There are two main uses for multiple Internet links connected to the same internal network. One common use is to select an outbound link based on the type of outbound service. The other is to split traffic arbitrarily across multiple ISPs for reasons like failover and to accommodate greater aggregate bandwidth than would be available on a single uplink.

If your need is the latter, please consult the documentation on the LARTC site, as it does a good job of summarizing the issues involved and describes how to accomplish this. This type of use of multiple Internet connections means that (from the perspective of the linux routing device), there is a multipath default route. The LARTC documentation remarks that Julian Anastasov's patches "make things nicer to work with." The patches to which the LARTC documents are referring are Julian's dead gateway detection patches (at least) which can help the linux routing device provide Internet service to the internal network when one of the links is down. See here for Julian's route work.

In the remainder of this section, we'll discuss how to classify traffic for different ISPs, how to handle the packet filtering for this sort of classification scheme, and how to create routing tables appropriate for the task at hand. If anything at all seems unclear in this section, you may find a quick re-reading of the advanced routing overview quite fruitful.

The simplest way to split Internet access into two separate groups is by source IP of the outbound packet. This can be done most simply with ip rule and a second routing table. We'll assume that masq-gw in the example network gets a second, low cost network connection through a DSL vendor.

The DSL IP on masq-gw will be 67.17.28.12 with a gateway of 67.17.28.14. We'll assume that this is for outbound connectivity only, and that the IP is active on eth4 of the masq-gw machine. Before beginning let's outline the process we are going to follow.

Here's a short snippet of shell which you may find handy for copying one routing table to another; see the full script for a more generalized example.

[root@masq-gw]# ip route show table main
192.168.100.0/30 dev eth3  scope link
67.17.28.0/28 dev eth4  scope link
205.254.211.0/24 dev eth1  scope link
192.168.100.0/24 dev eth0  scope link
192.168.99.0/24 dev eth0  scope link
192.168.98.0/24 via 192.168.99.1 dev eth0
10.38.0.0/16 via 192.168.100.1 dev eth3
127.0.0.0/8 dev lo  scope link 
default via 205.254.211.254 dev eth1
[root@masq-gw]# ip route flush table 4
[root@masq-gw]# ip route show table main | grep -Ev ^default \
>   | while read ROUTE ; do
>     ip route add table 4 $ROUTE
> done
[root@masq-gw]# ip route add table 4 default via 67.17.28.14
[root@masq-gw]# ip route show table 4
192.168.100.0/30 dev eth3  scope link
67.17.28.0/28 dev eth4  scope link
205.254.211.0/24 dev eth1  scope link
192.168.100.0/24 dev eth0  scope link
192.168.99.0/24 dev eth0  scope link
192.168.98.0/24 via 192.168.99.1 dev eth0
10.38.0.0/16 via 192.168.100.1 dev eth3
127.0.0.0/8 dev lo  scope link 
default via 67.17.28.14 dev eth4
        

Now, exactly what have we just done? We have created two routing tables on masq-gw each of which has a different default gateway. We have successfully accomplished the first part of our preparations.

Now, let's mark the traffic we would like to route in using conditional logic. We'll use iptables to select traffic bound for destination ports 80 and 443 originating in the main office desktop network.

[root@masq-gw]# iptables -t mangle -A PREROUTING -p tcp --dport 80 -s 192.168.99.0/24 -j MARK --set-mark 4
[root@masq-gw]# iptables -t mangle -A PREROUTING -p tcp --dport 443 -s 192.168.99.0/24 -j MARK --set-mark 4
[root@masq-gw]# iptables -t mangle -nvL
Chain PREROUTING (policy ACCEPT 0 packets, 0 bytes)
 pkts bytes target     prot opt in     out     source                destination         
    0     0 MARK       tcp  --  *      *       192.168.99.0/24       0.0.0.0/0          tcp dpt:80 MARK set 0x4 
    0     0 MARK       tcp  --  *      *       192.168.99.0/24       0.0.0.0/0          tcp dpt:443 MARK set 0x4 

Chain OUTPUT (policy ACCEPT 0 packets, 0 bytes)
  pkts bytes target     prot opt in     out     source               destination
[root@masq-gw]# iptables -t nat -A POSTROUTING -o eth4 -j SNAT --to-source 67.17.28.12
[root@masq-gw]# iptables -t nat -A POSTROUTING -o eth1 -j SNAT --to-source 205.254.211.179
Chain PREROUTING (policy ACCEPT 0 packets, 0 bytes)
 pkts bytes target     prot opt in     out     source               destination         

Chain POSTROUTING (policy ACCEPT 0 packets, 0 bytes)
 pkts bytes target     prot opt in     out     source               destination         
    0     0 SNAT       all  --  *      eth4    0.0.0.0/0            0.0.0.0/0          to:67.17.28.12
    0     0 SNAT       all  --  *      eth1    0.0.0.0/0            0.0.0.0/0          to:205.254.211.179

Chain OUTPUT (policy ACCEPT 0 packets, 0 bytes)
 pkts bytes target     prot opt in     out     source               destination
        

With these iptables lines we have instructed netfilter to mark packets matching these criteria with the fwmark and we have prepared the NAT rules so that our outbound packets will originate from the correct IPs.

Once again, it is important to realize that the fwmark added to a packet is only valid and discernible while the packet is still on the host running the packet filter. The fwmark is stored in a data structure the kernel uses to track the packet. Because the fwmark is not a part of the packet itself, the fwmark is lost as soon as the packet has left the local machine. For more detail on the use of fwmark, see Section 10.3.2, “Using fwmark for Policy Routing”.

iproute2 supports the use of fwmark as a selector for rule lookups, so we can use fwmarks in the routing policy database to cause packets to be conditionally routed based on that fwmark. This can lead to great complexity if a machine has multiple routing tables, packet filters, and other fancy networking tools, such as NAT or proxies. Caveat emptor.

A convention I find sensible is to use the same number for a routing table and fwmark where possible. This simplifies the maintenance of the systems which are using iproute2 and fwmark, especially if the table identifier and fwmark are set in a configuration file with the same variable name. Since we are testing this on the command line, we'll just make sure that we can add the rules first.

[root@masq-gw]# ip rule add fwmark 4 table 4
[root@masq-gw]# ip rule show
0:      from all lookup local 
32765:  from all fwmark        4 lookup 4 
32766:  from all lookup main 
32767:  from all lookup 253
[root@masq-gw]# ip route flush cache
        

The last piece is in place. Now, users in the 192.168.99.0/24 subnet who are browsing the Internet should be using the DSL line instead of the T1 line for connectivity.

In order to verify that traffic is indeed getting marked and routed appropriately, you should use tcpdump to profile the outbound traffic on each link at the same time as you generate outbound traffic on both links.

The above is a cookbook example of categorizing traffic, and sending the traffic out across different providers. To my knowledge, the commonest reason to use this sort of solution is to separate traffic by importance and use a reliable (and perhaps more costly) link for the more important traffic while reserving the less costly Internet connection for other connections. In the above illustrative case, we have simply selected the web traffic for the less reliable (DSL) provider.

Once again, if you would like to split load over multiple links regardless of classification of traffic, then you really want a multipath default route, which is described and documented very well in the LARTC HOWTO.

There are many different ways to handle hosting servers to multiple ISPs, and most of them are out of the scope of this document. If you are in need of this sort of advanced networking, you probably already know where to research. If not, I'd suggest starting your research in load balancing, global load balancing, failover, and layer 4-7 switching. These are networking tools which can facilitate the management of a highly available service.

Publishing the same service on two different ISPs is can be formidable challenge. While this is possible using some of the advanced networking features under linux, one should understand the greater issues involved with publishing a service on two public IPs, especially if the idea is to provide service to the general Internet even if one of the ISPs go down. For a thorough examination of the topics involved with load balancing of all kinds, see Chandra Kopparapu's book Load Balancing Servers, Firewalls and Caches.

If you are aware of the many difficult issues involved in handling inbound connections to a network, and still want to publish a service on two different ISPs (perhaps before you have a more robust load balancing/upper layer switching technology in place), you'll find the recipe below.

Before we examine the recipe, let's look at a complex scenario to see what the crucial points are. If you do not have the kernel packet traveling diagram memorized, you may wish to refer to it in the following discussion. One other item to remember is that routing decisions are stateless ^[37].

We'll assume that the client IP is a fixed IP (64.70.12.210) and we'll discuss how this client IP would reach each of the services published on masq-gw's two public networks. The IPs used for the services will be 67.17.28.10 and 205.254.211.17. Now, whether you are using NAT with iproute2 or with iptables, you'll run across the problem here outlined. Here is the flow of the packet through masq-gw to the server and back to the client.

This is the problem! The packet may have the correct source address, but it is leaving via the wrong interface. Many ISPs filter traffic entering their network and will block traffic from your network with source IPs outside your allocated range. To an ISP this looks like spoofed traffic.

The solution is marvelously elegant and simple. Select one IP on the internal server which will be reachable via one provider and one IP which will be reachable via the other provider. By using two IP addresses on the internal machine, we can use ip rule on masq-gw to select a routing table with a different default route based upon the source IP of the response packets to clients. Below, we'll assume the same routing tables as in the previous section (cf. Section 10.4.1, “Outbound traffic Using Multiple Connections to the Internet”).

Here we have a server isolde which needs to be accessible via two different public IP addresses. We'll add an IP address to isolde so that it is reachable on 192.168.100.10 as well as 192.168.100.17. Then, the following rules on masq-gw will ensure that packets are rewritten and routed in order to avoid the problem pointed out above.

[root@masq-gw]# ip route add nat 67.17.28.10 via 192.168.100.10
[root@masq-gw]# ip rule add nat 67.17.28.10 from 192.168.100.10 table 4
[root@masq-gw]# ip route add nat 205.254.211.17 via 192.168.100.17
[root@masq-gw]# ip rule add nat 205.254.211.17 from 192.168.100.17
[root@masq-gw]# ip rule show
0:      from all lookup local 
32765:  from 192.168.100.17 lookup main map-to 205.254.211.17
32765:  from 192.168.100.10 lookup 4 map-to 67.17.28.10
32766:  from all lookup main 
32767:  from all lookup 253
[root@masq-gw]# ip route show table local | grep ^nat
nat 205.254.211.17 via 192.168.100.17  scope host 
nat 67.17.28.10 via 192.168.100.10  scope host