# This test verifies that we can ping an IPv4-only server from an IPv6-only # client via a NAT64 router. The hosts and networks are configured as follows: # # +------ # Client | eth1 Address: 2001:db8::2/64 # | | Route: 64:ff9b::/96 via 2001:db8::1 # +--|--- # | VLAN 3 # +--|--- # | eth2 Address: 2001:db8::1/64 # Router | # | nat64 Address: 64:ff9b::1/128 # | Route: 64:ff9b::/96 # | Address: 192.0.2.0/32 # | Route: 192.0.2.0/24 # | # | eth1 Address: 100.64.0.1/24 # +--|--- # | VLAN 2 # +--|--- # Server | eth1 Address: 100.64.0.2/24 # | Route: 192.0.2.0/24 via 100.64.0.1 # +------ import ./make-test-python.nix ({ pkgs, lib, ... }: { name = "tayga"; meta = with pkgs.lib.maintainers; { maintainers = [ hax404 ]; }; nodes = { # The server is configured with static IPv4 addresses. RFC 6052 Section 3.1 # disallows the mapping of non-global IPv4 addresses like RFC 1918 into the # Well-Known Prefix 64:ff9b::/96. TAYGA also does not allow the mapping of # documentation space (RFC 5737). To circumvent this, 100.64.0.2/24 from # RFC 6589 (Carrier Grade NAT) is used here. # To reach the IPv4 address pool of the NAT64 gateway, there is a static # route configured. In normal cases, where the router would also source NAT # the pool addresses to one IPv4 addresses, this would not be needed. server = { virtualisation.vlans = [ 2 # towards router ]; networking = { useDHCP = false; interfaces.eth1 = lib.mkForce {}; }; systemd.network = { enable = true; networks."vlan1" = { matchConfig.Name = "eth1"; address = [ "100.64.0.2/24" ]; routes = [ { routeConfig = { Destination = "192.0.2.0/24"; Gateway = "100.64.0.1"; }; } ]; }; }; }; # The router is configured with static IPv4 addresses towards the server # and IPv6 addresses towards the client. For NAT64, the Well-Known prefix # 64:ff9b::/96 is used. NAT64 is done with TAYGA which provides the # tun-interface nat64 and does the translation over it. The IPv6 packets # are sent to this interfaces and received as IPv4 packets and vice versa. # As TAYGA only translates IPv6 addresses to dedicated IPv4 addresses, it # needs a pool of IPv4 addresses which must be at least as big as the # expected amount of clients. In this test, the packets from the pool are # directly routed towards the client. In normal cases, there would be a # second source NAT44 to map all clients behind one IPv4 address. router_systemd = { boot.kernel.sysctl = { "net.ipv4.ip_forward" = 1; "net.ipv6.conf.all.forwarding" = 1; }; virtualisation.vlans = [ 2 # towards server 3 # towards client ]; networking = { useDHCP = false; useNetworkd = true; firewall.enable = false; interfaces.eth1 = lib.mkForce { ipv4 = { addresses = [ { address = "100.64.0.1"; prefixLength = 24; } ]; }; }; interfaces.eth2 = lib.mkForce { ipv6 = { addresses = [ { address = "2001:db8::1"; prefixLength = 64; } ]; }; }; }; services.tayga = { enable = true; ipv4 = { address = "192.0.2.0"; router = { address = "192.0.2.1"; }; pool = { address = "192.0.2.0"; prefixLength = 24; }; }; ipv6 = { address = "2001:db8::1"; router = { address = "64:ff9b::1"; }; pool = { address = "64:ff9b::"; prefixLength = 96; }; }; }; }; router_nixos = { boot.kernel.sysctl = { "net.ipv4.ip_forward" = 1; "net.ipv6.conf.all.forwarding" = 1; }; virtualisation.vlans = [ 2 # towards server 3 # towards client ]; networking = { useDHCP = false; firewall.enable = false; interfaces.eth1 = lib.mkForce { ipv4 = { addresses = [ { address = "100.64.0.1"; prefixLength = 24; } ]; }; }; interfaces.eth2 = lib.mkForce { ipv6 = { addresses = [ { address = "2001:db8::1"; prefixLength = 64; } ]; }; }; }; services.tayga = { enable = true; ipv4 = { address = "192.0.2.0"; router = { address = "192.0.2.1"; }; pool = { address = "192.0.2.0"; prefixLength = 24; }; }; ipv6 = { address = "2001:db8::1"; router = { address = "64:ff9b::1"; }; pool = { address = "64:ff9b::"; prefixLength = 96; }; }; }; }; # The client is configured with static IPv6 addresses. It has also a static # route for the NAT64 IP space where the IPv4 addresses are mapped in. In # normal cases, there would be only a default route. client = { virtualisation.vlans = [ 3 # towards router ]; networking = { useDHCP = false; interfaces.eth1 = lib.mkForce {}; }; systemd.network = { enable = true; networks."vlan1" = { matchConfig.Name = "eth1"; address = [ "2001:db8::2/64" ]; routes = [ { routeConfig = { Destination = "64:ff9b::/96"; Gateway = "2001:db8::1"; }; } ]; }; }; environment.systemPackages = [ pkgs.mtr ]; }; }; testScript = '' # start client and server for machine in client, server: machine.wait_for_unit("network-online.target") machine.log(machine.execute("ip addr")[1]) machine.log(machine.execute("ip route")[1]) machine.log(machine.execute("ip -6 route")[1]) # test systemd-networkd and nixos-scripts based router for router in router_systemd, router_nixos: router.start() router.wait_for_unit("network-online.target") router.wait_for_unit("tayga.service") router.log(machine.execute("ip addr")[1]) router.log(machine.execute("ip route")[1]) router.log(machine.execute("ip -6 route")[1]) with subtest("Wait for tayga"): router.wait_for_unit("tayga.service") with subtest("Test ICMP"): client.wait_until_succeeds("ping -c 3 64:ff9b::100.64.0.2 >&2") with subtest("Test ICMP and show a traceroute"): client.wait_until_succeeds("mtr --show-ips --report-wide 64:ff9b::100.64.0.2 >&2") router.log(router.execute("systemd-analyze security tayga.service")[1]) router.shutdown() ''; })