File:  [NetBSD Developer Wiki] / wikisrc / guide / net-intro.mdwn
Revision 1.1: download - view: text, annotated - select for diffs
Mon Mar 11 20:54:42 2013 UTC (13 months, 1 week ago) by jdf
Branches: MAIN
CVS tags: HEAD
net-intro Guide -> wiki
On the way, remove bibliography from the index in favour of simple hyperlinks.

## Introduction to TCP/IP Networking

## Audience

This section explains various aspects of networking. It is intended to help
people with little knowledge about networks to get started. It is divided into
three big parts. We start by giving a general overview of how networking works
and introduce the basic concepts. Then we go into details for setting up various
types of networking in the second parts, and the third part of the networking
section covers a number of *advanced* topics that go beyond the basic operation
as introduced in the first two sections.

The reader is assumed to know about basic system administration tasks: how to
become root, edit files, change permissions, stop processes, etc. See the other
chapters of this NetBSD guide and
e.g.\[[[AeleenFrisch|guide/index#bibliography]]\]
for further information on this topic. Besides that, you should know how to
handle the utilities we're going to set up here, i.e. you should know how to
use telnet, FTP, ... I will not explain the basic features of those
utilities, please refer to the appropriate man-pages, the references listed
or of course the other parts of this document instead.

This introduction to TCP/IP Networking was written with the intention in mind to
give starters a basic knowledge. If you really want to know what it's all about,
read \[[[CraigHunt|guide/index#bibliography]]\]. This book does not only cover
the basics, but goes on and explains all the concepts, services and how to set
them up in detail. It's great, I love it! :-)

## Supported Networking Protocols

There are several protocol suites supported by NetBSD, most of which were
inherited from NetBSD's predecessor, 4.4BSD, and subsequently enhanced and
improved. The first and most important one today is DARPA's Transmission Control
Protocol/Internet Protocol (TCP/IP). Other protocol suites available in NetBSD
include the Xerox Network System (XNS) which was only implemented at UCB to
connect isolated machines to the net, Apple's AppleTalk protocol suite and the
ISO protocol suite, CCITT X.25 and ARGO TP. They are only used in some special
applications these days.

Today, TCP/IP is the most widespread protocol of the ones mentioned above. It is
implemented on almost every hardware and operating system, and it is also the
most-used protocol in heterogenous environments. So, if you just want to connect
your computer running NetBSD to some other machine at home or you want to
integrate it into your company's or university's network, TCP/IP is the right
choice. Besides the "old" IP version 4, NetBSD also supports the "new" IP
version 6 (IPv6) since NetBSD 1.5, thanks to code contributed by the KAME
project.

There are other protocol suites such as DECNET, Novell's IPX/SPX or Microsoft's
NetBIOS, but these are not currently supported by NetBSD. These protocols differ
from TCP/IP in that they are proprietary, in contrast to the others, which are
well-defined in several RFCs and other open standards.

## Supported Media

The TCP/IP protocol stack behaves the same regardless of the underlying media
used, and NetBSD supports a wide range of these, among them are Ethernet
(10/100/1000MBd), Arcnet, serial line, ATM, FDDI, Fiber Channel, USB, HIPPI,
FireWire (IEEE 1394), Token Ring, and serial lines.

### Serial Line

There are a couple of reasons for using TCP/IP over a serial line.

 * If your remote host is only reachable via telephone, you can use a modem to access it.
 * Many computers have a serial port, and the cable needed is rather cheap.

The disadvantage of a serial connection is that it's slower than other methods.
NetBSD can use at most 115200 bit/s, making it a lot slower than e.g. Ethernet's
minimum 10 Mbit/s and Arcnet's 4 Mbit/s.

There are two possible protocols to connect a host running NetBSD to another
host using a serial line (possibly over a phone-line):

 * Serial Line IP (SLIP)
 * Point to Point Protocol (PPP)

The choice here depends on whether you use a dial-up connection through a modem
or if you use a static connection (null-modem or leased line). If you dial up
for your IP connection, it's wise to use PPP as it offers some possibilities to
auto-negotiate IP-addresses and routes, which can be quite painful to do by
hand. If you want to connect to another machine which is directly connected,
use SLIP, as this is supported by about every operating system and more easy
to set up with fixed addresses and routes.

PPP on a direct connection is a bit difficult to setup, as it's easy to timeout
the initial handshake; with SLIP, there's no such initial handshake, i.e. you
start up one side, and when the other site has its first packet, it will send it
over the line.

[RFC1331](http://tools.ietf.org/html/rfc1331) ("The Point-to-Point Protocol
(PPP) for the Transmission of Multi-protocol Datagrams over Point-to-Point
Links") and [RFC1332](http://tools.ietf.org/html/rfc1332) ("The PPP Internet
Protocol Control Protocol (IPCP)") describe PPP and TCP/IP over PPP. SLIP is
defined in [RFC1055](http://tools.ietf.org/html/rfc1055) ("Nonstandard for
transmission of IP datagrams over serial lines: SLIP").

### Ethernet

Ethernet is the medium commonly used to build local area networks (LANs) of
interconnected machines within a limited area such as an office, company or
university campus. Ethernet is based on a bus structure to which many machines
can connect to, and communication always happens between two nodes at a time.
When two or more nodes want to talk at the same time, both will restart
communication after some timeout. The technical term for this is CSMA/CD
(Carrier Sense w/ Multiple Access and Collision Detection).

Initially, Ethernet hardware consisted of a thick (yellow) cable that machines
tapped into using special connectors that poked through the cable's outer
shielding. The successor of this was called 10base5, which used BNC-type
connectors for tapping in special T-connectors and terminators on both ends of
the bus. Today, ethernet is mostly used with twisted pair lines which are used
in a collapsed bus system that are contained in switches or hubs. The twisted
pair lines give this type of media its name - 10baseT for 10 Mbit/s networks,
and 100baseT for 100 MBit/s ones. In switched environments there's also the
distinction if communication between the node and the switch can happen in half-
or in full duplex mode.

## TCP/IP Address Format

TCP/IP uses 4-byte (32-bit) addresses in the current implementations (IPv4),
also called IP-numbers (Internet-Protocol numbers), to address hosts.

TCP/IP allows any two machines to communicate directly. To permit this all hosts
on a given network must have a unique IP address. To assure this, IP addresses
are administrated by one central organisation, the InterNIC. They give certain
ranges of addresses (network-addresses) directly to sites which want to
participate in the internet or to internet-providers, which give the addresses
to their customers.

If your university or company is connected to the Internet, it has (at least)
one such network-address for its own use, usually not assigned by the InterNIC
directly, but rather through an Internet Service Provider (ISP).

If you just want to run your private network at home, see below on how to
"build" your own IP addresses. However, if you want to connect your machine to
the (real :-) Internet, you should get an IP addresses from your local
network-administrator or -provider.

IP addresses are usually written in *dotted quad* notation - the four bytes are
written down in decimal (most significant byte first), separated by dots. For
example, 132.199.15.99 would be a valid address. Another way to write down
IP-addresses would be as one 32-bit hex-word, e.g. 0x84c70f63. This is not as
convenient as the dotted-quad, but quite useful at times, too. (See below!)

Being assigned a network means nothing else but setting some of the
above-mentioned 32 address-bits to certain values. These bits that are used for
identifying the network are called network-bits. The remaining bits can be used
to address hosts on that network, therefore they are called host-bits. The
following figure illustrates the separation.

![IPv4-addresses are divided into more significant network- and less significant hostbits](/guide/images/ipv4-en-0bits.png)  
**IPv4-addresses are divided into more significant network- and less significant hostbits**

In the above example, the network-address is 132.199.0.0 (host-bits are set to 0
in network-addresses) and the host's address is 15.99 on that network.

How do you know that the host's address is 16 bit wide? Well, this is assigned
by the provider from which you get your network-addresses. In the classless
inter-domain routing (CIDR) used today, host fields are usually between as
little as 2 to 16 bits wide, and the number of network-bits is written after the
network address, separated by a `/`, e.g. 132.199.0.0/16 tells that the network
in question has 16 network-bits. When talking about the *size* of a network,
it's usual to only talk about it as `/16`, `/24`, etc.

Before CIDR was used, there used to be four classes of networks. Each one starts
with a certain bit-pattern identifying it. Here are the four classes:

 * Class A starts with `0` as most significant bit. The next seven bits of a
   class A address identify the network, the remaining 24 bit can be used to
   address hosts. So, within one class A network there can be 2^24 hosts. It's
   not very likely that you (or your university, or company, or whatever) will
   get a whole class A address.

   The CIDR notation for a class A network with its eight network bits is an `/8`.

 * Class B starts with `10` as most significant bits. The next 14 bits are used
   for the networks address and the remaining 16 bits can be used to address
   more than 65000 hosts. Class B addresses are very rarely given out today,
   they used to be common for companies and universities before IPv4 address
   space went scarce.

   The CIDR notation for a class B network with its 16 network bits is an `/16`.

   Returning to our above example, you can see that 132.199.15.99 (or
   0x84c70f63, which is more appropriate here!) is on a class B network, as
   0x84... = **10**00... (base 2).

   Therefore, the address 132.199.15.99 can be split into an network-address of
   132.199.0.0 and a host-address of 15.99.

 * Class C is identified by the MSBs being `110`, allowing only 256 (actually:
   only 254, see below) hosts on each of the 2^21 possible class C networks.
   Class C addresses are usually found at (small) companies.

   The CIDR notation for a class C network with its 24 network bits is an `/24`.

 * There are also other addresses, starting with `111`. Those are used for
   special purposes (e. g. multicast-addresses) and are not of interest here.

Please note that the bits which are used for identifying the network-class are
part of the network-address.

When separating host-addresses from network-addresses, the *netmask* comes in
handy. In this mask, all the network-bits are set to `1`, the host-bits are `0`.
Thus, putting together IP-address and netmask with a logical AND-function, the
network-address remains.

To continue our example, 255.255.0.0 is a possible netmask for 132.199.15.99.
When applying this mask, the network-address 132.199.0.0 remains.

For addresses in CIDR notation, the number of network-bits given also says how
many of the most significant bits of the address must be set to `1` to get the
netmask for the corresponding network. For classful addressing, every
network-class has a fixed default netmask assigned:

 * Class A (/8): default-netmask: 255.0.0.0, first byte of address: 1-127
 * Class B (/16): default-netmask: 255.255.0.0, first byte of address: 128-191
 * Class C (/24): default-netmask: 255.255.255.0, first byte of address: 192-223

Another thing to mention here is the *broadcast-address*. When sending to this
address, *all* hosts on the corresponding network will receive the message sent.
The broadcast address is characterized by having all host-bits set to `1`.

Taking 132.199.15.99 with its netmask 255.255.0.0 again, the broadcast-address
would result in 132.199.255.255.

You'll ask now: But what if I want a host's address to be all bits `0` or `1`?
Well, this doesn't work, as network- and broadcast-address must be present!
Because of this, a class B (/16) network can contain at most 2^16-2 hosts, a
class C (/24) network can hold no more than 2^8-2 = 254 hosts.

Besides all those categories of addresses, there's the special IP-address
127.0.0.1 which always refers to the *local* host, i.e. if you talk to 127.0.0.1
you'll talk to yourself without starting any network-activity. This is sometimes
useful to use services installed on your own machine or to play around if you
don't have other hosts to put on your network.

Let's put together the things we've introduced in this section:

 * *IP-address* -- 32 bit-address, with network- and host-bits.
 * *Network-address* -- IP-address with all host bits set to `0`.
 * *Netmask* -- 32-bit mask with `1` for network- and `0` for host-bits.
 * *Broadcast* -- IP-address with all host bits set `1`.
 * *localhost's address* -- The local host's IP address is always 127.0.0.1.

## Subnetting and Routing

After talking so much about netmasks, network-, host- and other addresses, I
have to admit that this is not the whole truth.

Imagine the situation at your university, which usually has a class B (/16)
address, allowing it to have up to 2^16 \~= 65534 hosts on that net. Maybe it
would be a nice thing to have all those hosts on one single network, but it's
simply not possible due to limitations in the transport media commonly used
today.

For example, when using thinwire ethernet, the maximum length of the cable is
185 meters. Even with repeaters in between, which refresh the signals, this is
not enough to cover all the locations where machines are located. Besides that,
there is a maximum number of 1024 hosts on one ethernet wire, and you'll lose
quite a bit of performance if you go to this limit.

So, are you hosed now? Having an address which allows more than 60000 hosts, but
being bound to media which allows far less than that limit?

Well, of course not! :-)

The idea is to divide the *big* class B net into several smaller networks,
commonly called sub-networks or simply subnets. Those subnets are only allowed
to have, say, 254 hosts on them (i.e. you divide one big class B network into
several class C networks!).

To do this, you adjust your netmask to have more network- and less host-bits on
it. This is usually done on a byte-boundary, but you're not forced to do it
there. So, commonly your netmask will not be 255.255.0.0 as supposed by a class
B network, but it will be set to 255.255.255.0.

In CIDR notation, you now write a `/24` instead of the `/16` to show that 24
bits of the address are used for identifying the network and subnet, instead of
the 16 that were used before.

This gives you one additional network-byte to assign to each (physical!)
network. All the 254 hosts on that subnet can now talk directly to each other,
and you can build 256 such class C nets. This should fit your needs.

To explain this better, let's continue our above example. Say our host
132.199.15.99 (I'll call him `dusk` from now; we'll talk about assigning
hostnames later) has a netmask of 255.255.255.0 and thus is on the subnet
132.199.15.0/24. Let's furthermore introduce some more hosts so we have
something to play around with, see the next figure.

![Our demo-network](/guide/images/net-pic1.gif)  
**Our demo-network**

In the above network, `dusk` can talk directly to `dawn`, as they are both on
the same subnet. (There are other hosts attached to the 132.199.15.0/24-subnet
but they are not of importance for us now)

But what if `dusk` wants to talk to a host on another subnet?

Well, the traffic will then go through one or more gateways (routers), which are
attached to two subnets. Because of this, a router always has two different
addresses, one for each of the subnets it is on. The router is functionally
transparent, i.e. you don't have to address it to reach hosts on the *other*
side. Instead, you address that host directly and the packets will be routed to
it correctly.

Example. Let's say `dusk` wants to get some files from the local ftp-server. As
`dusk` can't reach `ftp` directly (because it's on a different subnet), all its
packets will be forwarded to its "defaultrouter" `rzi` (132.199.15.1), which
knows where to forward the packets.

`Dusk` knows the address of its defaultrouter in its network (`rzi`,
132.199.15.1), and it will forward any packets to it which are not on the same
subnet, i.e. it will forward all IP-packets in which the third address-byte
isn't 15.

The (default)router then gives the packets to the appropriate host, as it's also
on the FTP-server's network.

In this example, *all* packets are forwarded to the 132.199.1.0/24-network,
simply because it's the network's backbone, the most important part of the
network, which carries all the traffic that passes between several subnets.
Almost all other networks besides 132.199.15.0/24 are attached to the backbone
in a similar manner.

But what if we had hooked up another subnet to 132.199.15.0/24 instead of
132.199.1.0/24? Maybe something the situation displayed in the next figure.

![Attaching one subnet to another one](/guide/images/net-pic2.gif)  
**Attaching one subnet to another one**

When we now want to reach a host which is located in the 132.199.16.0/24-subnet
from `dusk`, it won't work routing it to `rzi`, but you'll have to send it
directly to `route2` (132.199.15.2). `Dusk` will have to know to forward those
packets to `route2` and send all the others to `rzi`.

When configuring `dusk`, you tell it to forward all packets for the
132.199.16.0/24-subnet to `route2`, and all others to `rzi`. Instead of
specifying this default as 132.199.1.0/24, 132.199.2.0/24, etc., 0.0.0.0 can be
used to set the default-route.

Returning to our demo network, there's a similar problem when `dawn` wants to
send to `noon`, which is connected to `dusk` via a serial line running. When
looking at the IP-addresses, `noon` seems to be attached to the
132.199.15.0-network, but it isn't really. Instead, `dusk` is used as gateway,
and `dawn` will have to send its packets to `dusk`, which will forward them to
`noon` then. The way `dusk` is forced into accepting packets that aren't
destined at it but for a different host (`noon`) instead is called *proxy arp*.

The same goes when hosts from other subnets want to send to `noon`. They have to
send their packets to `dusk` (possibly routed via `rzi`),

## Name Service Concepts

In the previous sections, when we talked about hosts, we referred to them by
their IP-addresses. This was necessary to introduce the different kinds of
addresses. When talking about hosts in general, it's more convenient to give
them *names*, as we did when talking about routing.

Most applications don't care whether you give them an IP address or a hostname.
However, they'll use IP addresses internally, and there are several methods for
them to map hostnames to IP addresses, each one with its own way of
configuration. In this section we'll introduce the idea behind each method, in
the next chapter, we'll talk about the configuration-part.

The mapping from hostnames (and domainnames) to IP-addresses is done by a piece
of software called the *resolver*. This is not an extra service, but some
library routines which are linked to every application using networking-calls.
The resolver will then try to resolve (hence the name ;-) the hostnames you give
into IP addresses. See [RFC1034](http://tools.ietf.org/html/rfc1034) ("Domain
names - concepts and facilities") and
[RFC1035](http://tools.ietf.org/html/rfc1035) ("Domain names - implementation
and specification") for details on the resolver.

Hostnames are usually up to 256 characters long, and contain letters, numbers
and dashes (`-`); case is ignored.

Just as with networks and subnets, it's possible (and desirable) to group hosts
into domains and subdomains. When getting your network-address, you usually also
obtain a domainname by your provider. As with subnets, it's up to you to
introduce subdomains. Other as with IP-addresses, (sub)domains are not directly
related to (sub)nets; for example, one domain can contain hosts from several
subnets.

Our demo-network shows this: Both subnets 132.199.1.0/24 and 132.199.15.0/24
(and others) are part of the subdomain `rz.uni-regensburg.de`. The domain the
University of Regensburg got from its IP-provider is `uni-regensburg.de` (`.de`
is for Deutschland, Germany), the subdomain `rz` is for Rechenzentrum, computing
center.

Hostnames, subdomain- and domainnames are separated by dots (`.`). It's also
possible to use more than one stage of subdomains, although this is not very
common. An example would be `fox_in.socs.uts.edu.au`.

A hostname which includes the (sub)domain is also called a fully qualified
domain name (FQDN). For example, the IP-address 132.199.15.99 belongs to the
host with the FQDN `dusk.rz.uni-regensburg.de`.

Further above I told you that the IP-address 127.0.0.1 always belongs to the
local host, regardless what's the `real` IP-address of the host. Therefore,
127.0.0.1 is always mapped to the name `localhost`.

The three different ways to translate hostnames into IP addresses are:
`/etc/hosts`, the Domain Name Service (DNS) and the Network Information Service
(NIS).

### `/etc/hosts`

The first and simplest way to translate hostnames into IP-addresses is by using
a table telling which IP address belongs to which hostname(s). This table is
stored in the file `/etc/hosts` and has the following format:

    IP-address        hostname [nickname [...]]

Lines starting with a hash mark (`#`) are treated as comments. The other lines
contain one IP-address and the corresponding hostname(s).

It's not possible for a hostname to belong to several IP addresses, even if I
made you think so when talking about routing. `rzi` for example has really two
distinct names for each of its two addresses: `rzi` and `rzia` (but please don't
ask me which name belongs to which address!).

Giving a host several nicknames can be convenient if you want to specify your
favorite host providing a special service with that name, as is commonly done
with FTP-servers. The first (leftmost) name is usually the real (canonical) name
of the host.

Besides giving nicknames, it's also convenient to give a host's full name
(including domain) as its canonical name, and using only its hostname (without
domain) as a nickname.

*Important:* There *must* be an entry mapping localhost to 127.0.0.1 in
`/etc/hosts`!

### Domain Name Service (DNS)

`/etc/hosts` bears an inherent problem, especially in big networks: when one
host is added or one host's address changes, all the `/etc/hosts` files on all
machines have to be changed! This is not only time-consuming, it's also very
likely that there will be some errors and inconsistencies, leading to problems.

Another approach is to hold only one hostnames-table (-database) for a network,
and make all the clients query that *nameserver*. Updates will be made only on
the nameserver.

This is the basic idea behind the Domain Name Service (DNS).

Usually, there's one nameserver for each domain (hence DNS), and every host
(client) in that domain knows which domain it is in and which nameserver to
query for its domain.

When the DNS gets a query about a host which is not in its domain, it will
forward the query to a DNS which is either the DNS of the domain in question or
knows which DNS to ask for the specified domain. If the DNS forwarded the query
doesn't know how to handle it, it will forward that query again to a DNS one
step higher. This is not ad infinitum, there are several *root*-servers, which
know about any domain.

See [[the separate article|guide/dns]] for details on DNS.

### Network Information Service (NIS/YP)

Yellow Pages (YP) was invented by Sun Microsystems. The name has been changed
into Network Information Service (NIS) because YP was already a trademark of the
British telecom. So, when I'm talking about NIS you'll know what I mean. ;-)

There are quite some configuration files on a Unix-system, and often it's
desired to maintain only one set of those files for a couple of hosts. Those
hosts are grouped together in a NIS-domain (which has *nothing* to do with the
domains built by using DNS!) and are usually contained in one workstation
cluster.

Examples for the config-files shared among those hosts are `/etc/passwd`,
`/etc/group` and - last but not least - `/etc/hosts`.

So, you can "abuse" NIS for getting a unique name-to-address-translation on all
hosts throughout one (NIS-)domain.

There's only one drawback, which prevents NIS from actually being used for that
translation: In contrast to the DNS, NIS provides no way to resolve hostnames
which are not in the hosts-table. There's no hosts "one level up" which the
NIS-server can query, and so the translation will fail! Suns NIS+ takes measures
against that problem, but as NIS+ is only available on Solaris-systems, this is
of little use for us now.

Don't get me wrong: NIS is a fine thing for managing e.g. user-information
(`/etc/passwd`, ...) in workstation-clusters, it's simply not too useful for
resolving hostnames.

### Other

The name resolving methods described above are what's used commonly today to
resolve hostnames into IP addresses, but they aren't the only ones. Basically,
every database mechanism would do, but none is implemented in NetBSD. Let's have
a quick look what you may encounter.

With NIS lacking hierarchy in data structures, NIS+ is intended to help out in
that field. Tables can be setup in a way so that if a query cannot be answered
by a domain's server, there can be another domain "above" that might be able to
do so. E.g. you could choose to have a domain that lists all the hosts (users,
groups, ...) that are valid in the whole company, one that defines the same
for each division, etc. NIS+ is not used a lot today, even Sun went back to
ship back NIS by default.

Last century, the X.500 standard was designed to accommodate both simple
databases like `/etc/hosts` as well as complex, hierarchical systems as can be
found e.g. in DNS today. X.500 wasn't really a success, mostly due to the fact
that it tried to do too much at the same time. A cut-down version is available
today as the Lightweight Directory Access Protocol (LDAP), which is becoming
popular in the last years to manage data like users but also hosts and others in
small to medium sized organisations.

## Next generation Internet protocol - IPv6

### The Future of the Internet

According to experts, the Internet as we know it will face a serious problem in
a few years. Due to its rapid growth and the limitations in its design, there
will be a point at which no more free addresses are available for connecting new
hosts. At that point, no more new web servers can be set up, no more users can
sign up for accounts at ISPs, no more new machines can be setup to access the
web or participate in online games - some people may call this a serious
problem.

Several approaches have been made to solve the problem. A very popular one is to
not assign a worldwide unique address to every user's machine, but rather to
assign them *private* addresses, and hide several machines behind one official,
globally unique address. This approach is called *Network Address Translation*
(NAT, also known as IP Masquerading). It has problems, as the machines hidden
behind the global address can't be addressed, and as a result of this, opening
connections to them - which is used in online gaming, peer to peer networking,
etc. - is not possible. For a more in-depth discussion of the drawbacks of NAT,
see \[[[RFC3027|guide/index#bilbiography]]\] ("Protocol Complications with the
IP Network Address Translator").

A different approach to the problem of internet addresses getting scarce is to
abandon the old Internet protocol with its limited addressing capabilities, and
use a new protocol that does not have these limitations. The protocol - or
actually, a set of protocols - used by machines connected to form today's
Internet is know as the TCP/IP (Transmission Control Protocol, Internet
Protocol) suite, and version 4 currently in use has all the problems described
above. Switching to a different protocol version that does not have these
problems of course requires for a 'better' version to be available, which
actually is. Version 6 of the Internet Protocol (IPv6) does fulfill any possible
future demands on address space, and also addresses further features such as
privacy, encryption, and better support of mobile computing.

Assuming a basic understanding of how today's IPv4 works, this text is intended
as an introduction to the IPv6 protocol. The changes in address formats and name
resolution are covered. With the background given here, the next sections will
show how to use IPv6 even if your ISP doesn't offer it by using a simple yet
efficient transition mechanism called 6to4. The goal is to get online with IPv6,
giving an example configuration for NetBSD.

### What good is IPv6?

When telling people to migrate from IPv4 to IPv6, the question you usually hear
is *why?*. There are actually a few good reasons to move to the new version:

 * Bigger address space
 * Support for mobile devices
 * Built-in security

#### Bigger Address Space

The bigger address space that IPv6 offers is the most obvious enhancement it has
over IPv4. While today's internet architecture is based on 32-bit wide
addresses, the new version has 128 bit available for addressing. Thanks to the
enlarged address space, work-arounds like NAT don't have to be used any more.
This allows full, unconstrained IP connectivity for today's IP based machines as
well as upcoming mobile devices like PDAs and cell phones will benefit from full
IP access through GPRS and UMTS.

#### Mobility

When mentioning mobile devices and IP, another important point to note is that
some special protocol is needed to support mobility, and implementing this
protocol - called *Mobile IP* - is one of the requirements for every IPv6 stack.
Thus, if you have IPv6 going, you have support for roaming between different
networks, with everyone being updated when you leave one network and enter the
other one. Support for roaming is possible with IPv4 too, but there are a number
of hoops that need to be jumped in order to get things working. With IPv6,
there's no need for this, as support for mobility was one of the design
requirements for IPv6. See [RFC3024](http://tools.ietf.org/html/rfc3024)
("Reverse Tunneling for Mobile IP") for some more information on the issues that
need to be addressed with Mobile IP on IPv4.

#### Security

Besides support for mobility, security was another requirement for the successor
to today's Internet Protocol version. As a result, IPv6 protocol stacks are
required to include IPsec. IPsec allows authentication, encryption and
compression of any IP traffic. Unlike application level protocols like SSL or
SSH, all IP traffic between two nodes can be handled, without adjusting any
applications. The benefit of this is that all applications on a machine can
benefit from encryption and authentication, and that policies can be set on a
per-host (or even per-network) base, not per application/service. An
introduction to IPsec with a roadmap to the documentation can be found in
[RFC2411](http://tools.ietf.org/html/rfc2411) ("IP Security Document Roadmap"),
the core protocol is described in [RFC2401](http://tools.ietf.org/html/rfc2401)
("Security Architecture for the Internet Protocol").  

### Changes to IPv4

After giving a brief overview of all the important features of IPv6, we'll go
into the details of the basics of IPv6 here. A brief understanding of how IPv4
works is assumed, and the changes in IPv6 will be highlighted. Starting with
IPv6 addresses and how they're split up we'll go into the various types of
addresses there are, what became of broadcasts, then after discussing the IP
layer go into changes for name resolving and what's new in DNS for IPv6.

#### Addressing

An IPv4 address is a 32 bit value, that's usually written in *dotted quad*
representation, where each *quad* represents a byte value between 0 and 255, for
example:

    127.0.0.1

This allows a theoretical number of 2^32 or \~4 billion hosts to be connected on
the internet today. Due to grouping, not all addresses are available today.

IPv6 addresses use 128 bit, which results in 2^128 theoretically addressable
hosts. This allows for a Really Big number of machines to addressed, and it sure
fits all of today's requirements plus all those nifty PDAs and cell phones with
IP phones in the near future without any sweat. When writing IPv6 addresses,
they are usually divided into groups of 16 bits written as four hex digits, and
the groups are separated by colons. An example is:

    fe80::2a0:d2ff:fea5:e9f5

This shows a special thing - a number of consecutive zeros can be abbreviated by
a single `::` once in the IPv6 address. The above address is thus equivalent to
`fe80:0:00:000:2a0:d2ff:fea5:e9f5` - leading zeros within groups can be omitted,
and only one `::` can be used in an IPv6 address.

To make addresses manageable, they are split in two parts, which are the bits
identifying the network a machine is on, and the bits that identify a machine on
a (sub)network. The bits are known as netbits and hostbits, and in both IPv4 and
IPv6, the netbits are the *left*, most significant bits of an IP address, and
the host bits are the *right*, least significant bits, as shown in the following
figure:

![IPv6-addresses are divided into more significant network- and less significant hostbits, too](/guide/images/ipv6-en-0bits.gif)  
**IPv6-addresses are divided into more significant network- and less significant hostbits, too**

In IPv4, the border is drawn with the aid of the netmask, which can be used to
mask all net/host bits. Typical examples are 255.255.0.0 that uses 16 bit for
addressing the network, and 16 bit for the machine, or 255.255.255.0 which takes
another 8 bit to allow addressing 256 subnets on e.g. a class B net.

When addressing switched from classful addressing to CIDR routing, the borders
between net and host bits stopped being on 8 bit boundaries, and as a result the
netmasks started looking ugly and not really manageable. As a replacement, the
number of network bits is used for a given address, to denote the border, e.g.

    10.0.0.0/24

is the same as a netmask of 255.255.255.0 (24 1-bits). The same scheme is used
in IPv6:

    2001:638:a01:2::/64

tells us that the address used here has the first (leftmost) 64 bits used as the
network address, and the last (rightmost) 64 bits are used to identify the
machine on the network. The network bits are commonly referred to as (network)
*prefix*, and the *prefixlen* here would be 64 bits.

Common addressing schemes found in IPv4 are the (old) class B and class C nets.
With a class C network (/24), you get 24 bits assigned by your provider, and it
leaves 8 bits to be assigned by you. If you want to add any subnetting to that,
you end up with *uneven* netmasks that are a bit nifty to deal with. Easier for
such cases are class B networks (/16), which only have 16 bits assigned by the
provider, and that allow subnetting, i.e. splitting of the rightmost bits into
two parts. One to address the on-site subnet, and one to address the hosts on
that subnet. Usually, this is done on byte (8 bit) boundaries. Using a netmask
of 255.255.255.0 (or a /24 prefix) allows flexible management even of bigger
networks here. Of course there is the upper limit of 254 machines per subnet,
and 256 subnets.

With 128 bits available for addressing in IPv6, the scheme commonly used is the
same, only the fields are wider. Providers usually assign /48 networks, which
leaves 16 bits for a subnetting and 64 hostbits.

![IPv6-addresses have a similar structure to class B addresses](/guide/images/ipv6-en-6adrformats.gif)
**IPv6-addresses have a similar structure to class B addresses**

Now while the space for network and subnets here is pretty much ok, using 64
bits for addressing hosts seems like a waste. It's unlikely that you will want
to have several billion hosts on a single subnet, so what is the idea behind
this?

The idea behind fixed width 64 bit wide host identifiers is that they aren't
assigned manually as it's usually done for IPv4 nowadays. Instead, IPv6 host
addresses are recommended (not mandatory!) to be built from so-called EUI64
addresses. EUI64 addresses are - as the name says - 64 bit wide, and derived
from MAC addresses of the underlying network interface. E.g. for ethernet, the 6
byte (48 bit) MAC address is usually filled with the hex bits `fffe` in the
middle and a bit is set to mark the address as unique (which is true for
Ethernet), e.g. the MAC address

    01:23:45:67:89:ab

results in the EUI64 address

    03:23:45:ff:fe:67:89:ab

which again gives the host bits for the IPv6 address as

    ::0323:45ff:fe67:89ab

These host bits can now be used to automatically assign IPv6 addresses to hosts,
which supports autoconfiguration of IPv6 hosts - all that's needed to get a
complete IPv6 address is the first (net/subnet) bits, and IPv6 also offers a
solution to assign them automatically.

When on a network of machines speaking IP, there's usually one router which acts
as the gateway to outside networks. In IPv6 land, this router will send *router
advertisement* information, which clients are expected to either receive during
operation or to solicit upon system startup. The router advertisement
information includes data on the router's address, and which address prefix it
routes. With this information and the host-generated EUI64 address, an IPv6-host
can calculate its IP address, and there is no need for manual address
assignment. Of course routers still need some configuration.

The router advertisement information they create are part of the Neighbor
Discovery Protocol (NDP, see [RFC2461](http://tools.ietf.org/html/rfc2461)
("Neighbor Discovery for IP Version 6 (IPv6)")), which is the successor to
IPv4's ARP protocol. In contrast to ARP, NDP does not only do lookup of IPv6
addresses for MAC addresses (the neighbor solicitation/advertisement part), but
also does a similar service for routers and the prefixes they serve, which is
used for autoconfiguration of IPv6 hosts as described in the previous paragraph.

#### Multiple Addresses

In IPv4, a host usually has one IP address per network interface or even per
machine if the IP stack supports it. Only very rare applications like web
servers result in machines having more than one IP address. In IPv6, this is
different. For each interface, there is not only a globally unique IP address,
but there are two other addresses that are of interest: The link local address,
and the site local address. The link local address has a prefix of fe80::/64,
and the host bits are built from the interface's EUI64 address. The link local
address is used for contacting hosts and routers on the same network only, the
addresses are not visible or reachable from different subnets. If wanted,
there's the choice of either using global addresses (as assigned by a provider),
or using site local addresses. Site local addresses are assigned the network
address fec0::/10, and subnets and hosts can be addressed just as for
provider-assigned networks. The only difference is, that the addresses will not
be visible to outside machines, as these are on a different network, and their
*site local* addresses are in a different physical net (if assigned at all). As
with the 10/8 network in IPv4, site local addresses can be used, but don't have
to. For IPv6 it's most common to have hosts assigned a link-local and a global
IP address. Site local addresses are rather uncommon today, and are no
substitute for globally unique addresses if global connectivity is required.

#### Multicasting

In IP land, there are three ways to talk to a host: unicast, broadcast and
multicast. The most common one is by talking to it directly, using its unicast
address. In IPv4, the unicast address is the *normal* IP address assigned to a
single host, with all address bits assigned. The broadcast address used to
address all hosts in the same IP subnet has the network bits set to the network
address, and all host bits set to `1` (which can be easily done using the
netmask and some bit operations). Multicast addresses are used to reach a number
of hosts in the same multicast group, which can be machines spread over the
whole internet. Machines must join multicast groups explicitly to participate,
and there are special IPv4 addresses used for multicast addresses, allocated
from the 224/8 subnet. Multicast isn't used very much in IPv4, and only few
applications like the MBone audio and video broadcast utilities use it.

In IPv6, unicast addresses are used the same as in IPv4, no surprise there - all
the network and host bits are assigned to identify the target network and
machine. Broadcasts are no longer available in IPv6 in the way they were in
IPv4, this is where multicasting comes into play. Addresses in the ff::/8
network are reserved for multicast applications, and there are two special
multicast addresses that supersede the broadcast addresses from IPv4. One is the
*all routers* multicast address, the others is for *all hosts*. The addresses
are specific to the subnet, i.e. a router connected to two different subnets can
address all hosts/routers on any of the subnets it's connected to. Addresses
here are:

 * ff0*`X`*::1 for all hosts and
 * ff0*`X`*::2 for all routers,

where `X` is the scope ID of the link here, identifying the network. Usually
this starts from `1` for the *node local* scope, `2` for the first link, etc.
Note that it's perfectly ok for two network interfaces to be attached to one
link, thus resulting in double bandwidth:

![Several interfaces attached to a link result in only one scope ID for the link](/guide/images/ipv6-en-4scope.gif)  
**Several interfaces attached to a link result in only one scope ID for the link**

One use of the *all hosts* multicast is in the neighbor solicitation code of
NDP, where any machine that wants to communicate with another machine sends out
a request to the *all hosts* group, and the machine in question is expected to
respond.

#### Name Resolving in IPv6

After talking a lot about addressing in IPv6, anyone still here will hope that
there's a proper way to abstract all these long & ugly IPv6 addresses with some
nice hostnames as one can do in IPv4, and of course there is.

Hostname to IP address resolving in IPv4 is usually done in one of three ways:
using a simple table in `/etc/hosts`, by using the Network Information Service
(NIS, formerly YP) or via the Domain Name System (DNS).

As of this writing, NIS/NIS+ over IPv6 is currently only available on Solaris 8,
for both database contents and transport, using a RPC extension.

Having a simple address<-\>name map like `/etc/hosts` is supported in all IPv6
stacks. With the KAME implementation used in NetBSD, `/etc/hosts` contains IPv6
addresses as well as IPv4 addresses. A simple example is the `localhost` entry
in the default NetBSD installation:

    127.0.0.1               localhost
    ::1                     localhost

For DNS, there are no fundamentally new concepts. IPv6 name resolving is done
with AAAA records that - as the name implies - point to an entity that's four
times the size of an A record. The AAAA record takes a hostname on the left
side, just as A does, and on the right side there's an IPv6 address, e.g.

    noon            IN      AAAA    3ffe:400:430:2:240:95ff:fe40:4385

For reverse resolving, IPv4 uses the in-addr.arpa zone, and below that it writes
the bytes (in decimal) in reversed order, i.e. more significant bytes are more
right. For IPv6 this is similar, only that hex digits representing 4 bits are
used instead of decimal numbers, and the resource records are also under a
different domain, ip6.int.

So to have the reverse resolving for the above host, you would put into your
`/etc/named.conf` something like:

    zone "0.3.4.0.0.0.4.0.e.f.f.3.IP6.INT" {
          type master;
          file "db.reverse";
    };

and in the zone file db.reverse you put (besides the usual records like SOA and
NS):

    5.8.3.4.0.4.e.f.f.f.5.9.0.4.2.0.2.0.0.0 IN   PTR   noon.ipv6.example.com.

The address is reversed here, and written down one hex digit after the other,
starting with the least significant (rightmost) one, separating the hex digits
with dots, as usual in zone files.

One thing to note when setting up DNS for IPv6 is to take care of the DNS
software version in use. BIND 8.x does understand AAAA records, but it does not
offer name resolving via IPv6. You need BIND 9.x for that. Beyond that, BIND 9.x
supports a number of resource records that are currently being discussed but not
officially introduced yet. The most noticeable one here is the A6 record which
allows easier provider/prefix changing.

To sum up, this section talked about the technical differences between IPv4 and
IPv6 for addressing and name resolving. Some details like IP header options, QoS
and flows were deliberately left out to not make this document more complex than
necessary.


CVSweb for NetBSD wikisrc <wikimaster@NetBSD.org> software: FreeBSD-CVSweb