Some consumer ISPs actually allow you to be assigned more than one public IP address when their DOCSIS modem is in what they call "Modem Mode".

Of course these addresses are not in any sane organization like a subnet or something, noooooo, you get a random assortment from their pool of public addresses.

This makes routing these in addition to your normal, private IPs, all via your normal gateway a bit of a challange.

While this certainly could be done it seems less than ideal to me. Sure you could set up a script that dynamically manages the address pool your DHCP server is handing out to clients, but since I just want to hand these out to a select few server machines anyway it's just not necessary and the router would be another single point of failure.

Instead my solution is to connect the machines which should get one of these public address "directly" to the modem via a VLAN on my switch, though an additional cable would work just as well ;)

Whether or not we're routing on the gateway we encounter the complication that bog standard IP routing simply isn't setup for having more than one IP addresses where each MUST be used as the source address only on the correct interface or else the paket will be dropped.

Normally when one machine has multiple IPs in a network they are either in the same Ethernet segment or at least handled by the same gateway. This is not so in our case though.

Fortunately the Linux kernel has facilities for crazy use-cases such as this. The general concept we need to apply is called "policy routing" or more specifically "source-specific routing" where the source address of a packet additionally determines to which router and out which interface it should be sent.

The idea is simple, we tell the kernel:

Every outgoing packet with a source address matching the public IP we got must be sent out via the interface having this address.

We do this by setting up one routing table for each interface which only contains a default route pointing at the gateway. We then use ip-rule policy routing to direct packets to the right table when we want to force them to egress via a particular interface.

For details on Linux Policy Routing see http://linux-ip.net/html/routing-rpdb.html, the rest of that site is also very helpful for IP routing on Linux in general.

Copy the accompanying script multihomed to /etc/dhcp/ and symlink it to /etc/dhcp/dhclient-exit-hooks.d.

The script will then add policy rules for all dhcp interface undiscriminately. Report an issue if this doesn't work for you.

Optionally it also supports setting up FWMARK based rules for WireGuard which doesn't (yet) allow binding it's UDP socket to an interface, but it does support fwmarks.

To enable this you can use an multihomed-fwmark <some-0xf00-hex-id> stanza in /etc/network/interfaces. We use ifquery to get an interface's options.

The script will also run hooks in /etc/dhcp/multihomed.d using run-parts(8) whenever the interface address might have changed. I use this to implement a workaround for WireGuard via a shitty 5G modem's NAT implementation.

Here's an example configuration and the rules we will install. In /etc/network/interfaces we have:

auto enp1s0
iface enp1s0 inet dhcp
	metric 1024

auto  enp2s0.5
iface enp2s0.5 inet dhcp
	metric 2048
	multihomed-fwmark 0x2005

Here we use the metric setting to determine which interface is used by default (greater number is less preferred).

And the resulting rules and per-interface route tables are as follows:

$ ip rule
32752:	from all fwmark 0x2005 lookup main suppress_prefixlength 0
32753:	from all fwmark 0x2005 lookup 16909060
32754:	from all fwmark 0x2005 unreachable
32755:	from 1.2.3.4 lookup main suppress_prefixlength 0
32756:	from 1.2.3.4 lookup 16909060
32757:	from 192.168.42.222 lookup main suppress_prefixlength 0
32758:	from 192.168.42.222 lookup 3232246494
32766:	from all lookup main
32767:	from all lookup default

$ ip route show table 3232246494
default via 192.168.42.1 dev enp1s0

$ ip route show table 16909060
default via 1.2.3.1 dev enp2s0.5

Notice that the lookup aka. table IDs are just the 32-bit integer representation of the associated IP address.

The lookup main suppress_prefixlength 0 rules are interesting too. See if we only insert the per-interface table lookup rules LAN destinations will become unreachable which is not quite what you'd want usually.

To fix this we use the suppress_prefixlength 0 match. When the default route matches in the lookup table it is ignored and policy rule evaluation continues. In effect you get uniform access to local LAN subnets while still routing via the right interface for global destinations.