. DNS works on both UDP for normal queries and TCP for zone transfers. If we don't want to allow zone transfers to the outside, we only open UDP. If we want to allow zone transfers, then we have to allow both. Assuming we want to allow both, we know that we can specify UDP, TCP or ICMP as protocols. You must specify -p (protocol) in order to specify a port. If you want both UDP and TCP, you should be able to say “not ICMP”, and the other two are assumed automatically. Unfortunately, it doesn't work that way.

When you specify a protocol, even if you say -p ! ICMP, it's the ICMP match module that is loaded, not the TCP and UDP match modules. So you'll get an error message when you specify a port. This is a danger with using a negative match; the match module that is loaded is the module specified, not the modules you may assume are loaded. You must specify positively each match you want so the corresponding match module is loaded.

For now, let's assume you are only interested in opening the UDP port:

iptables -t filter -I FORWARD -i eth0
-d
iptables -t nat -A POSTROUTING -i eth0
-d
--to-destination
iptables -t filter -I FORWARD -i eth0
-d
iptables -t nat -A POSTROUTING -i eth0
-d
--to-destination 192.168.0.8

And finally, we need to deal with our mail host:

iptables -t filter -I FORWARD -i eth0
-d 209.127.112.28 -i eth0 -p TCP --dport 25 -J
ACCEPT
iptables -t nat -A POSTROUTING -i eth0
-d
--to-destination
iptables -t filter -I FORWARD -i eth0
-d 209.127.112.29 -i eth0 -p UDP --dport 25 -J
ACCEPT
iptables -t nat -A POSTROUTING -i eth0
-d
--to-destination 192.168.0.9

If we test our outgoing mail using mail -v user@another.dom, our firewall will grab one of 209.127.112.25-209.127.112.30. If our DNS records say our MX host is 209.127.112.28, then we have only a 20% chance of grabbing that IP and an 80% chance that upstream mail hosts will bounce our mail as not being from a host with a DNS MX RR—not good.

So how do we fix this? If we have the luxury of adding all our IPs as MX hosts, that would solve part of the problem, but then upstream hosts might spend time connecting to IPs of ours that don't DNAT the mail through. And we really don't want all those IPs to appear as MX IPs.

The correct response is to add a more specific SNAT rule ahead of the general SNAT rules that will handle outgoing traffic on port 25. The danger here is that if we can't trust our internal users, we also must make sure that internal users can't abuse port 25. So we'll add three rules for outgoing traffic, one to SNAT port 25 traffic coming from our primary mail server (192.168.0.8) only out through 209.127.112.28, and two to block port 25 traffic from all other internal addresses except our true mail host:

iptables -t filter -I FORWARD -i eth1
-s
iptables -t filter -I FORWARD -i eth1 -p tcp
--dport 25 -s
-I POSTROUTING -o eth0 -p tcp --dport 25
-s 192.168.0.8 -j SNAT --to-source 209.127.112.28

Some of you may think: hah! caught him. The first two rules above are reversed. Well, yes, they are. But that's because we're inserting them one at a time as the first rule, so rule two above will really be rule one in the FORWARD chain after running them both. The SNAT rule is in another chain, so it could have gone anywhere. Also, I suggest you make sure the first rule above is correct for your system. If the untrusted network is 192.168.0.0/24, and the trusted network is 192.168.1.0/24, you may need the source (-s) to be 192.168.0.0/23 to cover both. Or, perhaps just drop the -s option and match on the inbound interface (-i).

I suggest that the best way to build a firewall is to walk through each chain to see where and how (or even if) a particular packet will be handled. Don't do as we've been doing here inserting rules seemingly willy-nilly. Build your chain on paper with all the rules in the correct order. Then you won't make mistakes. You can always check afterward to make sure the rules are as you think you wanted them with: iptables -t <table> -L -nv. The above rule with the -v included will show you how many packets and how many bytes are affected by this rule. If, after a week has gone by, you still have rules with 0 bytes affected by it, you might want to relook at that rule's position in the chain. But just because a rule has affected packets doesn't mean it's in the right place. It may have only affected half the packets that really should have been affected.

I'm waiting for someone to write “the killer app” for Netfilter, and that would be a utility that runs tests, analyzes the rules and allows you to move them around and test again. But until that day comes, you'll have to do it by hand.

Some of you may have noticed that I make heavy use of -i eth0, or -i ! eth0, but in general match an interface. Often, you can probably see that this isn't necessary because I've limited the source IP address or some other part of the packet header that pretty much ensures matching what we want. But I do this for a particular reason. I turn off rp_filters (reverse path filters). These tend to interfere with legitimate VPN packets. Besides, Linux's rp_filter is nowhere near as granular as iptables.