MAP66 (NAT from IPv6 to IPv6, NAT66) for Linux

Sven-Ola T cke

Freifunk

06-OCT-2010

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Table of Contents

Installation
DKMS Integration
Configuration

    Brief Version
    Detailed Version

Hints (Chapter is To-do)
Motivation

These files implement a Linux netfilter target to change the IPv6 address of
packets. The address change is done checksum neutral, thus no checksum
re-calculation for the IPv6 packet is necessary. You can change the IPv6 source
address of outgoing packets as well as the IPv6 destination address of incoming
packets. This allows you to map an internal IPv6 address range to a second,
externally used IPv6 address range. IPv6 address mapping is not very similar to
IPv4 network address translation, but one can describe it as some sort of
stateless NAT. The implementation is based on the expired the IETF discussion
paper published here:

http://tools.ietf.org/html/draft-mrw-behave-nat66-02

Installation

MAP66 implements two pieces of software: a shared library that extends the
ip6tables command and a Linux kernel module. The shared library file adds the
'-j MAP66' target to the ip6tables command. To build and install, you need
ip6tables installed as well as the necessary headers. The Linux kernel module
requires the Linux source file tree and Linux configuration files to compile.
On a Debian/(EKU)buntu, the following command prepares the build environment:

sudo apt-get install build-essential linux-headers iptables-dev

Unpack the source tgz archive, change to the directory and issue "make" to
build. If this compiles without errors, install the ip6tables extension with
the following command:

sudo make install

Note

The kernel module (ip6t_MAP66.ko for Linux-2.6 or ip6t_MAP66.o for Linux-2.4)
is not automatically installed nor loaded into the kernel. You can copy the
kernel module file manually, e.g. with sudo cp ip6t_MAP66.ko /lib/modules/$
(uname -r)/.

DKMS Integration

If the next system update needs to install a new kernel version, you also need
to re-compile/re-install the MAP66 kernel module. With Debian/(EKU)buntu, this
can be automated with the Dynamic Kernel Module Support Framework (DKMS). For
this, the dkms.conf file is included with the MAP66 source file package.
Install DKMS with the following command:

sudo apt-get install dkms

Move/unpack the MAP66 source files below /usr/src and adapt the directory name
to the version mentioned in the dkms.conf file. For example, issue mkdir /usr/
src/ip6t_MAP66-0.3 and cp * /usr/src/ip6t_MAP66-0.3/. To register the MAP66
source to DKMS and compile/install, issue these commands:

sudo dkms add -m ip6t_MAP66 -v 0.3
sudo dkms build -m ip6t_MAP66 -v 0.3
sudo dkms install -m ip6t_MAP66 -v 0.3

Read DKMS details here: https://wiki.kubuntu.org/Kernel/Dev/DKMSPackaging

Configuration

Brief Version

You always need to add two ip6tables-rules to your netfilter configuration. One
rule matches the outgoing packet and changes the IPv6 source address. The
second rule matches the incoming packet and revert the address change by
altering the IPv6 destination address. To following commands correspond to the
“Address Mapping Example” given in the IETF discussion paper:

ip6tables -t mangle -I POSTROUTING -o eth0 -s FD01:0203:0405::/48 -j MAP66 --to 2001:0DB8:0001::/48
ip6tables -t mangle -I PREROUTING  -i eth0 -d 2001:0DB8:0001::/48 -j MAP66 --to FD01:0203:0405::/48

This example is also printed to the screen if you issue ip6tables -j MAP66
--help. By design, you cannot use an arbitrary prefix length. Only /112, /96 ..
/16 are supported.

For each packet, the Linux kernel module also compares the packet's source
address to all IPv6 addresses assigned to the outgoing interface. It a match is
found, the packet's source address is not mapped. The same comparison happens
on the incoming packet's destination address. The comparison require some CPU
resources, especially if the interface has a large number of assigned IPv6
addresses. If you are sure, that the mapping cannot match the IPv6 address of
the interface (e.g. the mapping rule defines a mapping prefix that cannot match
the interface address) you can switch off the comparison. Add the --nocheck
parameter to the ip6tables command for this.

Detailed Version

The following explanation details a living example from the wireless mesh
network that is mentioned under Motivation (see below). The mesh network uses a
private IP range, the ULA prefix fdca:ffee:babe::/64. All mesh nodes derive
their IPv6 interface addresses by correlating the ULA prefix with the EUI48
(“MAC address”) of the respective network adapter.

There is a Debian based virtual machine that should act as one Internet gateway
for the mesh. You can reach this virtual machine's web service via IPv4 under
http://bbb-vpn.freifunk.net. To experiment with IPv6, a SIXXS static tunnel
setup has been added and there is also an experimental 6-to-4 configuration.
The following /etc/network/interfaces file provides the configuration for IPv6:

auto sixxs
iface sixxs inet6 v4tunnel
        address  2001:4dd0:ff00:2ee::2
        netmask 64
        local 77.87.48.7
        endpoint 78.35.24.124
        ttl 64
        up ip link set mtu 1280 dev $IFACE
        up ip route add default via  2001:4dd0:ff00:2ee::1 dev $IFACE
        up ip addr add 2001:4dd0:fe77::1/48 dev $IFACE

#auto tun6to4
iface tun6to4 inet6 v4tunnel
        # ipv6calc --quiet --action conv6to4 77.87.48.7
        address 2002:4d57:3007::1
        netmask 16
        local 77.87.48.7
        endpoint any
        ttl 64
        gateway ::192.88.99.1

As you can see, the virtual machine has an IPv6 prefix of 2001:4dd0:fe77::/48
and is reachable via http://[2001:4dd0:fe77::1]/. For experimental reasons, the
6-to-4 tunnel can be activated by issuing ifup tun6to4. The netfilter setup of
this machine includes the following command sequence to ensure mapping from the
private fdca:ffee:babe::/64 prefix to the globally valid IPv6 address:

ip6tables -t mangle -F POSTROUTING
ip6tables -t mangle -F PREROUTING
ip6tables -t mangle -F FORWARD

grep -q ^ip6t_MAP66 /proc/modules && rmmod ip6t_MAP66
insmod /usr/src/map66/ip6t_MAP66.ko

ip6tables -t mangle -A POSTROUTING -o sixxs -s fdca:ffee:babe::/64 -j MAP66 --to 2001:4dd0:fe77:1::/64 --nocheck
ip6tables -t mangle -A PREROUTING  -i sixxs -d 2001:4dd0:fe77:1::/64 -j MAP66 --to fdca:ffee:babe::/64 --nocheck
ip6tables -t mangle -A POSTROUTING -o tun6to4 -s fdca:ffee:babe::/64 -j MAP66 --to 2002:4d57:3007:1::/64 --nocheck
ip6tables -t mangle -A PREROUTING  -i tun6to4 -d 2002:4d57:3007:1::/64 -j MAP66 --to fdca:ffee:babe::/64 --nocheck
ip6tables -t mangle -A FORWARD -p tcp --tcp-flags SYN,RST SYN -j TCPMSS --clamp-mss-to-pmtu

Because for both IPv6 networks the external prefix length is smaller than the
internal prefix length, we can make sure that the mapped addresses cannot match
the interfaces address. For example: 2001:4dd0:fe77:1::/64 cannot match
2001:4dd0:fe77:0::1/128. For this reason, we can use the --nocheck speedup flag
here.

You may stumble over the MSS-clamping rule. While IPv6 defines, that path MTU
detection via ICMPv6 must be supported by any host, sometimes path MTU
detection does not work. The SIXXS tunnel uses an MTU of 1280 byte. To get the
following command working, I needed to add the above MSS-clamping rule on the
gateway:

wget --prefer-family=IPv6 -O - http://6to4.nro.net/

Note

The tun6to4 tunnel interface is disabled normally, because of the implicit
2002::/16 network route configured for that interface. This network route
ensures, that traffic between one 2002::/16 to another 2002::/16 travels
directly between the IPv4 hosts. Without this network route, any IPv6 traffic
will be routed via the 6-to-4 gateways which may not work and place a higher
load on those gateways.

However, if you ping the SIXXS IP address from another host that has a 6-to-4
address, you will get the answer packet back via 6-to-4. If the above network
mapping is configured, you ping one IPv6 address and get the answer from
another IPv6 address...

Hints (Chapter is To-do)

Note for me: svn propedit svn:ignore.

Note on ubuntu. One needs to enable prefer-family in /etc/wgetrc. after this
e.g. "wget -O /dev/null http://ftp.se.debian.org" should connect to the IPv6
addrs of that server. Alternative: change presedence of ULAs in /etc/gai.conf

Idea to think about: --salt 3b5b91c5a2 XOR client addresses for some more
privacy

Motivation

My Internet access at home is realized by a wireless community mesh network not
owned by me. The mesh is operated with small embedded devices (nodes aka. WLAN
routers) that are interconnected via radio links (WLAN IBSS / AdHoc). Routing
is done with a specialized protocol such as Batman or OLSR. The routing
protocol selects the nearest out of a dozen Internet gateways and configures a
default route or an IPIP tunnel accordingly. Each Internet gateway is connected
to a different ISP and provides the service with the help of IPv4 network
address translation (NAT). Using NAT has the following effects:

  ● Address amplification - something not necessary with IPv6 any more

  ● Anonymization - nice to have as an option but not mission critical

  ● ISP independence - no reverse routing, no "buy-a-number-range"

The last point _is_ mission critical. One can use a provider independent IPv6
address range, but you need the cooperation of an ISP to use that address range
for Internet connectivity. If you e.g. move to another ISP you need that
address range to be re-routed to your new location.

ISP independence is also possible with some tunneling technique, such as VPN or
mobile IP. Tunneling can be implemented on client PCs and Internet gateways/
servers one day. But there is no need to implement the same tunneling technique
on every mesh node. Why? Because the mesh nodes can use private IP addresses
(or "ULA") to transport the tunnel data between the client PC and the gateway/
server. Each tunneling technique typically needs a single instance (the
"server") which forms a single point of failure. Rule-of-thumb1: avoid a SPOF
for the infrastructure. Rule-of-thumb2: KISS.

Using private IP addresses on the mesh nodes has a drawback: mesh node software
updates e.g. via HTTP downloads from an Internet server is not possible. This
is where I start to think: "hey, some kind of address mapping may be nice to
have". While opening Pandora's NAT66 box, I discovered that IPv6 nerds do not
like the acronym. It is always a good tactic in info wars to rename, hence the
name "MAP66".

// Sven-Ola