tiny changes

[lartc.git] / 2.4routing.sgml

<!doctype linuxdoc system>

<!-- $Id$
 -->

<article>

<!-- Title information -->

<title>Linux 2.4 Advanced Routing HOWTO
<author>Netherlabs BV (bert hubert &lt;bert.hubert@netherlabs.nl&gt;)&nl;
Gregory Maxwell &lt;greg@linuxpower.cx&gt; &nl;
Remco van Mook &lt;remco@virtu.nl&gt; &nl;
Martijn van Oosterhout &lt;kleptog@cupid.suninternet.com&gt; &nl;
Paul B Schroeder  &lt;paulsch@us.ibm.com&gt; &nl;
Jasper Spaans &lt;jasper@spaans.ds9a.nl&gt; &nl;
howto@ds9a.nl
<date>v0.3.0 $Date$
<abstract>
A very hands-on approach to iproute2, traffic shaping and a bit of netfilter
</abstract>

<!-- Table of contents -->
<toc>

<!-- Begin the document -->

<sect>Dedication
<p>
This document is dedicated to lots of people, and is my attempt to do
something back. To list but a few:
<p>
<itemize>
        <item>Rusty Russell
        <item>Alexey N. Kuznetsov
        <item>The good folks from Google
        <item>The staff of Casema Internet
</itemize>      

<sect>Introduction
<p>
Welcome, gentle reader.
<p>
This document hopes to enlighten you on how to do more with Linux 2.2/2.4
routing. Unbeknownst to most users, you already run tools which allow you to
do spectacular things. Commands like 'route' and 'ifconfig' are actually
very thin wrappers for the very powerful iproute2 infrastructure
<p>
I hope that this HOWTO will become as readable as the ones by Rusty Russell
of (amongst other things) netfilter fame.

You can always reach us by writing to the <url name="HOWTO team"
url="mailto:HOWTO@ds9a.nl">. However, please consider posting to the mailing
list (see the relevant section) if you have questions which are not directly
related to this HOWTO.

Before losing your way in this HOWTO, if all you want to do is simple
traffic shaping, skip everything and head to the 'Other possibilties'
chapter, and read about CBQ.init.


<sect1>Disclaimer &amp; License
<p>
This document is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

In short, if your STM-64 backbone breaks down and distributes pornography to
your most esteemed customers - it's never our fault. Sorry.

Copyright (c) 2001 by bert hubert, Gregory Maxwell, Martijn van
Oosterhout, Remco van Mook, Paul B. Schroeder and others.

Please freely copy and distribute (sell or give away) this document in any
format. It's requested that corrections and/or comments be fowarded to the
document maintainer. You may create a derivative work and distribute it
provided that you:

<enum>
<item>Send your derivative work (in the most suitable format such as sgml) to
the LDP (Linux Documentation Project) or the like for posting on the Internet. 
If not the LDP, then let the LDP know where it is available. 

<item>License the derivative work with this same license or use GPL. Include a
copyright notice and at least a pointer to the license used. 

<item>Give due credit to previous authors and major contributors. 

</enum>
If you're considering making a derived work other than a translation, it's
requested that you discuss your plans with the current maintainer. 

It is also requested that if you publish this HOWTO in hardcopy that you
send the authors some samples for 'review purposes' :-) 


<sect1>Prior knowledge
<p>
As the title implies, this is the 'Advanced' HOWTO. While by no means rocket
science, some prior knowledge is assumed. 

Here are some other references which might help learn you more:
<descrip>
<tag><url
url="http://netfilter.samba.org/unreliable-guides/networking-concepts-HOWTO/index.html"
name="Rusty Russell's networking-concepts-HOWTO"></tag>
Very nice introduction, explaining what a network is, and how it is
connected to other networks
<tag>Linux Networking-HOWTO (Previously the Net-3 HOWTO)</tag>
Great stuff, although very verbose. It learns you a lot of stuff that's
already configured if you are able to connect to the internet. 
Should be located in <file>/usr/doc/HOWTO/NET3-4-HOWTO.txt</file> but can be also be found 
<url url="http://www.linuxports.com/howto/networking"
name="online">
</descrip>

<sect1>What Linux can do for you
<p>
A small list of things that are possible:
<p>
<itemize>
        <item>Throttle bandwidth for certain computers
        <item>Throttle bandwidth TO certain computers
        <item>Help you to fairly share your bandwidth
        <item>Protect your network from DoS attacks
        <item>Protect the internet from your customers
        <item>Multiplex several servers as one, for load balancing or
enhanced availability
        <item>Restrict access to your computers
        <item>Limit access of your users to other hosts
        <item>Do routing based on user id (yes!), MAC address, source IP
address, port, type of service, time of day or content
</itemize>
<p>
Currently, not many people are using these advanced features. This is for
several reasons. While the provided documentation is verbose, it is not very
hands-on. Traffic control is almost undocumented.
<sect1>Housekeeping notes
<p>
There are several things which should be noted about this document. While I
wrote most of it, I really don't want it to stay that way. I am a strong
believer in Open Source, so I encourage you to send feedback, updates,
patches etcetera. Do not hesitate to inform me of typos or plain old errors.
If my English sounds somewhat wooden, please realise that I'm not a native
speaker. Feel free to send suggestions.

If you feel to you are better qualified to maintain a section, or think that
you can author and maintain new sections, you are welcome to do so. The SGML
of this HOWTO is available via CVS, I very much envision more people
working on it.

In aid of this, you will find lots of FIXME notices. Patches are always
welcome! Wherever you find a FIXME, you should know that you are treading in
unknown territory. This is not to say that there are no errors elsewhere,
but be extra careful. If you have validated something, please let us know so
we can remove the FIXME notice.

About this HOWTO, I will take some liberties along the road. For example, I
postulate a 10Mbit internet connection, while I know full well that those
are not very common.
<sect1>Access, CVS &amp; submitting updates
<p>
The canonical location for the HOWTO is <url
url="http://www.ds9a.nl/2.4Routing" name="here">.

We now have anonymous CVS access available to the world at large. This is
good in a number of ways. You can easily upgrade to newer versions of this
HOWTO and submitting patches is no work at all.

Furthermore, it allows the authors to work on the source independently,
which is good too.

<tscreen><verb>
$ export CVSROOT=:pserver:anon@outpost.ds9a.nl:/var/cvsroot
$ cvs login
CVS password: [enter 'cvs' (without 's)]
$ cvs co 2.4routing
cvs server: Updating 2.4routing
U 2.4routing/2.4routing.sgml
</verb></tscreen>

If you spot an error, or want to add something, just fix it locally, and run
cvs diff -u, and send the result off to us.

A Makefile is supplied which should help you create postscript, dvi, pdf,
html and plain text. You may need to install sgml-tools, ghostscript and
tetex to get all formats.

<sect1>Mailing list
<p>
<label id="MLIST">
The authors receive an increasing amount of mail about this HOWTO. Because
of the clear interest of the community, it has been decided to start a
mailinglist where people can talk to each other about Advanced Routing and
Traffic Control. You can subscribe to the list
<url url="http://mailman.ds9a.nl/mailman/listinfo/lartc" name="here">.
<p>
It should be pointed out that the authors are very hesitant of answering
questions not asked on the list. We would like the archive of the list to
become some kind of knowledge base. If you have a question, please search
the archive, and then post to the mailinglist.

<sect1>Layout of this document
<p>
We will be doing interesting stuff almost immediately, which also means that
there will initially be parts that are explained incompletely or are not
perfect. Please gloss over these parts and assume that all will become clear.

Routing and filtering are two distinct things. Filtering is documented very
well by Rusty's HOWTOs, available here:

<itemize>
        <item><url url="http://netfilter.samba.org/unreliable-guides/"
name="Rusty's Remarkably Unreliable Guides">
</itemize>

We will be focusing mostly on what is possible by combining netfilter and
iproute2.
<sect>Introduction to iproute2
<sect1>Why iproute2?
<p>
Most Linux distributions, and most UNIX's, currently use the 
venerable 'arp', 'ifconfig' and 'route' commands. While these tools work,
they show some unexpected behaviour under Linux 2.2 and up. For example, GRE
tunnels are an integral part of routing these days, but require completely
different tools.

With iproute2, tunnels are an integral part of the tool set.

The 2.2 and above Linux kernels include a completely redesigned network
subsystem. This new networking code brings Linux performance and a feature
set with little competition in the general OS arena. In fact, the new
routing, filtering, and classifying code is more featureful than the one
provided by many dedicated routers and firewalls and traffic shaping
products.

As new networking concepts have been invented, people have found ways to
plaster them on top of the existing framework in existing OSes. This
constant layering of cruft has lead to networking code that is filled with
strange behaviour, much like most human languages. In the past, Linux
emulated SunOS's handling of many of these things, which was not ideal.  

This new framework makes it possible to clearly express features
previously beyond Linux's reach.

<sect1>iproute2 tour
<p>
Linux has a sophisticated system for bandwidth provisioning called Traffic
Control. This system supports various method for classifying, prioritizing,
sharing, and limiting both inbound and outbound traffic.


We'll start off with a tiny tour of iproute2 possibilities.
<sect1>Prerequisites
<p>
You should make sure that you have the userland tools installed. This
package is called 'iproute' on both RedHat and Debian, and may otherwise be
found at <tt>ftp://ftp.inr.ac.ru/ip-routing/iproute2-2.2.4-now-ss??????.tar.gz"</tt>. 

You can also try <url name="here" url="ftp://ftp.inr.ac.ru/ip-routing/iproute2-current.tar.gz"> 
for the latest version.

Some parts of iproute require you to have certain kernel options enabled. It
should also be noted that all releases of RedHat up to and including 6.2
come without most of the traffic control features in the default kernel. 

RedHat 7.2 has everything in by default.

Also make sure that you have netlink support, should you choose to roll your
own kernel. Iproute2 needs it.

<sect1>Exploring your current configuration
<p>
This may come as a surprise, but iproute2 is already configured! The current
commands <tt>ifconfig</tt> and <tt>route</tt> are already using the advanced
syscalls, but mostly with very default (ie. boring) settings.

The <tt>ip</tt> tool is central, and we'll ask it to display our interfaces
for us.
<sect2><tt>ip</tt> shows us our links
<p>
<tscreen><verb>
[ahu@home ahu]$ ip link list
1: lo: <LOOPBACK,UP> mtu 3924 qdisc noqueue 
    link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00
2: dummy: <BROADCAST,NOARP> mtu 1500 qdisc noop 
    link/ether 00:00:00:00:00:00 brd ff:ff:ff:ff:ff:ff
3: eth0: <BROADCAST,MULTICAST,PROMISC,UP> mtu 1400 qdisc pfifo_fast qlen 100
    link/ether 48:54:e8:2a:47:16 brd ff:ff:ff:ff:ff:ff
4: eth1: <BROADCAST,MULTICAST,PROMISC,UP> mtu 1500 qdisc pfifo_fast qlen 100
    link/ether 00:e0:4c:39:24:78 brd ff:ff:ff:ff:ff:ff
3764: ppp0: <POINTOPOINT,MULTICAST,NOARP,UP> mtu 1492 qdisc pfifo_fast qlen 10
    link/ppp 

</verb></tscreen>
<p>Your mileage may vary, but this is what it shows on my NAT router at
home. I'll only explain part of the output as not everything is directly
relevant.

We first see the loopback interface. While your computer may function
somewhat without one, I'd advise against it. The MTU size (Maximum Transfer
Unit) is 3924 octets, and it is not supposed to queue. Which makes sense
because the loopback interface is a figment of your kernel's imagination.

I'll skip the dummy interface for now, and it may not be present on your
computer. Then there are my two physical network interfaces, one at the side
of my cable modem, the other one serves my home ethernet segment.
Furthermore, we see a ppp0 interface.

Note the absence of IP addresses. iproute disconnects the concept of 'links'
and 'IP addresses'. With IP aliasing, the concept of 'the' IP address had
become quite irrelevant anyhow. 

It does show us the MAC addresses though, the hardware identifier of our
ethernet interfaces.
<sect2><tt>ip</tt> shows us our IP addresses
<p>
<tscreen><verb>
[ahu@home ahu]$ ip address show        
1: lo: <LOOPBACK,UP> mtu 3924 qdisc noqueue 
    link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00
    inet 127.0.0.1/8 brd 127.255.255.255 scope host lo
2: dummy: <BROADCAST,NOARP> mtu 1500 qdisc noop 
    link/ether 00:00:00:00:00:00 brd ff:ff:ff:ff:ff:ff
3: eth0: <BROADCAST,MULTICAST,PROMISC,UP> mtu 1400 qdisc pfifo_fast qlen 100
    link/ether 48:54:e8:2a:47:16 brd ff:ff:ff:ff:ff:ff
    inet 10.0.0.1/8 brd 10.255.255.255 scope global eth0
4: eth1: <BROADCAST,MULTICAST,PROMISC,UP> mtu 1500 qdisc pfifo_fast qlen 100
    link/ether 00:e0:4c:39:24:78 brd ff:ff:ff:ff:ff:ff
3764: ppp0: <POINTOPOINT,MULTICAST,NOARP,UP> mtu 1492 qdisc pfifo_fast qlen 10
    link/ppp 
    inet 212.64.94.251 peer 212.64.94.1/32 scope global ppp0
</verb></tscreen>
<p>
This contains more information. It shows all our addresses, and to which
cards they belong. 'inet' stands for Internet (IPv4). There are lots of other
address families, but these don't concern us right now.

Let's examine eth0 somewhat closer. It says that it is related to the inet
address '10.0.0.1/8'. What does this mean? The /8 stands for the number of
bits that are in the Network Address. There are 32 bits, so we have 24 bits
left that are part of our network. The first 8 bits of 10.0.0.1 correspond
to 10.0.0.0, our Network Address, and our netmask is 255.0.0.0.

The other bits are connected to this interface, so 10.250.3.13 is directly
available on eth0, as is 10.0.0.1 for example. 

With ppp0, the same concept goes, though the numbers are different. Its
address is 212.64.94.251, without a subnet mask. This means that we have a
point-to-point connection and that every address, with the exception of
212.64.94.251, is remote. There is more information however, it tells us
that on the other side of the link there is yet again only one address,
212.64.94.1. The /32 tells us that there are no 'network bits'.

It is absolutely vital that you grasp these concepts. Refer to the
documentation mentioned at the beginning of this HOWTO if you have trouble.

You may also note 'qdisc', which stands for Queueing Discipline. This will
become vital later on. 

<sect2><tt>ip</tt> shows us our routes
<p>
Well, we now know how to find 10.x.y.z addresses, and we are able to reach
212.64.94.1. This is not enough however, so we need instructions on how to
reach the world. The internet is available via our ppp connection, and it
appears that 212.64.94.1 is willing to spread our packets around the
world, and deliver results back to us.

<tscreen><verb>
[ahu@home ahu]$ ip route show
212.64.94.1 dev ppp0  proto kernel  scope link  src 212.64.94.251 
10.0.0.0/8 dev eth0  proto kernel  scope link  src 10.0.0.1 
127.0.0.0/8 dev lo  scope link 
default via 212.64.94.1 dev ppp0 
</verb></tscreen>

This is pretty much self explanatory. The first 4 lines of output explicitly
state what was already implied by <tt>ip address show</tt>, the last line
tells us that the rest of the world can be found via 212.64.94.1, our
default gateway. We can see that it is a gateway because of the word
via, which tells us that we need to send packets to 212.64.94.1, and that it
will take care of things.

For reference, this is what the old 'route' utility shows us:
<tscreen><verb>
[ahu@home ahu]$ route -n
Kernel IP routing table
Destination     Gateway         Genmask         Flags Metric Ref    Use
Iface
212.64.94.1     0.0.0.0         255.255.255.255 UH    0      0        0 ppp0
10.0.0.0        0.0.0.0         255.0.0.0       U     0      0        0 eth0
127.0.0.0       0.0.0.0         255.0.0.0       U     0      0        0 lo
0.0.0.0         212.64.94.1     0.0.0.0         UG    0      0        0 ppp0
</verb></tscreen>

<sect1>ARP
<p>
ARP is the Address Resolution Protocol as described in
<url url="http://www.faqs.org/rfcs/rfc826.html" name="RFC 826">.
ARP is used by a networked machine to resolve the hardware location/address of
another machine on the same
local network.  Machines on the Internet are generally known by their names
which resolve to IP
addresses.  This is how a machine on the foo.com network is able to communicate
with another machine which is on the bar.net network.  An IP address, though,
cannot tell you the physical location of a machine.  This is where ARP comes
into the picture.

Let's take a very simple example.  Suppose I have a network composed of several
machines.  Two of the machines which are currently on my network are foo
with an IP address of 10.0.0.1 and bar with an IP address of 10.0.0.2.
Now foo wants to ping bar to see that he is alive, but alas, foo has no idea
where bar is.  So when foo decides to ping bar he will need to send
out an ARP request.
This ARP request is akin to foo shouting out on the network "Bar (10.0.0.2)!
Where are you?"  As a result of this every machine on the network will hear
foo shouting, but only bar (10.0.0.2) will respond.  Bar will then send an
ARP reply directly back to foo which is akin
bar saying,
"Foo (10.0.0.1) I am here at 00:60:94:E9:08:12."  After this simple transaction
that's used to locate his friend on the network, foo is able to communicate
with bar until he (his arp cache) forgets where bar is (typically after
15 minutes on Unix).

Now let's see how this works.
You can view your machines current arp/neighbor cache/table like so:
<tscreen><verb>
[root@espa041 /home/src/iputils]# ip neigh show
9.3.76.42 dev eth0 lladdr 00:60:08:3f:e9:f9 nud reachable
9.3.76.1 dev eth0 lladdr 00:06:29:21:73:c8 nud reachable
</verb></tscreen>

As you can see my machine espa041 (9.3.76.41) knows where to find espa042 
(9.3.76.42) and
espagate (9.3.76.1).  Now let's add another machine to the arp cache.

<tscreen><verb>
[root@espa041 /home/paulsch/.gnome-desktop]# ping -c 1 espa043
PING espa043.austin.ibm.com (9.3.76.43) from 9.3.76.41 : 56(84) bytes of data.
64 bytes from 9.3.76.43: icmp_seq=0 ttl=255 time=0.9 ms

--- espa043.austin.ibm.com ping statistics ---
1 packets transmitted, 1 packets received, 0% packet loss
round-trip min/avg/max = 0.9/0.9/0.9 ms

[root@espa041 /home/src/iputils]# ip neigh show
9.3.76.43 dev eth0 lladdr 00:06:29:21:80:20 nud reachable
9.3.76.42 dev eth0 lladdr 00:60:08:3f:e9:f9 nud reachable
9.3.76.1 dev eth0 lladdr 00:06:29:21:73:c8 nud reachable
</verb></tscreen>

As a result of espa041 trying to contact espa043, espa043's hardware
address/location has now been added to the arp/neighbor cache.
So until the entry for
espa043 times out (as a result of no communication between the two) espa041
knows where to find espa043 and has no need to send an ARP request.

Now let's delete espa043 from our arp cache:

<tscreen><verb>
[root@espa041 /home/src/iputils]# ip neigh delete 9.3.76.43 dev eth0
[root@espa041 /home/src/iputils]# ip neigh show
9.3.76.43 dev eth0  nud failed
9.3.76.42 dev eth0 lladdr 00:60:08:3f:e9:f9 nud reachable
9.3.76.1 dev eth0 lladdr 00:06:29:21:73:c8 nud stale
</verb></tscreen>

Now espa041 has again forgotten where to find espa043 and will need to send
another ARP request the next time he needs to communicate with espa043.
You can also see from the above output that espagate (9.3.76.1) has been
changed to the "stale" state.  This means that the location shown is still
valid, but it will have to be confirmed at the first transaction to that
machine.

<sect>Rules - routing policy database
<p>
If you have a large router, you may well cater for the needs of different
people, who should be served differently. The routing policy database allows
you to do this by having multiple sets of routing tables. 

If you want to use this feature, make sure that your kernel is compiled with
the "IP: advanced router" and "IP: policy routing" features.

When the kernel needs to make a routing decision, it finds out which table
needs to be consulted. By default, there are three tables. The old 'route'
tool modifies the main and local tables, as does the ip tool (by default).

The default rules:
<tscreen><verb>
[ahu@home ahu]$ ip rule list
0:      from all lookup local 
32766:  from all lookup main 
32767:  from all lookup default
</verb></tscreen>

This lists the priority of all rules. We see that all rules apply to all
packets ('from all'). We've seen the 'main' table before, it is output by
<tt>ip route ls</tt>, but the 'local' and 'default' table are new.

If we want to do fancy things, we generate rules which point to different
tables which allow us to override system wide routing rules.

For the exact semantics on what the kernel does when there are more matching
rules, see Alexey's ip-cref documentation. 

<sect1>Simple source routing
<p>
Let's take a real example once again, I have 2 (actually 3, about time I
returned them) cable modems, connected to a Linux NAT ('masquerading')
router. People living here pay me to use the internet. Suppose one of my
house mates only visits hotmail and wants to pay less. This is fine with me,
but you'll end up using the low-end cable modem.

The 'fast' cable modem is known as 212.64.94.251 and is a PPP link to
212.64.94.1. The 'slow' cable modem is known by various ip addresses,
212.64.78.148 in this example and is a link to 195.96.98.253.

The local table:
<tscreen><verb>
[ahu@home ahu]$ ip route list table local
broadcast 127.255.255.255 dev lo  proto kernel  scope link  src 127.0.0.1 
local 10.0.0.1 dev eth0  proto kernel  scope host  src 10.0.0.1 
broadcast 10.0.0.0 dev eth0  proto kernel  scope link  src 10.0.0.1 
local 212.64.94.251 dev ppp0  proto kernel  scope host  src 212.64.94.251 
broadcast 10.255.255.255 dev eth0  proto kernel  scope link  src 10.0.0.1 
broadcast 127.0.0.0 dev lo  proto kernel  scope link  src 127.0.0.1 
local 212.64.78.148 dev ppp2  proto kernel  scope host  src 212.64.78.148 
local 127.0.0.1 dev lo  proto kernel  scope host  src 127.0.0.1 
local 127.0.0.0/8 dev lo  proto kernel  scope host  src 127.0.0.1 
</verb></tscreen>

Lots of obvious things, but things that need to be specified somewhere.
Well, here they are. The default table is empty.

Let's view the 'main' table:
<tscreen><verb>
[ahu@home ahu]$ ip route list table main 
195.96.98.253 dev ppp2  proto kernel  scope link  src 212.64.78.148 
212.64.94.1 dev ppp0  proto kernel  scope link  src 212.64.94.251 
10.0.0.0/8 dev eth0  proto kernel  scope link  src 10.0.0.1 
127.0.0.0/8 dev lo  scope link 
default via 212.64.94.1 dev ppp0 
</verb></tscreen>

We now generate a new rule which we call 'John', for our hypothetical
house mate. Although we can work with pure numbers, it's far easier if we add
our tables to <file>/etc/iproute2/rt_tables</file>. 

<tscreen><verb>
# echo 200 John >> /etc/iproute2/rt_tables
# ip rule add from 10.0.0.10 table John
# ip rule ls
0:      from all lookup local 
32765:  from 10.0.0.10 lookup John
32766:  from all lookup main 
32767:  from all lookup default
</verb></tscreen>

Now all that is left is to generate Johns table, and flush the route cache:
<tscreen><verb>
# ip route add default via 195.96.98.253 dev ppp2 table John
# ip route flush cache
</verb></tscreen>

And we are done. It is left as an exercise for the reader to implement this
in ip-up.
<sect>GRE and other tunnels
<p>
There are 3 kinds of tunnels in Linux. There's IP in IP tunneling, GRE tunneling and tunnels that live outside the kernel (like, for example PPTP). 
<sect1>A few general remarks about tunnels:
<p>
Tunnels can be used to do some very unusual and very cool stuff. They can also make things go horribly wrong when you don't configure them right. Don't point your default route to a tunnel device unless you know <bf>exactly</bf> what you are doing :-). Furthermore, tunneling increases overhead, because it needs an extra set of IP headers. Typically this is 20 bytes per packet, so if the normal packet size (MTU) on a network is 1500 bytes, a packet that is sent through a tunnel can only be 1480 bytes big. This is not necessarily a problem, but be sure to read up on IP packet fragmentation/reassembly when you plan to connect large networks with tunnels. Oh, and of course, the fastest way to dig a tunnel is to dig at both sides.
<p>
<sect1>IP in IP tunneling
<p>
This kind of tunneling has been available in Linux for a long time. It requires 2 kernel modules,
ipip.o and new_tunnel.o.

Let's say you have 3 networks: Internal networks A and B, and intermediate network C (or let's say, Internet). 
So we have network A:

<tscreen><verb>
network 10.0.1.0
netmask 255.255.255.0
router  10.0.1.1
</verb></tscreen>
The router has address 172.16.17.18 on network C.

and network B:
<tscreen><verb>
network 10.0.2.0
netmask 255.255.255.0
router  10.0.2.1
</verb></tscreen>
The router has address 172.19.20.21 on network C.

As far as network C is concerned, we assume that it will pass any packet sent
from A to B and vice versa. You might even use the Internet for this.

Here's what you do:

First, make sure the modules are installed:

<tscreen><verb>
insmod ipip.o
insmod new_tunnel.o
</verb></tscreen>
Then, on the router of network A, you do the following:
<tscreen><verb>
ifconfig tunl0 10.0.1.1 pointopoint 172.19.20.21
route add -net 10.0.2.0 netmask 255.255.255.0 dev tunl0
</verb></tscreen>
And on the router of network B:
<tscreen><verb>
ifconfig tunl0 10.0.2.1 pointopoint 172.16.17.18
route add -net 10.0.1.0 netmask 255.255.255.0 dev tunl0
</verb></tscreen>
And if you're finished with your tunnel:
<tscreen><verb>
ifconfig tunl0 down
</verb></tscreen>
Presto, you're done. You can't forward broadcast or IPv6 traffic through
an IP-in-IP tunnel, though. You just connect 2 IPv4 networks that normally wouldn't be able to talk to each other, that's all. As far as compatibility goes, this code has been around a long time, so it's compatible all the way back to 1.3 kernels. Linux IP-in-IP tunneling doesn't work with other Operating Systems or routers, as far as I know. It's simple, it works. Use it if you have to, otherwise use GRE.

<sect1>GRE tunneling
<p>
GRE is a tunneling protocol that was originally developed by Cisco, and it
can do a few more things than IP-in-IP tunneling. For example, you can also
transport multicast traffic and IPv6 through a GRE tunnel.

In Linux, you'll need the ip_gre.o module.

<sect2>IPv4 Tunneling
<p>
Let's do IPv4 tunneling first:

Let's say you have 3 networks: Internal networks A and B, and intermediate network C (or let's say, Internet). 

So we have network A:
<tscreen><verb>
network 10.0.1.0
netmask 255.255.255.0
router  10.0.1.1
</verb></tscreen>
The router has address 172.16.17.18 on network C.
Let's call this network neta (ok, hardly original)

and network B:
<tscreen><verb>
network 10.0.2.0
netmask 255.255.255.0
router  10.0.2.1
</verb></tscreen>
The router has address 172.19.20.21 on network C.
Let's call this network netb (still not original)

As far as network C is concerned, we assume that it will pass any packet sent
from A to B and vice versa. How and why, we do not care.
<p>
On the router of network A, you do the following:
<tscreen><verb>
ip tunnel add netb mode gre remote 172.19.20.21 local 172.16.17.18 ttl 255
ip link set netb up
ip addr add 10.0.1.1 dev netb
ip route add 10.0.2.0/24 dev netb
</verb></tscreen>

Let's discuss this for a bit. In line 1, we added a tunnel device, and
called it netb (which is kind of obvious because that's where we want it to
go). Furthermore we told it to use the GRE protocol (mode gre), that the
remote address is 172.19.20.21 (the router at the other end), that our
tunneling packets should originate from 172.16.17.18 (which allows your
router to have several IP addresses on network C and let you decide which
one to use for tunneling) and that the TTL field of the packet should be set
to 255 (ttl 255).

The second line enables the device.

In the third line we gave the newly born interface netb the address
10.0.1.1. This is OK for smaller networks, but when you're starting up a
mining expedition (LOTS of tunnels), you might want to consider using
another IP range for tunneling interfaces (in this example, you could use
10.0.3.0).

<p>In the fourth line we set the route for network B. Note the different notation for the netmask. If you're not familiar with this notation, here's how it works: you write out the netmask in binary form, and you count all the ones. If you don't know how to do that, just remember that 255.0.0.0 is /8, 255.255.0.0 is /16 and 255.255.255.0 is /24. Oh, and 255.255.254.0 is /23, in case you were wondering.
<p>
But enough about this, let's go on with the router of network B.
<tscreen><verb>
ip tunnel add neta mode gre remote 172.16.17.18 local 172.19.20.21 ttl 255
ip link set neta up
ip addr add 10.0.2.1 dev neta
ip route add 10.0.1.0/24 dev neta
</verb></tscreen>
And when you want to remove the tunnel on router A:
<tscreen><verb>
ip link set netb down
ip tunnel del netb
</verb></tscreen>
Of course, you can replace netb with neta for router B.

<sect2>IPv6 Tunneling
<p>

A short bit about IPv6 addresses:<p>
IPv6 addresses are, compared to IPv4 addresses, monstrously big. An example:
<verb>3ffe:2502:200:40:281:48fe:dcfe:d9bc</verb>
So, to make writing them down easier, there are a few rules:
<itemize>
<item>Don't use leading zeroes. Same as in IPv4.
<item>Use colons to separate every 16 bits or two bytes.
<item>When you have lots of consecutive zeroes, you can write this down as ::. You can only do this once in an address and only for quantities of 16 bits, though.
</itemize>
Using these rules, the address 3ffe:0000:0000:0000:0000:0020:34A1:F32C can be written down as 3ffe::20:34A1:F32C, which is a lot shorter. 
<p>
On with the tunnels.

Let's assume that you have the following IPv6 network, and you want to connect it to 6bone, or a friend.

<tscreen><verb>
Network 3ffe:406:5:1:5:a:2:1/96
</verb></tscreen>
Your IPv4 address is 172.16.17.18, and the 6bone router has IPv4 address 172.22.23.24. 
<p>
<tscreen><verb>
ip tunnel add sixbone mode sit remote 172.22.23.24 local 172.16.17.18 ttl 255
ip link set sixbone up
ip addr add 3ffe:406:5:1:5:a:2:1/96 dev sixbone
ip route add 3ffe::/15 dev sixbone 
</verb></tscreen>

Let's discuss this. In the first line, we created a tunnel device called sixbone. We gave it mode sit (which is IPv6 in IPv4 tunneling) and told it where to go to (remote) and where to come from (local). TTL is set to maximum, 255. Next, we made the device active (up). After that, we added our own network address, and set a route for 3ffe::/15 (which is currently all of 6bone) through the tunnel.
<p>
GRE tunnels are currently the preferred type of tunneling. It's a standard that's also widely adopted outside the Linux community and therefore a Good Thing.
<p>
<sect1>Userland tunnels
<p>
There are literally dozens of implementations of tunneling outside the kernel. Best known are of course PPP and PPTP, but there are lots more (some proprietary, some secure, some that don't even use IP) and that is really beyond the scope of this HOWTO.

<sect>IPv6 tunneling with Cisco and/or 6bone
<p>
By Marco Davids &lt;marco@sara.nl&gt;

NOTE to maintainer:

As far as I am concerned, this IPv6-IPv4 tunneling is not per definition
GRE tunneling. You could tunnel IPv6 over IPv4 by means of GRE tunnel devices
(GRE tunnels ANY to IPv4), but the device used here ("sit") only tunnels
IPv6 over IPv4 and is therefore something different.

<sect1>IPv6 Tunneling
<p>
This is another application of the tunneling capabilities of Linux. It is
popular among the IPv6 early adopters, or pioneers if you like.
The 'hands-on' example described below is certainly not the only way
to do IPv6 tunneling. However, it is the method that is often used to tunnel
between Linux and a Cisco IPv6 capable router and experience learned that
this is just the thing many people are after. Ten to one this applies to
you too ;-)

A short bit about IPv6 addresses:

IPv6 addresses are, compared to IPv4 addresses, really big: 128 bits
against 32 bits. And this provides us just with the thing we need: many, many
IP-addresses: 340,282,266,920,938,463,463,374,607,431,768,211,465 to be
precise. Apart from this, IPv6 (or IPng, for IP Next Generation) is supposed
to provide for smaller routing tables on the Internet's backbone routers,
simpler configuration of equipment, better security at the IP level and
better support for QoS.

An example: 2002:836b:9820:0000:0000:0000:836b:9886

Writing down IPv6 addresses can be quite a burden. Therefore, to make
life easier there are some rules:

<itemize>
<item>
Don't use leading zeroes. Same as in IPv4.

<item>Use colons to separate every 16 bits or two bytes.

<item>When you have lots of consecutive zeroes,
you can write this down as ::. You can only do this once in an
address and only for quantities of 16 bits, though.
</itemize>

The address 2002:836b:9820:0000:0000:0000:836b:9886 can be written down
as 2002:836b:9820::836b:9886, which is somewhat friendlier.

IPv6 is intended to be the successor of the current IPv4. Because it
is relatively new technology, there is no worldwide native IPv6 network
yet. To be able to move forward swiftly, the 6bone was introduced. 

Native IPv6 networks are connected to each other by encapsulating the IPv6
protocol in IPv4 packets and sending them over the existing IPv4 infrastructure
from one IPv6 site to another. 

That is precisely where the tunnel steps in.

To be able to use IPv6, we should have a kernel that supports it. There
are many good documents on how to achieve this. But it all comes down to
a few steps:
<itemize>
<item>Get yourself a recent Linux distribution, with suitable glibc.
<item>Then get yourself an up-to-date kernel source.
</itemize>
If you are all set, then you can go ahead and compile an IPv6 capable
kernel:
<itemize>
<item>Go to /usr/src/linux and type:
<item>make menuconfig
<item>Choose "Networking Options"
<item>Select "The IPv6 protocol", "IPv6: enable EUI-64 token format", "IPv6:
disable provider based addresses"
</itemize>
HINT: Don't go for the 'module' option. Often this won't work well.

In other words, compile IPv6 as 'built-in' in your kernel.
You can then save your config like usual and go ahead with compiling
the kernel.

HINT: Before doing so, consider editing the Makefile:
EXTRAVERSION = -x ; --> ; EXTRAVERSION = -x-IPv6

There is a lot of good documentation about compiling and installing
a kernel, however this document is about something else. If you run into
problems at this stage, go and look for documentation about compiling a
Linux kernel according to your own specifications.

The file /usr/src/linux/README might be a good start.
After you acomplished all this, and rebooted with your brand new kernel,
you might want to issue an '/sbin/ifconfig -a' and notice the brand 
new 'sit0-device'. SIT stands for Simple Internet Transition. You may give
yourself a compliment; you are now one major step closer to IP, the Next
Generation ;-)

Now on to the next step. You want to connect your host, or maybe even
your entire LAN to another IPv6 capable network. This might be the "6bone"
that is setup especially for this particular purpose.

Let's assume that you have the following IPv6 network: 3ffe:604:6:8::/64 and
you want to connect it to 6bone, or a friend. Please note that the /64
subnet notation works just like with regular IP adresses.

Your IPv4 address is 145.100.24.181 and the 6bone router has IPv4 address
145.100.1.5
<tscreen><verb>
# ip tunnel add sixbone mode sit remote 145.100.1.5 [local 145.100.24.181 ttl 225]
# ip link set sixbone up
# ip addr add 3FFE:604:6:7::2/126 dev sixbone
# ip route add 3ffe::0/16 dev sixbone
</verb></tscreen>

Let's discuss this. In the first line, we created a tunnel device called
sixbone. We gave it mode sit (which is IPv6 in IPv4 tunneling) and told it
where to go to (remote) and where to come from (local). TTL is set to
maximum, 255. 

Next, we made the device active (up). After that, we added our own network
address, and set a route for 3ffe::/15 (which is currently all of 6bone)
through the tunnel. If the particular machine you run this on is your IPv6
gateway, then consider adding the following lines:

<tscreen><verb>
# echo 1 >/proc/sys/net/ipv6/conf/all/forwarding
# /usr/local/sbin/radvd
</verb></tscreen>
The latter, radvd is -like zebra- a router advertisement daemon, to
support IPv6's autoconfiguration features. Search for it with your favourite
search-engine if you like.
You can check things like this:

<tscreen><verb>
# /sbin/ip -f inet6 addr
</verb></tscreen>

If you happen to have radvd running on your IPv6 gateway and boot your
IPv6 capable Linux on a machine on your local LAN, you would be able to
enjoy the benefits of IPv6 autoconfiguration:
<tscreen><verb>
# /sbin/ip -f inet6 addr
1: lo: &lt;LOOPBACK,UP&gt; mtu 3924 qdisc noqueue inet6 ::1/128 scope host

3: eth0: &lt;BROADCAST,MULTICAST,UP&gt; mtu 1500 qdisc pfifo_fast qlen 100
inet6 3ffe:604:6:8:5054:4cff:fe01:e3d6/64 scope global dynamic
valid_lft forever preferred_lft 604646sec inet6 fe80::5054:4cff:fe01:e3d6/10 
scope link
</verb></tscreen>

You could go ahead and configure your bind for IPv6 addresses. The A
type has an equivalent for IPv6: AAAA. The in-addr.arpa's equivalent is:
ip6.int. There's a lot of information available on this topic.

There is an increasing number of IPv6-aware applications available,
including secure shell, telnet, inetd, Mozilla the browser, Apache the
websever and a lot of others. But this is all outside the scope of this
Routing document ;-)

On the Cisco side the configuration would be something like this:
<tscreen><verb>
!
interface Tunnel1
description IPv6 tunnel
no ip address
no ip directed-broadcast
ipv6 enable
ipv6 address 3FFE:604:6:7::1/126
tunnel source Serial0
tunnel destination 145.100.24.181
tunnel mode ipv6ip
!
ipv6 route 3FFE:604:6:8::/64 Tunnel1
</verb></tscreen>
But if you don't have a Cisco at your disposal, try one of the many
IPv6 tunnel brokers available on the Internet. They are willing to configure
their Cisco with an extra tunnel for you. Mostly by means of a friendly
web interface. Search for "ipv6 tunnel broker" on your favourite search engine.

<sect>IPsec: secure IP over the internet
<p>
FIXME: editor vacancy.
In the meantime, see: <url url="http://www.freeswan.org/" name="The
FreeS/WAN project">. Another IPSec implementation for Linux is Cerberus,
by NIST. However, their web pages have not been updated in over a year,
and their version tended to trail well behind the current Linux kernel.
USAGI, an alternative IPv6 implementation for Linux, also includes an
IPSec implementation, but that might only be for IPv6.

<sect>Multicast routing
<p>
FIXME: Editor Vacancy!
<p>
The Multicast-HOWTO is ancient (relatively-speaking) and may be inaccurate
or misleading in places, for that reason.
<p>
Before you can do any multicast routing, you need to configure the Linux
kernel to support the type of multicast routing you want to do. This, in
turn, requires you to decide what type of multicast routing you expect to
be using. There are essentially four "common" types - DVMRP (the Multicast
version of the RIP unicast protocol), MOSPF (the same, but for OSPF), PIM-SM
("Protocol Independent Multicasting - Sparse Mode", which assumes that users
of any multicast group are spread out, rather than clumped) and PIM-DM (the
same, but "Dense Mode", which assumes that there will be significant clumps
of users of the same multicast group).
<p>
In the Linux kernel, you will notice that these options don't appear. This is
because the protocol itself is handled by a routing application, such as
Zebra, mrouted, or pimd. However, you still have to have a good idea of which
you're going to use, to select the right options in the kernel.
<p>
For all multicast routing, you will definitely need to enable "multicasting"
and "multicast routing". For DVMRP and MOSPF, this is sufficient. If you are
going to use PIM, you must also enable PIMv1 or PIMv2, depending on whether
the network you are connecting to uses version 1 or 2 of the PIM protocol.
<p>
Once you have all that sorted out, and your new Linux kernel compiled, you
will see that the IP protocols listed, at boot time, now include IGMP. This
is a protocol for managing multicast groups. At the time of writing, Linux
supports IGMP versions 1 and 2 only, although version 3 does exist and has
been documented. This doesn't really affect us that much, as IGMPv3 is still
new enough that the extra capabilities of IGMPv3 aren't going to be that
much use. Because IGMP deals with groups, only the features present in the
simplest version of IGMP over the entire group are going to be used. For the
most part, that will be IGMPv2, although IGMPv1 is sill going to be
encountered.
<p>
So far, so good. We've enabled multicasting. Now, we have to tell the Linux
kernel to actually do something with it, so we can start routing. This means
adding the Multicast virtual network to the router table:
<p>
ip route add 224.0.0.0/4 dev eth0
<p>
(Assuming, of course, that you're multicasting over eth0! Substitute the
device of your choice, for this.)
<p>
Now, tell Linux to forward packets...
<p>
echo 1 > /proc/sys/net/ipv4/ip_forward
<p>
At this point, you may be wondering if this is ever going to do anything. So,
to test our connection, we ping the default group, 224.0.0.1, to see if anyone
is alive. All machines on your LAN with multicasting enabled <em>should</em>
respond, but nothing else. You'll notice that none of the machines that
respond have an IP address of 224.0.0.1. What a surprise! :) This is a group
address (a "broadcast" to subscribers), and all members of the group will
respond with their own address, not the group address.
<p>
ping -c 2 224.0.0.1
<p>
At this point, you're ready to do actual multicast routing. Well, assuming
that you have two networks to route between.

(To Be Continued!)

<sect>Queueing Disciplines for Bandwidth Management
<p>
Now, when I discovered this, it <em>really</em> blew me away. Linux 2.2/2.4
comes with everything to manage bandwidth in ways comparable to high-end
dedicated bandwidth management systems.

Linux even goes far beyond what Frame and ATM provide. 

<sect1>Queues and Queueing Disciplines explained
<p>
With queueing we determine the way in which data is <em>sent</em>. It is
important to realise that we can only shape data that we transmit.

With the way the internet works, we have no direct control of what people
send us. It's a bit like your (physical!) mailbox at home. There is no way
you can influence the world to modify the amount of mail they send you,
short of contacting everybody.

However, the internet is mostly based on TCP/IP which has a few features
that help us. TCP/IP has no way of knowing the capacity of the network
between two hosts, so it just starts sending data faster and faster ('slow
start') and when packets start getting lost, because there is no room to
send them, it will slow down. In fact it is a bit smarter than this, but
more about that later.

This is the equivalent of not reading half of your mail, and hoping that
people will stop sending it to you. With the difference that it works for
the Internet :-)

If you have a router and wish to prevent certain hosts of networks from
downloading too fast, you need to do your shaping on the *inner* interface
of your router, the one that sends data to your own computers.

<sect1>Simple, classless Queueing Disciplines
<p>
As said, with queueing disciplines, we change the way data is sent.
Classless queueing disciplines are those that, by and large accept data and
only reorder, delay or drop it.

These can be used to shape traffic for an entire interface, without any
subdivisions. It is vital that you understand this part of queueing before
we go on the the classful qdisc-containing-qdiscs!

By far the most widely used discipline is the pfifo_fast queue - this is the
default. This also explains why these advanced features are so robust. They
are nothing more than 'just another queue'.

Each of these queues has specific strengths and weaknesses. Not all of them
may be as well tested.

<sect2>pfifo_fast
<p>
This queue is, as the name says, First In, First Out, which means that no
packet receives special treatment. At least, not quite. This queue has 3 so
called 'bands'. Within each band, FIFO rules apply. However, as long as
there are packets waiting in band 0, band 1 won't be processed. Same goes
for band 1 and band 2.

The kernel honors the so called Type of Service flag of packets, and takes
care to insert 'minimum delay' packets in band 0. 

<sect2>Token Bucket Filter
<p>
The Token Bucket Filter (TBF) is a simple queue that only passes packets
arriving at a rate which is not exceeding some administratively set rate,
with the possibility to allow short bursts in excess of this rate.

TBF is very precise, network- and processor friendly. It should be your
first choice if you simple want to slow an interface down!

The TBF implementation consists of a buffer (bucket), constantly filled by
some virtual pieces of information called tokens, at a specific rate (token
rate). The most important parameter of the bucket is its size, that is the
number of tokens it can store.

Each arriving token collects one incoming data packet from the data queue
and is then deleted from the bucket. Associating this algorithm
with the two flows -- token and data, gives us three possible scenarios:

<itemize>
<item> The data arrives in TBF at a rate that's <em>equal</em> to the rate
of incoming tokens. In this case each incoming packet has its matching token
and passes the queue without delay.

<item> The data arrives in TBF at a rate that's <em>smaller</em> than the
token rate. Only a part of the tokens are deleted at output of each data packet
that's sent out the queue, so the tokens accumulate, up to the bucket size.
The unused tokens can then be used to send data a a speed that's exceeding the
standard token rate, in case short data bursts occur.

<item> The data arrives in TBF at a rate <em>bigger</em> than the token rate.
This means that the bucket will soon be devoid of tokens, which causes the
TBF to throttle itself for a while. This is called an 'overlimit situation'.
If packets keep coming in, packets will start to get dropped.
</itemize>

The last scenario is very important, because it allows to
administratively shape the bandwidth available to data that's passing
the filter.

The accumulation of tokens allows a short burst of overlimit data to be
still passed without loss, but any lasting overload will cause packets to be
constantly delayed, and then dropped.

<sect2>Stochastic Fairness Queueing
<p>
Stochastic Fairness Queueing (SFQ) is a simple implementation of the fair
queueing algorithms family. It's less accurate than others, but it also
requires less calculations while being almost perfectly fair.

The key word in SFQ is conversation (or flow), which mostly corresponds to a
TCP session or a UDP stream. Traffic is divided into a pretty large number
of FIFO queues, one for each conversation. Traffic is then sent in a round
robin fashion, giving each session the chance to send data in turn.

This leads to very fair behaviour and disallows any single conversation from
drowning out the rest. SFQ is called 'Stochastic' because it doesn't really
allocate a queue for each session, it has an algorithm which divides traffic
over a limited number of queues using a hashing algorithm. 

Because of the hash, multiple sessions might end up in the same bucket, which
would halve each session's chance of sending a packet, thus halving the
effective speed available. To prevent this situation from becoming
noticeable, SFQ changes its hashing algorithm quite often so that any two
colliding sessions will only do so for a small number of seconds.

It is important to note that SFQ is only useful in case your actual outgoing
interface is really full! If it isn't then there will be no queue on your
linux machine and hence no effect. Later on we will describe how to combine
SFQ with other qdiscs to get a best-of-both worlds situation.

Specifically, setting SFQ on the ethernet interface heading to your
cablemodem or DSL router is pointless without further shaping!
<sect2>Random Early Drop
<p>
RED has some extra smartness built in. When a TCP/IP session starts, neither
end knows the amount of bandwidth available. So TCP/IP starts to transmit
slowly and goes faster and faster, though limited by the latency at which
ACKs return.

Once a link is filling up, RED starts dropping packets, which indicate to
TCP/IP that the link is congested, and that it should slow down. The smart
bit is that RED simulates real congestion, and starts to drop some packets
some time before the link is entirely filled up. Once the link is completely
saturated, it behaves like a normal policer.

For more information on this, see the Backbone chapter.

<sect2>Ingress policer qdisc
<p> 
The Ingress qdisc comes in handy if you need to ratelimit a host without
help from routers or other Linux boxes. You can police incoming bandwidth
and drop packets when this bandwidth exceeds your desired rate. This can
save your host from a SYN flood, for example, and also works to slow down
TCP/IP, which responds to dropped packets by reducing speed.

FIXME: shaping by dropping packets seems less desirable than using, for
example, a token bucket filter. Not sure though, Cisco CAR works this
way, and people appear to be happy with it.

See the reference to <ref id="CAR" name="IOS Committed Access Rate"> at the
end of this document.

In short: you can use this to limit how fast your computer downloads files,
thus leaving more of the available bandwidth for others.

See the section on 'Protecting your host from SYN floods' for an example on
how this works.

<sect2>Summary
<p>
Summarizing, these are the simple queues that actually manage traffic by
reordering, slowing or dropping packets.

Advice:
<itemize>
<item>
To purely slow down outgoing traffic, use the Token Bucket Filter.
<item>
If your link is truly full and you want to make sure that no single session
can dominate your outgoing bandwidth, use Stochastical Fairness Queueing.
<item>
If you have a big backbone and know what you are doing, consider Random
Early Drop - see the backbone chapter.
<item>
To 'shape' incoming traffic which you are not forwarding, use the Ingress
Policer. If you *are* forwarding it, use a TBF on the interface you are
forwarding the data to.
</itemize>
<sect1>Classful Queueing Disciplines
<p>
Some queueing disciplines can contain other queueing disciplines, which are
then suddenly called 'classes'. A class is nothing short of a qdisc, except
that it lives within another qdisc. We use the terms 'inner qdisc'
, 'sub-qdisc' and 'class' interchangeably.

Classful qdiscs are very useful if you have different kinds of traffic which
should have differing treatment. One of the classful qdiscs is called 'CBQ'
, 'Class Based Queueing' - it is so widely mentioned that people identify
queueing with classes solely with CBQ, but this is not the case.

CBQ is merely the oldest kid on the block - yet it is by far the least
useful qdisc and also the most complex one. I advise *against* using it.
This may come as something of a shock to many who fell for the 'sendmail
effect', which learns us that any complex technology which doesn't come with
documentation must be the best available.

<sect2>Flow within classful qdiscs &amp; classes
<p>
When traffic enters a classful qdisc, it needs to be sent to any of the
qdiscs within - the classes. To determine what to do with a packet, the so
called 'filters' are consulted. It is important to know that the filters are
called from within a qdisc, and not the other way around!

The filters then return with a decision, and the qdisc uses this to enqueue
the packet into one of the classes. These classes don't know that they are
part of an outer-qdisc, they act as they normally do: accepting packets on
one end and outputting them again when asked.

Besides containing other qdiscs, most classful qdiscs also perform shaping.
This is useful to perform both packet reordering (with SFQ, for example) and
rate control. You need this in cases where you have a high speed
interface (for example, ethernet) to a slower device (a cable modem).

If you were only to run SFQ, nothing would happen, as packets enter &amp;
leave your router without delay: the output interface is far faster than
your actual link speed. There is no queue to process then.

<sect2>The qdisc family: roots, handles, siblings and parents
<p>
Each interface has a 'root qdisc', by default the earlier mentioned
classless pfifo_fast queueing discipline. Each qdisc can be assigned a
handle, which can be used by later configuration statements to refer to that
qdisc.

These handles consist of two parts, a major number and a minor number. It is
habitual to name the root qdisc '1:', which is equal to '1:0'. 

Because classful qdiscs can have a lot of children, you should give them
their own namespace by giving them a separate major number.

Recapping, a typical hierarchy might look like this:
<verb>
                    root 1:
                      |
                    _1:1_
                   /  |  \
                  /   |   \
                 /    |    \
               10:   11:   12:
              /   \       /   \
           10:1  10:2   12:1  12:2
</verb>

But don't let this tree fool you! You should *not* imagine the kernel to be
at the apex of the tree and the network below, that is just not the case.
Packets get enqueued and dequeued at the root qdisc, which is the only thing
the kernel talks to. 

A packet might get enqueued in a chain like this:

        1: -> 1:1 -> 12: -> 12:2

The packet now resides in a queue in qdisc 12:2.

When the kernel decides that it needs to extract packets to send to the
interface, the root qdisc 1: gets a dequeue request, which is passed to
1:1, which is in turn passed to 10:, 11: and 12:, which each query their
siblings, and try to dequeue() from them. In this case, the kernel needs to
walk the entire tree, because only 12:2 contains a packet. 

In short, nested qdiscs ONLY talk to their parent qdiscs, never to an
interface. Only the root qdisc gets dequeued by the kernel! 

The upshot of this is that sub-qdiscs never get dequeued faster than their
parents allow. And this is exactly what we want: this way we can have SFQ as
an inner class, which doesn't do any shaping, only reordering, and have a
shaping outer class, which does the shaping.
<sect2>The PRIO qdisc
<p>
The PRIO qdisc doesn't actually shape, it only subdivides traffic based on
how you configured your filters. You can consider the PRIO qdisc a kind
of pfifo_fast on stereoids, whereby each band is a separate qdisc instead of
a simple FIFO.

When a packet is enqueued to the PRIO qdisc, a sub-qdisc is chosen based on
the filter commands you gave. By default, three pfifo sub-qdiscs are
created. These sub-qdiscs are pure FIFO queues with no internal structure,
but you can replace them by any qdisc you have available.

This queue is very useful in case you want to prioritize certain kinds of
traffic without using TOS-flags but using all the power of the tc filters.
Because it doesn't actually shape, the same warning as for SFQ holds: either
use it only if your physical link is really full or wrap it inside a
classful qdisc that does shape.

The last holds for almost all cablemodems and DSL devices.
<sect2>The famous CBQ qdisc
<p>
As said before, CBQ is the most complex qdisc available, the most hyped, the
least understood and possibly the worst queueing discipline for you in the
entire Linux kernel. This is not because the authors are evil or
incompetent, far from it, it's just that the CBQ algorithm isn't all that
precise and doesn't really match the way Linux works.

Besides being classful, CBQ is also a shaper and it is in that aspect that
it really doesn't work very well. It should work like this. If you try to
shape a 10mbit/s connection to 1mbit/s, the link should be idle 90% of the
time. If it isn't, we need to throttle so that it IS idle 90% of the time.

This is pretty hard to measure, so CBQ also needs to know how big an average
packet is going to be, and instead derives the idle time from the number of
microseconds that elapse between requests from the hardware layer for more
data. Combined, this can be used to approximate how full or empty the link
is.

This is rather circumspect and doesn't always arrive at proper results. For
example, what is the actual link speed of an interface that is not really
able to transmit the full 100mbit/s of data, perhaps because of a badly
implemented driver? A PCMCIA network card will also never achieve 100mbit/s
because of the way the bus is designed - again, how do we calculate the idle
time?

It gets even worse if we consider not-quite-real network devices like PPP
over Ethernet or PPTP over TCP/IP. The effective bandwidth in that case is
probably determined by the efficiency of pipes to userspace - which is huge.

People who have done measurements discover that CBQ is not very accurate and
sometimes completely misses the mark.

Besides not being all that good, it also comes with little or no
documentation AND with about 20 knobs to tune. In fact, CBQ is so hard to
configure that people use scripts to generate the needed commands.

In short - I would advise AGAINST using CBQ if you want to have accurate
results and want to understand what you are doing.
<sect3>CBQ shaping in detail
<p>
Because lots of people are using CBQ anyhow, possibly because they don't
have anything else available, we will describe it here.

As said before, CBQ works by making sure that the link is idle just long
enough to bring down the real bandwidth to the configured rate. To do so, it
calculates the time that should pass between average packets. 

During operations, the effective idletime is measured using an exponential
weighted moving average (EWMA), which considers recent packets to be
exponentially more important than past ones. The unix loadaverage is
calculated in the same way.

The calculated idle time is substracted from the EWMA measured one, the
resulting number is called 'avgidle'. A perfectly loaded link has an avgidle
of zero: packets arrive exactly once every calculated interval.  

An overloaded link has a negative avgidle and if it gets too negative, CBQ
shuts down for a while and is then 'overlimit'.

Conversely, an idle link might amass a huge avgidle, which would then allow
infinite bandwidths after a few hours of silence. To prevent this, avgidle is
capped at maxidle.

If overlimit, in theory, the CBQ could throttle itself for exactly the
amount of time that was calculated to pass between packets, and then pass
one packet, and throttle again. But see the 'minburst' parameter below.

These are parameters you can specify in order to configure shaping:

<descrip>
<tag>avpkt</tag>
Average size of a packet, measured in bytes. Needed for idle time
approximation.
<tag>bandwidth</tag>
The physical bandwidth of your device, also needed for idle time
calculations.
<tag>cell</tag>
The time a packet takes to be transmitted over an ethernet device grows in
steps, based on the packet size. An 800 and a 806 size packet may take just
as long to send, for example - this sets the granularity. Most often set 
to '8'. Must be an integral power of two.
<tag>maxburst</tag>
This number of packets is used to calculate maxidle so that when avgidle is
at maxidle, this number of average packets can be burst before avgidle drops
to 0. Set it higher to be more tolerant of bursts. You can't set maxidle
directly, only via this parameter.
<tag>minburst</tag>
As mentioned before, CBQ needs to throttle in case of overlimit. The ideal
solution is to do so for exactly the calculated idle time, and pass 1
packet. However, Unix kernels generally have a hard time scheduling events
shorter than 10ms, so it is better to throttle for a longer period, and then
pass minburst packets in one go, and then sleep minburst times longer.

The time to wait is called the offtime. Higher values of minburst lead to
more accurate shaping in the long term, but to bigger bursts at millisecond
timescales.
<tag>minidle</tag>
If avgidle is below 0, we are overlimits and need to wait until avgidle will
be big enough to send one packet. To prevent a sudden burst from shutting
down the link for a prolonged period of time, avgidle is reset to minidle if
it gets too low.

Minidle is specified in negative microseconds, so 10 means that avgidle is
capped at -10us.
<tag>mpu</tag>
Mininum packet size - needed because even a zero size packet is padded
to 64 bytes on ethernet, and so takes a certain time to transmit. CBQ needs
to know this to accurately measure the idle time.
<tag>rate</tag>
Desired rate of traffic leaving this qdisc - this is the 'speed knob'!
</descrip>

Internally, CBQ has a lot of finetuning. For example, classes which are
known not to have data enqueued to them aren't queried. Overlimit classes
are penalized by lowering their effective priority. All very smart &amp;
complicated.

<sect3>CBQ classful behaviour
<p>
Besides shaping, using the aforementioned idletime approximations, CBQ also
acts like the PRIO queue in the sense that classes can have differing
priorities and that lower priority numbers will be polled before the higher
priority ones.

Each time a packet is requested by the hardware layer to be sent out to the
network, a weighted round robin process starts, beginning with the lower
priority classes. 

These are then grouped and queried if they have data available. If so, it is
returned. After a class has been allowed to dequeue a number of bytes, the
next class within that priority is tried.

The following parameters control the WRR process:
<descrip>
<tag>allot</tag>
When the outer cbq is asked for a packet to send out on the interface, it
will try all inner qdiscs (classes) in turn, in order of the 'priority'
parameter. Each time a class gets its turn, it can only send out a limited
amount of data. 'Allot' is the base unit of this amount. See the 'weight'
parameter for more information.


<tag>prio</tag>
The CBQ can also act like the PRIO device. Inner classes with lower priority
are tried first and as long as they have traffic, other classes are not
polled for traffic.

<tag>weight</tag>
Weight helps in the Weighted Round Robin process. Each class gets a chance
to send in turn. If you have classes with significantly more bandwidth than
other classes, it makes sense to allow them to send more data in one round
than the others.

A CBQ adds up all weights within a class, and normalizes them, so you can
use arbitrary numbers: only the ratios are important. People have been 
using 'rate/10' as a rule of thumb and it appears to work well. The
renormalized weight is multiplied by the 'allot' parameter to determine how
much data can be sent in one round.
</descrip>

<sect3>CBQ parameters that determine link sharing &amp; borrowing
<p>
Besides purely limiting certain kinds of traffic, it is also possible to
specify which classes can borrow capacity from other classes or, conversely,
lend out bandwidth.

<descrip>
<tag>Isolated/sharing</tag>
A class that is configured with 'isolated' will not lend out bandwidth to
sibling classes. Use this if you have competing or mutually-unfriendly
agencies on your link who do want to give eachother freebies.

The control program tc also knows about 'sharing', which is the reverse 
of 'isolated'.
<tag>bounded/borrow</tag>
A class can also be 'bounded', which means that it will not try to borrow
bandwidth from sibling classes. tc also knows about 'borrow', which is the
reverse of 'bounded'.
</descrip>
A typical situation might be where you have two agencies on your link which
are both 'isolated' and 'bounded', which means that they are really limited
to their assigened rate, and also won't allow each other to borrow.

Within such an agency class, there might be other classes which are allowed
to swap bandwidth.

<sect3>Other CBQ parameters: split &amp; defmap
<p>
As said before, a classful qdisc needs to call the filter chain to determine
which class a packet will be enqueued to. 

Besides calling the filter, CBQ offers other options, defmap &amp; split.
This is pretty complicated to understand, and it is not vital. But as this
is the only known place where defmap &amp; split are properly explained, I'm
doing my best. 

As you will often want to filter on the Type of Service field only, a special
syntax is provided. Whenever the CBQ needs to figure out where a packet
needs to be enqueued, it checks if this node is a 'split node'. If so, one
of the sub-qdiscs has indicated that it wishes to receive all packets with
a certain configured priority, as might be derived from the TOS field, or
socket options set by applications.

The packets' priority bits are or-ed with the defmap field to see if a match
exists. In other words, this is a short-hand way of creating a very fast
filter, which only matches certain priorities. A defmap of ff (hex) will
match everything, a map of 0 nothing. A sample configuration may help make
things clearer:

<verb>
# tc qdisc add dev eth1 root handle 1: cbq bandwidth 10Mbit allot 1514 \
  cell 8 avpkt 1000 mpu 64
 
# tc class add dev eth1 parent 1:0 classid 1:1 cbq bandwidth 10Mbit    \
  rate 10Mbit allot 1514 cell 8 weight 1Mbit prio 8 maxburst 20        \
  avpkt 1000
</verb>
Standard CBQ preamble. I never get used to the sheer amount of numbers
required!

Defmap refers to TC_PRIO bits, which are defined as follows:

<verb>
TC_PRIO_BESTEFFORT              0
TC_PRIO_FILLER                  1
TC_PRIO_BULK                    2
TC_PRIO_INTERACTIVE_BULK        4
TC_PRIO_INTERACTIVE             6
TC_PRIO_CONTROL                 7
</verb>

This corresponds to bits, counted from the right. Now the interactive, and
the bulk classes:

<verb>
# tc class add dev eth1 parent 1:1 classid 1:2 cbq bandwidth 10Mbit     \
  rate 1Mbit allot 1514 cell 8 weight 100Kbit prio 3 maxburst 20        \
  avpkt 1000 split 1:0 defmap c0

# tc class add dev eth1 parent 1:1 classid 1:3 cbq bandwidth 10Mbit     \
  rate 8Mbit allot 1514 cell 8 weight 800Kbit prio 7 maxburst 20        \
  avpkt 1000 split 1:0 defmap 3f
</verb>

The 'split qdisc' is 1:0, which is where the choice will be made. C0 is
binary for 11000000, 3F for 00111111, so these two together will match
everything. The first class matches bits 7 & 6, and thus corresponds 
to 'interactive' and 'control' traffic. The second class matches the rest.

Node 1:0 now has a table like this:
<verb>
priority        send to
0               1:3
1               1:3
2               1:3
3               1:3
4               1:3
5               1:3
6               1:2
7               1:2
</verb>


For additional fun, you can also pass a 'change mask', which indicates
exactly which priorities you wish to change. You only need to use this if you
are running 'tc class change'. For example, to add best effort traffic to
1:2, we could run this:

<verb>
# tc class change dev eth1 classid 1:2 cbq defmap 01/01
</verb>

The priority map over at 1:0 now looks like this:

<verb>
priority        send to
0               1:2
1               1:3
2               1:3
3               1:3
4               1:3
5               1:3
6               1:2
7               1:2
</verb>

FIXME: did not test this, only looked at the source.
<sect2>Hierarchical Token Bucket 
<p>
Martin Devera (&lt;devik&gt;) rightly realised that CBQ is complex and does
not seem optimized for many typical situations. His Hierarchial approach is
well suited for setups where you have a fixed amount of bandwidth which you
want to divide for different purposes, giving each purpose a guaranteed
bandwidth, with the possibility of specifying how much bandwidth can be
borrowed.

HTB works just like CBQ but does not resort to idle time calculations to
shape. Instead, it is a classful Token Bucket Filter - hence the name. It
has only a few parameters, which are well documented on his 
<url url="http://luxik.cdi.cz/~devik/qos/htb/"
 name="site">.

As your HTB configuration gets more complex, your configuration scales
well. With CBQ it is already complex even in simple cases! HTB is not yet a
part of the standard kernel, but it should soon be!

If you are in a position to patch your kernel, by all means use HTB instead
of CBQ.
<sect1>Classifying packets with filters
<p>
To determine which class shall process a packet, the so-called 'classifier
chain' is called. A filter is attached to a queueing discipline, the chain
starts at the top with the outer-most qdisc, and continues until a
classifier matched and returned an answer.

To reiterate the tree, which is not a tree:
<verb>
                    root 1:
                      |
                    _1:1_
                   /  |  \
                  /   |   \
                 /    |    \
               10:   11:   12:
              /   \       /   \
           10:1  10:2   12:1  12:2
</verb>

When enqueueing a packet, at each branch the filter chain is consulted for a
relevant instruction. A typical setup might be to have a filter in 1:1 that
directs a packet to 12: and a filter on 12: that sends the packet to 12:2.

You might also attach this latter rule to 1:1, but you can make efficiency
gains by having more specific tests lower in the chain.

You can't filter a packet 'upwards', by the way. 

And again - packets are only enqueued downwards! When they are dequeued,
they go up again, where the interface lives. They do NOT fall off the end of
the tree to the network adaptor!

<sect2>Some simple filtering examples
<p>
As explained in the Classifier chapter, you can match on literally anything,
using a very complicated syntax. To start, we will show how to do the
obvious things, which luckily are quite easy.

Let's say we have a PRIO qdisc called '10:' which contains three classes, and
we want to assign all traffic from and to port 22 to the highest priority
band, the filters would be:

<tscreen><verb>
# tc filter add dev eth0 protocol ip parent 10: prio 1 u32 match \ 
  ip dport 22 0xffff flowid 10:1
# tc filter add dev eth0 protocol ip parent 10: prio 1 u32 match \
  ip sport 80 0xffff flowid 10:1
# tc filter add dev eth0 protocol ip parent 10: prio 2 flowid 10:2
</verb></tscreen>

What does this say? It says: attach to eth0, node 10: a  priority 1 u32
filter that matches on IP destination port 22 *exactly* and send it to band
10:1. And it then repeats the same for source port 80. The last command says
that anything unmatched so far should go to band 10:2, the next-highest
priority.

You need to add 'eth0', or whatever your interface is called, because each
interface has a unique namespace of handles.

To select on an IP address, use this:
<tscreen><verb>
# tc filter add dev eth0 parent 10:0 protocol ip prio 1 u32 \ 
  match ip dst 4.3.2.1/32 flowid 10:1
# tc filter add dev eth0 parent 10:0 protocol ip prio 1 u32 \
  match ip src 1.2.3.4/32 flowid 10:1
# tc filter add dev eth0 protocol ip parent 10: prio 2      \
  flowid 10:2
</verb></tscreen>

This assigns traffic to 4.3.2.1 and traffic from 1.2.3.4 to the highest
priority queue, and the rest to the next-highest one.

To can concatenate matches, to match on traffic from 1.2.3.4 and from port
80, do this:
<tscreen><verb>
# tc filter add dev eth0 parent 10:0 protocol ip prio 1 u32 match ip src 4.3.2.1/32
  match ip sport 80 flowid 10:1
</verb></tscreen>

For far more information, see the 'Using filters to classify packets to
qdiscs' chapter.
<sect1>Commandlines to configure qdiscs, classes and filters
<p>
By now the concepts should be pretty clear. If not, reread the previous
sections until you are at least nearly there. Perhaps the following
commandlines will complete your understanding then.

<sect2>CBQ
<p>
This configuration limits webserver traffic to 5mbit and smtp traffic to 3
mbit, and limits the sum to 5mbit:
<verb>
# tc qdisc add dev eth0 root handle 1:0 cbq bandwidth 100Mbit         \ 
  avpkt 1000 cell 8
# tc class add dev eth0 parent 1:0 classid 1:1 cbq bandwidth 100Mbit  \
  rate 5Mbit weight 0.5Mbit prio 8 allot 1514 cell 8 maxburst 20      \
  avpkt 1000
</verb>
This part installs the root and the customary 1:0 sub-root. While it is
possible to remove the first line, and attach the second line directly to
the root, there are some subtleties involved which are avoided if adding
this extra layer.

As said before, CBQ requires a *lot* of knobs. All parameters are explained
above, however. The corresponding HTB configuration is lots simpler.

<verb>
# tc class add dev eth0 parent 1:1 classid 10: cbq bandwidth 100Mbit  \
  rate 5Mbit weight 0.5Mbit prio 5 allot 1514 cell 8 maxburst 20      \
  avpkt 1000 bounded                                                    
# tc class add dev eth0 parent 1:1 classid 20: cbq bandwidth 100Mbit  \
  rate 3Mbit weight 0.3Mbit prio 5 allot 1514 cell 8 maxburst 20      \
  avpkt 1000 bounded
</verb>

These are our two classes. Note how we scale the weight with the configured
rate. Also note that both classes are bounded and won't therefore try to
borrow traffic.

<verb>
# tc qdisc add dev eth0 parent 10: tbf rate 5Mbit buffer 10Kb/8 limit \
  15Kb mtu 1500
# tc qdisc add dev eth0 parent 20: tbf rate 3Mbit buffer 10Kb/8 limit \
  15Kb mtu 1500
</verb>

Here we install token bucket filters in the two configured subclasses. The
/8 corresponds to the cell size we mentioned earlier for CBQ. We create a
bucket of 10kbytes of tokens, a maximum 'pre-bucket' backlog of 15kbyte.

<verb>
# tc filter add dev eth0 parent 1:0 protocol ip prio 1 u32 match ip \
  sport 80 0xffff flowid 10:0
# tc filter add dev eth0 parent 1:0 protocol ip prio 1 u32 match ip \
  sport 25 0xffff flowid 20:0
</verb>

These commands, attached directly to the root, send traffic to the right
qdiscs.

Note that we use 'tc class add' to CREATE classes within a qdisc, but that
we use 'tc qdisc add' to actually configure these classes.

You may wonder what happens to traffic that is not classified by any of the
two rules. It appears that in this case, data will then be processed within
1:0, and be unlimited. This can be configured in a variety of ways, which I
do not understand. Use HTB :-)

If smtp+web together try to exceed the set limit of 5mbit/s, bandwidth will
be divided according to the weight parameter, giving 5/8 of traffic to  the
webserver and 3/8 to the mailserver.
<sect2>HTB
<p>
The functionally almost identical, but *better*, HTB configuraton is:

<tscreen><verb>
# tc qdisc add dev eth0 root handle 1: htb default 30:

# tc class add dev eth0 parent 1: classid 1:1 htb rate 5mbit burst 15k

# tc class add dev eth0 parent 1:1 classid 10: htb rate 5mbit burst 15k
# tc class add dev eth0 parent 1:1 classid 20: htb rate 3mbit ceil 5mbit burst 15k
# tc class add dev eth0 parent 1:1 classid 30: htb rate 1kbit ceil 5mbit burst 15k
</verb></tscreen>

The author then recommends SFQ for beneath these classes:
<tscreen><verb>
# tc qdisc add dev eth0 parent 10: sfq perturb 10
# tc qdisc add dev eth0 parent 20: sfq perturb 10
# tc qdisc add dev eth0 parent 30: sfq perturb 10
</verb></tscreen> 

Add the filters which direct traffic to the right classes:
<tscreen><verb>
# U32="tc filter add dev eth0 protocol ip parent 1:0 prio 1 u32"
# $U32 match ip dport 80 0xffff flowid 10:0
# $U32 match ip sport 25 0xffff flowid 20:0
</verb></tscreen>
And that's it - no unsightly unexplained numbers, no undocumented
parameters. 

HTB certainly looks wonderful - if 10: and 20: both have their guaranteed
bandwidth, and more is left to divide, they borrow in a 5:3 ratio, just as
you would expect.

Unclassified traffic gets routed to 30:, which has little bandwidth of its
own but can borrow everything that is left over. Because we chose SFQ
internally, we get fairness thrown in for free!

<sect>Loadsharing over multiple interfaces
<p>
There are several ways of doing this. One of the easiest and straightforward
ways is 'TEQL' - "True" (or "trivial") link equalizer. Like most things
having to do with queueing, loadsharing goes both ways. Both ends of a link
may need to participate for full effect.

Imagine this situation:

<tscreen><verb>
                 +-------+   eth1   +-------+
                 |       |==========|       |
 'network 1' ----|   A   |          |   B   |---- 'network 2'
                 |       |==========|       |
                 +-------+   eth2   +-------+
</verb></tscreen>

A and B are routers, and for the moment we'll assume both run Linux. If
traffic is going from network 1 to network 2, router A needs to distribute
the packets over both links to B. Router B needs to be configured to accept
this. Same goes the other way around, when packets go from network 2 to
network 1, router B needs to send the packets over both eth1 and eth2.

The distributing part is done by a 'TEQL' device, like this (it couldn't be
easier):

<tscreen><verb>
# tc qdisc add dev eth1 root teql0
# tc qdisc add dev eth2 root teql0
</verb></tscreen>

This needs to be done on both hosts. The device teql0 is basically a
roundrobbin distributor over eth1 and eth2, for sending packets. No data
ever comes in over an teql device, that just appears on the 'raw' eth1 and
eth2.

But now we just have devices, we also need proper routing. One way to do
this is to assign a /31 network to both links, and a /31 to the teql0 device
as well:

FIXME: does this need something like 'nobroadcast'? A /31 is too small to
house a network address and a broadcast address - if this doesn't work as
planned, try a /30, and adjust the ip adresses accordingly. You might even
try to make eth1 and eth2 do without an IP address!

On router A:
<tscreen><verb>
# ip addr add dev eth1 10.0.0.0/31
# ip addr add dev eth2 10.0.0.2/31
# ip addr add dev teql0 10.0.0.4/31
</verb></tscreen>

On router B:
<tscreen><verb>
# ip addr add dev eth1 10.0.0.1/31
# ip addr add dev eth2 10.0.0.3/31
# ip addr add dev teql0 10.0.0.5/31
</verb></tscreen>

Router A should now be able to ping 10.0.0.1, 10.0.0.3 and 10.0.0.5 over the
2 real links and the 1 equalized device. Router B should be able to ping
10.0.0.0, 10.0.0.2 and 10.0.0.4 over the links.

If this works, Router A should make 10.0.0.5 its route for reaching network
2, and Router B should make 10.0.0.4 its route for reaching network 1. For
the special case where network 1 is your network at home, and network 2 is
the internet, Router A should make 10.0.0.5 its default gateway.

<sect1>Caveats
<p>
Nothing is as easy as it seems. eth1 and eth2 on both router A and B need to
have return path filtering turned off, because they will otherwise drop
packets destined for ip addresses other than their own:

<tscreen><verb>
# echo 0 > /proc/net/ipv4/conf/eth1/rp_filter
# echo 0 > /proc/net/ipv4/conf/eth2/rp_filter
</verb></tscreen>

Then there is the nasty problem of packet reordering. Let's say 6 packets
need to be sent from A to B - eth1 might get 1, 3 and 5. eth2 would then do
2, 4 and 6. In an ideal world, router B would receive this in order, 1, 2,
3, 4, 5, 6. But the possibility is very real that the kernel gets it like
this: 2, 1, 4, 3, 6, 5. The problem is that this confuses TCP/IP. While not
a problem for links carrying many different TCP/IP sessions, you won't be
able to to a bundle multiple links and get to ftp a single file lots faster,
except when your receiving or sending OS is Linux, which is not easily
shaken by some simple reordering.

However, for lots of applications, link loadbalancing is a great idea.

<sect>Special Queues
<sect1>DSMARK
<p>
Esteve Camps Chust &lt;marvin@grn.es&gt;&nl;
This text is an extract from my thesis on "QoS Support in Linux", September 2000.&nl;
<P>
Source documents:&nl;
<itemize>
<item><url url="http://ica1www.epfl.ch/~almesber" name="Draft-almesberger-wajhak-diffserv-linux-01.txt">.
<item>Examples in iproute2 distribution.
<item><url url="http://www.qosforum.com/white-papers/qosprot_v3.pdf" name="White Paper-QoS protocols and architectures"> and
<url url="http://www.qosforum.com/docs/faq" name="IP QoS Frequently Asked Questions"> both by <em>Quality of Service Forum</em>.
</itemize>
<p>
This chapter was written by Esteve Camps &lt;esteve@hades.udg.es&gt;.
<sect2>Introduction
<p>

First of all, first of all, it would be a great idea for you to read RFCs
written about this (RFC2474, RFC2475, RFC2597 and RFC2598) at <url
url="http://www.ietf.org/html.carters/diffserv-charter.html" name="IETF
DiffServ working Group web site"> and <url
url="http://ica1www.epfl.ch/~almesber" name="Werner Almesberger web site">
(he wrote the code to support Differentiated Services on Linux).

<sect2>What is Dsmark related to?
<p>
Dsmark is a queueing discipline that offers the capabilities needed in
Differentiated Services (also called DiffServ or, simply, DS). DiffServ is
one of two actual QoS architectures (the other one is called Integrated
Services) that is based on a value carried by packets in the DS field of the
IP header.

<p>
One of the first solutions in IP designed to offer some QoS level was
the Type of Service field (TOS byte) in IP header. By changing that value,
we could choose a high/low level of throughput, delay or reliability.
But this didn't provide sufficient flexibility to the needs of new
services (such as real-time applications, interactive applications and
others). After this, new architectures appeared. One of these was DiffServ
which kept TOS bits and renamed DS field.
<sect2>Differentiated Services guidelines
<p>
Differentiated Services is group-oriented. I mean, we don't know nothing
about flows (this will be the Integrated Services purpose); we know about
flow aggregations and we will apply different behaviours depending on which
aggregation a packet belongs to.
<p>

When a packet arrives to an edge node (entry node to a DiffServ domain)
entering to a DiffServ Domain we'll have to policy, shape and/or mark those
packets (marking refers to assigning a value to the DS field. It's just like the
cows :-) ). This will be the mark/value that the internal/core nodes on our
DiffServ Domain will look at to determine which behaviour or QoS level
apply.

<p>
As you can deduce, Differentiated Services involves a domain on which
all DS rules will have to be applied. In fact you can think &dquot;I
will classify all the packets entering my domain. Once they enter my
domain they will be subjected to the rules that my classification dictates
and every traversed node will apply that QoS level&dquot;.
<p>
In fact, you can apply your own policies into your local domains, but some
<em>Service Level Agreements</em> should be considered when connecting to
other DS domains.
<p>

At this point, you maybe have a lot of questions. DiffServ is more than I've
explained. In fact, you can understand that I can not resume more than 3
RFC's in just 50 lines :-).

<sect2>Working with Dsmark
<p>

As the DiffServ bibliography specifies, we differentiate boundary nodes and
interior nodes. These are two important points in the traffic path. Both
types perform a classification when the packets arrive. Its result may be
used in different places along the DS process before the packet is released
to the network. It's just because of this that the diffserv code supplies an
structure called sk_buff, including a new field called skb-&gt;tc_index
where we'll store the result of initial classification that may be used in
several points in DS treatment.

<p>
The skb-&gt;tc_index value will be initially set by the DSMARK qdisc,
retrieving it from the DS field in IP header of every received packet.
Besides, cls_tcindex classifier will read all or part of skb-&gt;tcindex
value and use it to select classes.

<p>
But, first of all, take a look at DSMARK qdisc command and its parameters:
<tscreen><verb>
... dsmark indices INDICES [ default_index DEFAULT_INDEX ] [ set_tc_index ]
</verb></tscreen>
What do these parameters mean?
<itemize>
<item><bf>indices</bf>: size of table of (mask,value) pairs. Maximum value is 2^n, where n&gt=0.
<item><bf>Default_index</bf>: the default table entry index if classifier finds no match.
<item><bf>Set_tc_index</bf>: instructs dsmark discipline to retrieve the DS field and store it onto skb-&gt;tc_index.
</itemize>
Let's see the DSMARK process.

<sect2>How SCH_DSMARK works.
<p>
This qdisc will apply the next steps:
<itemize>
<item>If we have declared set_tc_index option in qdisc command, DS field is retrieved and stored onto
skb-&gt;tc_index variable.
<item>Classifier is invoked. The classifier will be executed and it will return a class ID that will be stored in
skb-&gt;tc_index variable.If no filter matches are found, we consider the default_index option to be the
classId to store. If neither set_tc_index nor default_index has been declared results may be
impredictibles.
<item>After been sent to internal qdisc's where you can reuse the result of the filter, the classid returned by
the internal qdisc is stored into skb-&gt;tc_index. We will use this value in the future to index a mask-
value table. The final result to assign to the packet will be that resulting from next operation:
<tscreen><verb>
New_Ds_field = ( Old_DS_field & mask ) | value
</verb></tscreen>

<item>Thus, new value will result from "anding" ds_field and mask values and next, this result "ORed" with
value parameter. See next diagram to understand all this process:
</itemize>
<tscreen>
<verb>
                         skb-&gt;ihp-&gt;tos
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - &gt;
     |                                                       |     ^
     | -- If you declare set_tc_index, we set DS             |     |  &lt;-----May change
     |    value into skb-&gt;tc_index variable                  |     |O       DS field
     |                                                      A|     |R
   +-|-+      +------+    +---+-+    Internal   +-+     +---N|-----|----+
   | | |      | tc   |---&gt;|   | |--&gt;  . . .  --&gt;| |     |   D|     |    |
   | | |-----&gt;|index |---&gt;|   | |     Qdisc     | |----&gt;|    v     |    |
   | | |      |filter|---&gt;| | | +---------------+ |   ----&gt;(mask,value) |
--&gt;| O |      +------+    +-|-+--------------^----+  /  |  (.  ,  .)    |
   | | |          ^         |                |       |  |  (.  ,  .)    |
   | | +----------|---------|----------------|-------|--+  (.  ,  .)    |
   | | sch_dsmark |         |                |       |                  |
   +-|------------|---------|----------------|-------|------------------+
     |            |         | &lt;- tc_index -&gt; |       |
     |            |(read)   |    may change  |       |  &lt;--------------Index to the
     |            |         |                |       |                    (mask,value)
     v            |         v                v       |                    pairs table
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -&gt;
                         skb-&gt;tc_index
</verb>
</tscreen>
<P>
How to do marking? Just change the mask and value of the class you want to remark. See next line of code:
<tscreen>
tc class change dev eth0 classid 1:1 dsmark mask 0x3 value 0xb8
</tscreen>
This changes the (mask,value) pair in hash table, to remark packets belonging to class 1:1.You have to "change" this values
because of default values that (mask,value) gets initially (see table below).
<p>
Now, we'll explain how TC_INDEX filter works and how fits into this. Besides, TCINDEX filter can be
used in other configurations rather than those including DS services.


<sect2>TC_INDEX Filter
<p>
This is the basic command to declare a TC_INDEX filter:
<tscreen>
<verb>
... tcindex [ hash SIZE ] [ mask MASK ] [ shift SHIFT ]
            [ pass_on | fall_through ]
            [ classid CLASSID ] [ police POLICE_SPEC ]
</verb>
</tscreen>
Next, we show the example used to explain TC_INDEX operation mode. Pay attention to bolded words:
&nl;&nl;
tc qdisc add dev eth0 handle 1:0 root dsmark indices 64 <bf>set_tc_index</bf>&nl;
tc filter add dev eth0 parent 1:0 protocol ip prio 1 tcindex <bf>mask 0xfc  shift 2</bf>&nl;
tc qdisc add dev eth0 parent 1:0 handle 2:0 cbq bandwidth 10Mbit cell 8 avpkt 1000 mpu 64&nl;
# EF traffic class&nl;
tc class add dev eth0 parent 2:0 classid 2:1 cbq bandwidth 10Mbit rate 1500Kbit avpkt 1000 prio 1 bounded isolated allot 1514 weight 1 maxburst 10&nl;
# Packet fifo qdisc for EF traffic&nl;
tc qdisc add dev eth0 parent 2:1 pfifo limit 5&nl;
tc filter add dev eth0 parent 2:0 protocol ip prio 1 <bf>handle 0x2e</bf> tcindex <bf>classid 2:1 pass_on</bf>&nl;
&nl;&nl;
(This code is not complete. It's just an extract from EFCBQ example included in iproute2 distribution).
<p>
First of all, suppose we receive a packet marked as EF .  If you read RFC2598, you'll see that DSCP
recommended value for EF traffic is 101110. This means that DS field will be 10111000 (remember that
less signifiant bits in TOS byte are not used in DS) or 0xb8 in hexadecimal codification.
<p>
<tscreen>
<verb>
              TC INDEX
              FILTER
   +---+      +-------+    +---+-+    +------+                +-+    +-------+
   |   |      |       |    |   | |    |FILTER|  +-+    +-+    | |    |       |
   |   |-----&gt;| MASK  | -&gt; |   | | -&gt; |HANDLE|-&gt;| |    | | -&gt; | | -&gt; |       |
   |   |  .   | =0xfc |    |   | |    |0x2E  |  | +----+ |    | |    |       |
   |   |  .   |       |    |   | |    +------+  +--------+    | |    |       |
   |   |  .   |       |    |   | |                            | |    |       |
--&gt;|   |  .   | SHIFT |    |   | |                            | |    |       |--&gt;
   |   |  .   | =2    |    |   | +----------------------------+ |    |       |
   |   |      |       |    |   |       CBQ 2:0                  |    |       |
   |   |      +-------+    +---+--------------------------------+    |       |
   |   |                                                             |       |
   |   +-------------------------------------------------------------+       |
   |                          DSMARK 1:0                                     |
   +-------------------------------------------------------------------------+

</verb>
</tscreen>
<p>
The packet arrives, then, set with 0xb8 value at DS field. As we explained before, dsmark qdisc identified
by 1:0 id in the example, retrieves DS field and store it in skb-&gt;tc_index variable.
Next step in the example will correspond to the filter associated to this qdisc (second line in the example).
This will perform next operations:
<tscreen>
<verb>
Value1 = skb->tc_index & MASK
Key = Value1 >> SHIFT
</verb>
</tscreen>

<p>
In the example, MASK=0xFC i SHIFT=2.
<tscreen>
<verb>
Value1 = 10111000 & 11111100 = 10111000
Key = 10111000 >> 2 = 00101110 -> 0x2E in hexadecimal
</verb>
</tscreen>

<p>
The returned value will correspond to a qdisc interal filter handle (in the example, identifier 2:0). If a
filter with this id exists, policing and metering conditions will be verified (in case that filter includes this)
and the classid will be returned (in our example, classid 2:1) and stored in skb-&gt;tc_index variable.

<P>
But if any filter with that identifier is found, the result will depend on fall_through flag declaration. If so,
value key is returned as classid. If not, an error is returned and process continues with the rest filters. Be
careful if you use fall_through flag; this can be done if a simple relation exists between values
&nl;of skb-&gt;tc_index variable and class id's.

<P>
The latest parameters to comment on are hash and pass_on. The first one
relates to hash table size. Pass_on will be used to indicate that if no classid
equal to the result of this filter is found, try next filter.
The default action is fall_through (look at next table).

<P>
Finally, let's see which possible values can be set to all this TCINDEX parameters:
<tscreen>
<verb>
TC Name                 Value           Default
-----------------------------------------------------------------
Hash                    1...0x10000     Implementation dependent
Mask                    0...0xffff      0xffff
Shift                   0...15          0
Fall through / Pass_on  Flag            Fall_through
Classid                 Major:minor     None
Police                  .....           None
</verb>
</tscreen>
<P>
This kind of filter is very powerful. It's necessary to explore all possibilities. Besides, this filter is not only used in DiffServ configurations.
You can use it as any other kind of filter.
<p>
I recommend you to look at all DiffServ examples included in iproute2 distribution. I promise I will try to
complement this text as soon as I can. Besides, all I have explained is the result of a lot of tests.
I would thank you tell me if I'm wrong in any point.
<sect1>WRR
<p>
This qdisc is not included in the standard kernels but can be downloaded from 
<url url="http://wipl-wrr.dkik.dk/wrr/">. 
Currently the qdisc is only tested with Linux 2.2 kernels but it will 
probably work with 2.4 kernels too.

The WRR qdisc distributes bandwidth between its classes using the weighted 
round robin scheme. That is, like the CBQ qdisc it contains classes 
into which arbitrary qdiscs can be plugged. All classes which have sufficient 
demand will get bandwidth proportional to the weights associated with the classes.
The weights can be set manually using the <tt>tc</tt> program. But they
can also be made automatically decreasing for classes transferring much data.

The qdisc has a built-in classifier which assigns packets coming from or 
sent to different machines to different classes. Either the MAC or IP and 
either source or destination addresses can be used. The MAC address can only 
be used when the Linux box is acting as an ethernet bridge, however. The 
classes are automatically assigned to machines based on the packets seen.
 
The qdisc can be very useful at sites such as dorms where a lot of unrelated 
individuals share an Internet connection. A set of scripts setting up a 
relevant behavior for such a site is a central part of the WRR distribution.

<sect>Netfilter &amp; iproute - marking packets
<p>
So far we've seen how iproute works, and netfilter was mentioned a few
times. This would be a good time to browse through <url name="Rusty's Remarkably
Unreliable Guides"
url="http://netfilter.samba.org/unreliable-guides/">. Netfilter itself
can be found <url name="here"
url="http://netfilter.filewatcher.org/">.

Netfilter allows us to filter packets, or mangle their headers. One special
feature is that we can mark a packet with a number. This is done with the
--set-mark facility. 

As an example, this command marks all packets destined for port 25, outgoing
mail:

<tscreen><verb>
# iptables -A PREROUTING -i eth0 -t mangle -p tcp --dport 25 \
 -j MARK --set-mark 1
</verb></tscreen>

Let's say that we have multiple connections, one that is fast (and
expensive, per megabyte) and one that is slower, but flat fee. We would most
certainly like outgoing mail to go via the cheap route.

We've already marked the packets with a '1', we now instruct the routing
policy database to act on this:

<tscreen><verb>
# echo 201 mail.out >> /etc/iproute2/rt_tables
# ip rule add fwmark 1 table mail.out
# ip rule ls
0:      from all lookup local 
32764:  from all fwmark        1 lookup mail.out 
32766:  from all lookup main 
32767:  from all lookup default 
</verb></tscreen>

Now we generate the mail.out table with a route to the slow but cheap link:
<tscreen><verb>
# /sbin/ip route add default via 195.96.98.253 dev ppp0 table mail.out
</verb></tscreen>

And we are done. Should we want to make exceptions, there are lots of ways
to achieve this. We can modify the netfilter statement to exclude certain
hosts, or we can insert a rule with a lower priority that points to the main
table for our excepted hosts.

We can also use this feature to honour TOS bits by marking packets with a
different type of service with different numbers, and creating rules to act
on that. This way you can even dedicate, say, an ISDN line to interactive
sessions.

Needless to say, this also works fine on a host that's doing NAT
('masquerading').

IMPORTANT: We received a report that MASQ and SNAT at least collide
with marking packets. Rusty Russell explains it in
<url
url="http://lists.samba.org/pipermail/netfilter/2000-November/006089.html" 
name="this posting">. Turn off the reverse path filter to make it work
properly.

Note: to mark packets, you need to have some options enabled in your
kernel:

<tscreen><verb>
IP: advanced router (CONFIG_IP_ADVANCED_ROUTER) [Y/n/?]
 IP: policy routing (CONFIG_IP_MULTIPLE_TABLES) [Y/n/?]
  IP: use netfilter MARK value as routing key (CONFIG_IP_ROUTE_FWMARK) [Y/n/?]
</verb></tscreen>

See also <ref id="SQUID" name="Transparent web-caching using netfilter, iproute2, ipchains and squid">
in the Cookbook.
<sect>Using filters to classify packets to qdiscs
<p>
As explained in the section on classful queueing disciplines, filters are
needed to classify packets into any of the sub-queues. These filters are
called from within the classful qdisc.

Here is an incomplete list of classifiers available:
<descrip>
  <tag>fw</tag>
       Bases the decision on how the firewall has marked the packet.
       
  <tag>u32</tag>
       Bases the decision on fields within the packet (i.e. source IP address, etc)
       
  <tag>route</tag>
       Bases the decision on which route the packet will be routed by.
       
  <tag>rsvp, rsvp6</tag>
       Bases the decision on the target (destination,protocol) and optionally the source as well. (I think)
       
  <tag>tcindex</tag>
       FIXME: Fill me in
</descrip>

Note that in general there are many ways in which you can classify packet
and that it generally comes down to preference as to which system you wish
to use.

Classifiers in general accept a few arguments in common. They are listed here for convenience:

<descrip>
  <tag>protocol</tag>
      The protocol this classifier will accept. Generally you will only be
      accepting only IP traffic. Required.

  <tag>parent</tag>
      The handle this classifier is to be attached to. This handle must be
      an already existing class. Required.

  <tag>prio</tag>
      The priority of this classifier. Lower numbers get tested first.

  <tag>handle</tag>
      This handle means different things to different filters.

      FIXME: Add more
</descrip>

All the following sections will assume you are trying to shape the traffic
going to <tt>HostA</tt>. They will assume that the root class has been
configured on 1: and that the class you want to send the selected traffic to
is 1:1.

<sect1>The "fw" classifier
<p>  The "fw" classifier relies on the firewall tagging the packets to be shaped. So,
  first we will setup the firewall to tag them:

<tscreen><verb>
# iptables -I PREROUTING -t mangle -p tcp -d HostA \
 -j MARK --set-mark 1
</verb></tscreen>

  Now all packets to that machine are tagged with the mark 1. Now we build
  the packet shaping rules to actually shape the packets.  Now we just need
  to indicate that we want the packets that are tagged with the mark 1 to go
  to class 1:1. This is accomplished with the command:

<tscreen><verb>
# tc filter add dev eth1 protocol ip parent 1:0 prio 1 handle 1 fw classid 1:1
</verb></tscreen>

  This should be fairly self-explanatory. Attach to the 1:0 class a filter
  with priority 1 to filter all packet marked with 1 in the firewall to
  class 1:1. Note how the handle here is used to indicate what the mark
  should be.

  That's all there is to it! This is the (IMHO) easy way, the other ways are
  harder to understand, I think. Note that you can apply the full power of
  the firewalling code with this classifier, including matching MAC
  addresses, user IDs and anything else the firewall can match.

<sect1>The "u32" classifier
<p>
The U32 filter is the most advanced filter available in the current
implementation. It entirely based on hashing tables, which make it
robust when there are many filter rules.

In its simplest form the U32 filter is a list of records, each
consisting of two fields: a selector and an action. The selectors,
described below, are compared with the currently processed IP packet
until the first match occurs, and then the associated action is performed.
The simplest type of action would be directing the packet into defined
CBQ class.

The commandline of <tt>tc filter</tt> program, used to configure the filter,
consists of three parts: filter specification, a selector and an action.
The filter specification can be defined as:

<tscreen><verb>
tc filter add dev IF [ protocol PROTO ]
                     [ (preference|priority) PRIO ]
                     [ parent CBQ ]
</verb></tscreen>

The <tt>protocol</tt> field describes protocol that the filter will be
applied to. We will only discuss case of <tt>ip</tt> protocol. The
<tt>preference</tt> field (<tt>priority</tt> can be used alternatively)
sets the priority of currently defined filter. This is important, since
you can have several filters (lists of rules) with different priorities.
Each list will be passed in the order the rules were added, then list with
lower priority (higher preference number) will be processed. The <tt>parent</tt>
field defines the CBQ tree top (e.g. 1:0), the filter should be attached
to.

The options decribed apply to all filters, not only U32.

<sect2>U32 selector
<p>
The U32 selector contains definition of the pattern, that will be matched
to the currently processed packet. Precisely, it defines which bits are
to be matched in the packet header and nothing more, but this simple
method is very powerful. Let's take a look at the following examples,
taken directly from a pretty complex, real-world filter:

<tscreen><verb>
# filter parent 1: protocol ip pref 10 u32 fh 800::800 order 2048 key ht 800 bkt 0 flowid 1:3 \
  match 00100000/00ff0000 at 0
</verb></tscreen>

<p>
For now, leave the first line alone - all these parameters describe
the filter's hash tables. Focus on the selector line, containing
<tt>match</tt> keyword. This selector will match to IP headers, whose
second byte will be 0x10 (0010). As you can guess, the 00ff number is
the match mask, telling the filter exactly which bits to match. Here
it's 0xff, so the byte will match if it's exactly 0x10. The <tt>at</tt>
keyword means that the match is to be started at specified offset (in
bytes) -- in this case it's beginning of the packet.  Translating all
that to human language, the packet will match if its Type of Service
field will have `low delay' bits set. Let's analyze another rule:

<tscreen><verb>
# filter parent 1: protocol ip pref 10 u32 fh 800::803 order 2051 key ht 800 bkt 0 flowid 1:3 \
  match 00000016/0000ffff at nexthdr+0
</verb></tscreen>

<p>
The <tt>nexthdr</tt> option means next header encapsulated in the IP packet,
i.e. header of upper-layer protocol. The match will also start here
at the beginning of the next header. The match should occur in the
second, 32-bit word of the header. In TCP and UDP protocols this field
contains packet's destination port. The number is given in big-endian
format, i.e. older bits first, so we simply read 0x0016 as 22 decimal,
which stands for SSH service if this was TCP. As you guess, this match
is ambigous without a context, and we will discuss this later.

<p>
Having understood all the above, we will find the following selector
quite easy to read: <tt>match c0a80100/ffffff00 at 16</tt>. What we
got here is a three byte match at 17-th byte, counting from the IP
header start. This will match for packets with destination address
anywhere in 192.168.1/24 network. After analyzing the examples, we
can summarize what we have learnt.

<sect2>General selectors

<p>
General selectors define the pattern, mask and offset the pattern
will be matched to the packet contents. Using the general selectors
you can match virtually any single bit in the IP (or upper layer)
header. They are more difficult to write and read, though, than
specific selectors that described below. The general selector syntax
is:

<tscreen><verb>
match [ u32 | u16 | u8 ] PATTERN MASK [ at OFFSET | nexthdr+OFFSET]
</verb></tscreen>

<p>
One of the keywords <tt>u32</tt>, <tt>u16</tt> or <tt>u8</tt> specifies
length of the pattern in bits. PATTERN and MASK should follow, of length
defined by the previous keyword. The OFFSET parameter is the offset,
in bytes, to start matching. If <tt>nexthdr+</tt> keyword is given,
the offset is relative to start of the upper layer header.

<p>
Some examples:

<tscreen><verb>
# tc filter add dev ppp14 parent 1:0 prio 10 u32 \
     match u8 64 0xff at 8 \
     flowid 1:4
</verb></tscreen>

<p>
Packet will match to this rule, if its time to live (TTL) is 64.
TTL is the field starting just after 8-th byte of the IP header.

<tscreen><verb>
# tc filter add dev ppp14 parent 1:0 prio 10 u32 \
     match u8 0x10 0xff at nexthdr+13 \
     protocol tcp \
     flowid 1:3 \
</verb></tscreen>

<p>
This rule will only match TCP packets with ACK bit set. Here we can see
an example of using two selectors, the final result will be logical AND
of their results. If we take a look at TCP header diagram, we can see
that the ACK bit is second older bit (0x10) in the 14-th byte of the TCP
header (<tt>at nexthdr+13</tt>).  As for the second selector, if we'd like
to make our life harder, we could write <tt>match u8 0x06 0xff at 9</tt>
instead of using the specific selector <tt>protocol tcp</tt>, because
6 is the number of TCP protocol, present in 10-th byte of the IP header.
On the other hand, in this example we couldn't use any specific selector
for the first match - simply because there's no specific selector to match
TCP ACK bits.

<sect2>Specific selectors
<p>
The following table contains a list of all specific selectors 
the author of this section has found in the <tt>tc</tt> program
source code. They simply make your life easier and increase readability
of your filter's configuration.

FIXME: table placeholder - the table is in separate file ,,selector.html''

FIXME: it's also still in Polish :-(

FIXME: must be sgml'ized

Some examples:


<tscreen><verb>
# tc filter add dev ppp0 parent 1:0 prio 10 u32 \
     match ip tos 0x10 0xff \
     flowid 1:4
</verb></tscreen>

The above rule will match packets which have the TOS field set to 0x10.
The TOS field starts at second byte of the packet and is one byte big,
so we could write an equivalent general selector: <tt>match u8 0x10 0xff
at 1</tt>. This gives us hint to the internals of U32 filter -- the
specific rules are always translated to general ones, and in this
form they are stored in the kernel memory. This leads to another conclusion
-- the <tt>tcp</tt> and <tt>udp</tt> selectors are exactly the same
and this is why you can't use single <tt>match tcp dst 53 0xffff</tt>
selector to match TCP packets sent to given port -- they will also
match UDP packets sent to this port. You must remember to also specify
the protocol and end up with the following rule:

<tscreen><verb>
# tc filter add dev ppp0 parent 1:0 prio 10 u32 \
        match tcp dst 53 0xffff \
        match ip protocol 0x6 0xff \
        flowid 1:2
</verb></tscreen>

<!--
TODO:

describe more options

match
offset
hashkey
classid | flowid
divisor
order
link
ht
sample
police

-->

<sect1>The "route" classifier
<p>

  This classifier filters based on the results of the routing tables. When a
  packet that is traversing through the classes reaches one that is marked
  with the "route" filter, it splits the packets up based on information in
  the routing table.

<tscreen><verb>
# tc filter add dev eth1 parent 1:0 protocol ip prio 100 route
</verb></tscreen>

  Here we add a route classifier onto the parent node 1:0 with priority 100. 
  When a packet reaches this node (which, since it is the root, will happen
  immediately) it will consult the routing table and if one matches will
  send it to the given class and give it a priority of 100. Then, to finally
  kick it into action, you add the appropriate routing entry:

  The trick here is to define 'realm' based on either destination or source. 
  The way to do it is like this:

<tscreen><verb>
# ip route add Host/Network via Gateway dev Device realm RealmNumber
</verb></tscreen>
  
  For instance, we can define our destination network 192.168.10.0 with a realm
  number 10:
  
<tscreen><verb>
# ip route add 192.168.10.0/24 via 192.168.10.1 dev eth1 realm 10
</verb></tscreen>  

  When adding route filters, we can use realm numbers to represent the
  networks or hosts and specify how the routes match the filters.

<tscreen><verb>
# tc filter add dev eth1 parent 1:0 protocol ip prio 100 \
  route to 10 classid 1:10
</verb></tscreen>  

  The above rule says packets going to the network 192.168.10.0 match class id
  1:10.
  
  Route filter can also be used to match source routes. For example, there is 
  a subnetwork attached to the Linux router on eth2.

<tscreen><verb>
# ip route add 192.168.2.0/24 dev eth2 realm 2
# tc filter add dev eth1 parent 1:0 protocol ip prio 100 \
  route from 2 classid 1:2
</verb></tscreen>

  Here the filter specifies that packets from the subnetwork 192.168.2.0
  (realm 2) will match class id 1:2.

<sect1>The "rsvp" classifier
<p>FIXME: Fill me in


<sect1>The "tcindex" classifier
<p>FIXME: Fill me in

<sect>Kernel network parameters
<p>
The kernel has lots of parameters which can be tuned for different
circumstances. While, as usual, the default parameters serve 99% of
installations very well, we don't call this the Advanced HOWTO for the fun
of it!

The interesting bits are in /proc/sys/net, take a look there. Not everything
will be documented here initially, but we're working on it.

(FIXME)

<sect1>Reverse Path Filtering
<p>
By default, routers route everything, even packets which 'obviously' don't
belong on your network. A common example is private IP space escaping onto
the internet. If you have an interface with a route of 195.96.96.0/24 to it,
you do not expect packets from 212.64.94.1 to arrive there.

Lots of people will want to turn this feature off, so the kernel hackers
have made it easy. There are files in <file>/proc</file> where you can tell
the kernel to do this for you. The method is called "Reverse Path
Filtering". Basically, if the reply to this packet wouldn't go out the
interface this packet came in, then this is a bogus packet and should be
ignored.

The following fragment will turn this on for all current and future
interfaces.

<tscreen><verb>
# for i in /proc/sys/net/ipv4/conf/*/rp_filter ; do
&gt;  echo 2 > $i 
&gt; done
</verb></tscreen>

Going by the example above, if a packet arrived on the Linux router on eth1
claiming to come from the Office+ISP subnet, it would be dropped. Similarly,
if a packet came from the Office subnet, claiming to be from somewhere
outside your firewall, it would be dropped also.

The above is full reverse path filtering. The default is to only filter
based on IPs that are on directly connected networks. This is because the
full filtering breaks in the case of asymmetric routing (where packets come
in one way and go out another, like satellite traffic, or if you have
dynamic (bgp, ospf, rip) routes in your network. The data comes down
through the satellite dish and replies go back through normal land-lines).

If this exception applies to you (and you'll probably know if it does) you
can simply turn off the <file>rp_filter</file> on the interface where the
satellite data comes in. If you want to see if any packets are being
dropped, the <file>log_martians</file> file in the same directory will tell
the kernel to log them to your syslog.

<tscreen><verb>
# echo 1 >/proc/sys/net/ipv4/conf/<interfacename>/log_martians
</verb></tscreen>

FIXME: is setting the conf/{default,all}/* files enough? - martijn

<sect1>Obscure settings
<p>
Ok, there are a lot of parameters which can be modified. We try to list them
all. Also documented (partly) in <file>Documentation/ip-sysctl.txt</file>.

Some of these settings have different defaults based on whether you 
answered 'Yes' to 'Configure as router and not host' while compiling your
kernel.

<sect2>Generic ipv4
<p>
As a generic note, most rate limiting features don't work on loopback, so
don't test them locally. The limits are supplied in 'jiffies', and are
enforced using the earlier mentioned token bucket filter.

The kernel has an internal clock which runs at 'HZ' ticks (or 'jiffies') per
second. On intel, 'HZ' is mostly 100. So setting a *_rate file to, say 50,
would allow for 2 packets per second. The token bucket filter is also
configured to allow for a burst of at most 6 packets, if enough tokens have
been earned.

Several entries in the following list have been copied from
/usr/src/linux/Documentation/networking/ip-sysctl.txt, written by Alexey
Kuznetsov &lt;kuznet@ms2.inr.ac.ru&gt; and Andi Kleen &lt;ak@muc.de&gt;
<descrip>
<tag>/proc/sys/net/ipv4/icmp_destunreach_rate</tag>
If the kernel decides that it can't deliver a packet, it will drop it, and
send the source of the packet an ICMP notice to this effect.
<tag>/proc/sys/net/ipv4/icmp_echo_ignore_all</tag>
Don't act on echo packets at all. Please don't set this by default, but if
you are used as a relay in a DoS attack, it may be useful.
<tag>/proc/sys/net/ipv4/icmp_echo_ignore_broadcasts [Useful]</tag>
If you ping the broadcast address of a network, all hosts are supposed to
respond. This makes for a dandy denial-of-service tool. Set this to 1 to
ignore these broadcast messages.
<tag>/proc/sys/net/ipv4/icmp_echoreply_rate</tag>
The rate at which echo replies are sent to any one destination.
<tag>/proc/sys/net/ipv4/icmp_ignore_bogus_error_responses</tag>
Set this to ignore ICMP errors caused by hosts in the network reacting badly
to frames sent to what they perceive to be the broadcast address.
<tag>/proc/sys/net/ipv4/icmp_paramprob_rate</tag>
A relatively unknown ICMP message, which is sent in response to incorrect
packets with broken IP or TCP headers. With this file you can control the
rate at which it is sent.
<tag>/proc/sys/net/ipv4/icmp_timeexceed_rate</tag>
This the famous cause of the 'Solaris middle star' in traceroutes. Limits
number of ICMP Time Exceeded messages sent. 
<tag>/proc/sys/net/ipv4/igmp_max_memberships</tag>
Maximum number of listening igmp (multicast) sockets on the host.
FIXME: Is this true?
<tag>/proc/sys/net/ipv4/inet_peer_gc_maxtime</tag>
FIXME: Add a little explanation about the inet peer storage?&nl;
Minimum interval between garbage collection passes.  This interval is in
effect under low (or absent) memory pressure on the pool. Measured in
jiffies.
<tag>/proc/sys/net/ipv4/inet_peer_gc_mintime</tag>
Minimum interval between garbage collection passes.  This interval is in
effect under high memory pressure on the pool. Measured in jiffies.
<tag>/proc/sys/net/ipv4/inet_peer_maxttl</tag>
Maximum time-to-live of entries.  Unused entries will expire after this
period of time if there is no memory pressure on the pool (i.e. when the
number of entries in the pool is very small). Measured in jiffies.
<tag>/proc/sys/net/ipv4/inet_peer_minttl</tag>
Minimum time-to-live of entries.  Should be enough to cover fragment
time-to-live on the reassembling side.  This minimum time-to-live
is guaranteed if the pool size is less than inet_peer_threshold.
Measured in jiffies.
<tag>/proc/sys/net/ipv4/inet_peer_threshold</tag>
The approximate size of the INET peer storage.  Starting from this threshold
entries will be thrown aggressively.  This threshold also determines
entries' time-to-live and time intervals between garbage collection passes. 
More entries, less time-to-live, less GC interval.
<tag>/proc/sys/net/ipv4/ip_autoconfig</tag>
This file contains the number one if the host received its IP configuration by
RARP, BOOTP, DHCP or a similar mechanism. Otherwise it is zero.
<tag>/proc/sys/net/ipv4/ip_default_ttl</tag>
Time To Live of packets. Set to a safe 64. Raise it if you have a huge
network. Don't do so for fun - routing loops cause much more damage that
way. You might even consider lowering it in some circumstances.
<tag>/proc/sys/net/ipv4/ip_dynaddr</tag>
You need to set this if you use dial-on-demand with a dynamic interface
address. Once your demand interface comes up, any local TCP sockets which haven't seen replies will be rebound to have the right address. This solves the problem that the
connection that brings up your interface itself does not work, but the
second try does.
<tag>/proc/sys/net/ipv4/ip_forward</tag>
If the kernel should attempt to forward packets. Off by default.
<tag>/proc/sys/net/ipv4/ip_local_port_range</tag>
Range of local ports for outgoing connections. Actually quite small by
default, 1024 to 4999.
<tag>/proc/sys/net/ipv4/ip_no_pmtu_disc</tag>
Set this if you want to disable Path MTU discovery - a technique to
determine the largest Maximum Transfer Unit possible on your path. See also
the section on Path MTU discovery in the cookbook chapter.
<tag>/proc/sys/net/ipv4/ipfrag_high_thresh</tag>
Maximum memory used to reassemble IP fragments. When 
ipfrag_high_thresh bytes of memory is allocated for this purpose,
the fragment handler will toss packets until ipfrag_low_thresh
is reached.
<tag>/proc/sys/net/ipv4/ip_nonlocal_bind</tag>
Set this if you want your applications to be able to bind to an address
which doesn't belong to a device on your system. This can be useful when
your machine is on a non-permanent (or even dynamic) link, so your services
are able to start up and bind to a specific address when your link is down.
<tag>/proc/sys/net/ipv4/ipfrag_low_thresh</tag>
Minimum memory used to reassemble IP fragments.
<tag>/proc/sys/net/ipv4/ipfrag_time</tag>
Time in seconds to keep an IP fragment in memory.
<tag>/proc/sys/net/ipv4/tcp_abort_on_overflow</tag>
A boolean flag controlling the behaviour under lots of incoming connections.
When enabled, this causes the kernel to actively send RST packets when a
service is overloaded.
<tag>/proc/sys/net/ipv4/tcp_fin_timeout</tag>
Time to hold socket in state FIN-WAIT-2, if it was closed by our side. Peer
can be broken and never close its side, or even died unexpectedly. Default
value is 60sec. Usual value used in 2.2 was 180 seconds, you may restore it,
but remember that if your machine is even underloaded WEB server, you risk
to overflow memory with kilotons of dead sockets, FIN-WAIT-2 sockets are
less dangerous than FIN-WAIT-1, because they eat maximum 1.5K of memory, but
they tend to live longer. Cf. tcp_max_orphans.
<tag>/proc/sys/net/ipv4/tcp_keepalive_time</tag>
How often TCP sends out keepalive messages when keepalive is enabled. &nl;
Default: 2hours.
<tag>/proc/sys/net/ipv4/tcp_keepalive_intvl</tag>
How frequent probes are retransmitted, when a probe isn't acknowledged. &nl;
Default: 75 seconds.
<tag>/proc/sys/net/ipv4/tcp_keepalive_probes</tag>
How many keepalive probes TCP will send, until it decides that the
connection is broken. &nl;
Default value: 9. &nl;
Multiplied with tcp_keepalive_intvl, this gives the time a link can be
nonresponsive after a keepalive has been sent.
<tag>/proc/sys/net/ipv4/tcp_max_orphans</tag>
Maximal number of TCP sockets not attached to any user file handle, held by
system. If this number is exceeded orphaned connections are reset
immediately and warning is printed. This limit exists only to prevent simple
DoS attacks, you _must_ not rely on this or lower the limit artificially,
but rather increase it (probably, after increasing installed memory), if
network conditions require more than default value, and tune network
services to linger and kill such states more aggressively. Let me remind you
again: each orphan eats up to ~64K of unswappable memory.
<tag>/proc/sys/net/ipv4/tcp_orphan_retries</tag>
How may times to retry before killing TCP connection, closed by our side.
Default value 7 corresponds to ~50sec-16min depending on RTO. If your machine
is a loaded WEB server, you should think about lowering this value, such
sockets may consume significant resources. Cf. tcp_max_orphans.
<tag>/proc/sys/net/ipv4/tcp_max_syn_backlog</tag>
Maximal number of remembered connection requests, which still did not
receive an acknowledgement from connecting client. Default value is 1024 for
systems with more than 128Mb of memory, and 128 for low memory machines. If
server suffers of overload, try to increase this number. Warning! If you
make it greater than 1024, it would be better to change TCP_SYNQ_HSIZE in
include/net/tcp.h to keep TCP_SYNQ_HSIZE*16<=tcp_max_syn_backlog and to
recompile kernel.
<tag>/proc/sys/net/ipv4/tcp_max_tw_buckets</tag>
Maximal number of timewait sockets held by system simultaneously. If this
number is exceeded time-wait socket is immediately destroyed and warning is
printed. This limit exists only to prevent simple DoS attacks, you _must_
not lower the limit artificially, but rather increase it (probably, after
increasing installed memory), if network conditions require more than
default value.
<tag>/proc/sys/net/ipv4/tcp_retrans_collapse</tag>
Bug-to-bug compatibility with some broken printers.
On retransmit try to send bigger packets to work around bugs in
certain TCP stacks.
<tag>/proc/sys/net/ipv4/tcp_retries1</tag>
How many times to retry before deciding that something is wrong
and it is necessary to report this suspection to network layer.
Minimal RFC value is 3, it is default, which corresponds
to ~3sec-8min depending on RTO.
<tag>/proc/sys/net/ipv4/tcp_retries2</tag>
How may times to retry before killing alive TCP connection.
RFC1122 says that the limit should be longer than 100 sec.
It is too small number. Default value 15 corresponds to ~13-30min
depending on RTO.
<tag>/proc/sys/net/ipv4/tcp_rfc1337</tag>
This boolean enables a fix for 'time-wait assassination hazards in tcp', described
in RFC 1337. If enabled, this causes the kernel to drop RST packets for
sockets in the time-wait state.&nl;
Default: 0
<tag>/proc/sys/net/ipv4/tcp_sack</tag>
Use Selective ACK which can be used to signify that specific packets are
missing - therefore helping fast recovery.
<tag>/proc/sys/net/ipv4/tcp_stdurg</tag>
Use the Host requirements interpretation of the TCP urg pointer
field. &nl;
Most hosts use the older BSD interpretation, so if you turn this on
Linux might not communicate correctly with them. &nl;
Default: FALSE 
<tag>/proc/sys/net/ipv4/tcp_syn_retries</tag>
Number of SYN packets the kernel will send before giving up on the new
connection.
<tag>/proc/sys/net/ipv4/tcp_synack_retries</tag>
To open the other side of the connection, the kernel sends a SYN with a
piggybacked ACK on it, to acknowledge the earlier received SYN. This is part
2 of the threeway handshake. This setting determines the number of SYN+ACK
packets sent before the kernel gives up on the connection.
<tag>/proc/sys/net/ipv4/tcp_timestamps</tag>
Timestamps are used, amongst other things, to protect against wrapping
sequence numbers. A 1 gigabit link might conceivably re-encounter a previous
sequence number with an out-of-line value, because it was of a previous
generation. The timestamp will let it recognise this 'ancient packet'.
<tag>/proc/sys/net/ipv4/tcp_tw_recycle</tag>
Enable fast recycling TIME-WAIT sockets. Default value is 1.
It should not be changed without advice/request of technical experts.

<tag>/proc/sys/net/ipv4/tcp_window_scaling</tag>
TCP/IP normally allows windows up to 65535 bytes big. For really fast
networks, this may not be enough. The window scaling options allows for
almost gigabyte windows, which is good for high bandwidth*delay products.

</descrip>
<sect2>Per device settings
<p>
DEV can either stand for a real interface, or for 'all' or 'default'.
Default also changes settings for interfaces yet to be created.
<descrip>
<tag>/proc/sys/net/ipv4/conf/DEV/accept_redirects</tag>
If a router decides that you are using it for a wrong purpose (ie, it needs
to resend your packet on the same interface), it will send us a ICMP
Redirect. This is a slight security risk however, so you may want to turn it
off, or use secure redirects.
<tag>/proc/sys/net/ipv4/conf/DEV/accept_source_route</tag>
Not used very much anymore. You used to be able to give a packet a list of
IP addresses it should visit on its way. Linux can be made to honor this IP
option.
<tag>/proc/sys/net/ipv4/conf/DEV/bootp_relay</tag>
FIXME: fill this in
<tag>/proc/sys/net/ipv4/conf/DEV/forwarding</tag>
FIXME:
<tag>/proc/sys/net/ipv4/conf/DEV/log_martians</tag>
See the section on reverse path filters.
<tag>/proc/sys/net/ipv4/conf/DEV/mc_forwarding</tag>
If we do multicast forwarding on this interface
<tag>/proc/sys/net/ipv4/conf/DEV/proxy_arp</tag>
If you set this to 1, all other interfaces will respond to arp queries
destined for addresses on this interface. Can be very useful when building 'ip 
pseudo bridges'. Do take care that your netmasks are very correct before
enabling this!
<tag>/proc/sys/net/ipv4/conf/DEV/rp_filter</tag>
See the section on reverse path filters.
<tag>/proc/sys/net/ipv4/conf/DEV/secure_redirects</tag>
FIXME: fill this in
<tag>/proc/sys/net/ipv4/conf/DEV/send_redirects</tag>
If we send the above mentioned redirects.
<tag>/proc/sys/net/ipv4/conf/DEV/shared_media</tag>
FIXME: fill this in
<tag>/proc/sys/net/ipv4/conf/DEV/tag</tag>
FIXME: fill this in

</descrip>

<sect2> Neighbor pollicy
<p>
Dev can either stand for a real interface, or for 'all' or 'default'.
Default also changes settings for interfaces yet to be created.
<descrip>
<tag>/proc/sys/net/ipv4/neigh/DEV/anycast_delay</tag>
FIXME: fill this in
<tag>/proc/sys/net/ipv4/neigh/DEV/app_solicit</tag>
FIXME: fill this in
<tag>/proc/sys/net/ipv4/neigh/DEV/base_reachable_time</tag>
FIXME: fill this in
<tag>/proc/sys/net/ipv4/neigh/DEV/delay_first_probe_time</tag>
FIXME: fill this in
<tag>/proc/sys/net/ipv4/neigh/DEV/gc_stale_time</tag>
FIXME: fill this in
<tag>/proc/sys/net/ipv4/neigh/DEV/locktime</tag>
FIXME: fill this in
<tag>/proc/sys/net/ipv4/neigh/DEV/mcast_solicit</tag>
FIXME: fill this in
<tag>/proc/sys/net/ipv4/neigh/DEV/proxy_delay</tag>
FIXME: fill this in
<tag>/proc/sys/net/ipv4/neigh/DEV/proxy_qlen</tag>
FIXME: fill this in
<tag>/proc/sys/net/ipv4/neigh/DEV/retrans_time</tag>
FIXME: fill this in
<tag>/proc/sys/net/ipv4/neigh/DEV/ucast_solicit</tag>
FIXME: fill this in
<tag>/proc/sys/net/ipv4/neigh/DEV/unres_qlen</tag>
FIXME: fill this in

</descrip>

<sect2>Routing settings
<p>
<descrip>
<tag>/proc/sys/net/ipv4/route/error_burst</tag>
FIXME: fill this in
<tag>/proc/sys/net/ipv4/route/error_cost</tag>
FIXME: fill this in
<tag>/proc/sys/net/ipv4/route/flush</tag>
FIXME: fill this in
<tag>/proc/sys/net/ipv4/route/gc_elasticity</tag>
FIXME: fill this in
<tag>/proc/sys/net/ipv4/route/gc_interval</tag>
FIXME: fill this in
<tag>/proc/sys/net/ipv4/route/gc_min_interval</tag>
FIXME: fill this in
<tag>/proc/sys/net/ipv4/route/gc_thresh</tag>
FIXME: fill this in
<tag>/proc/sys/net/ipv4/route/gc_timeout</tag>
FIXME: fill this in
<tag>/proc/sys/net/ipv4/route/max_delay</tag>
FIXME: fill this in
<tag>/proc/sys/net/ipv4/route/max_size</tag>
FIXME: fill this in
<tag>/proc/sys/net/ipv4/route/min_adv_mss</tag>
FIXME: fill this in
<tag>/proc/sys/net/ipv4/route/min_delay</tag>
FIXME: fill this in
<tag>/proc/sys/net/ipv4/route/min_pmtu</tag>
FIXME: fill this in
<tag>/proc/sys/net/ipv4/route/mtu_expires</tag>
FIXME: fill this in
<tag>/proc/sys/net/ipv4/route/redirect_load</tag>
FIXME: fill this in
<tag>/proc/sys/net/ipv4/route/redirect_number</tag>
FIXME: fill this in
<tag>/proc/sys/net/ipv4/route/redirect_silence</tag>
FIXME: fill this in
</descrip>


<sect>Backbone applications of traffic control
<p>
This chapter is meant as an introduction to backbone routing, which often
involves >100 megabit bandwidths, which requires a different approach than
your ADSL modem at home.

<sect1>Router queues
<p>
The normal behaviour of router queues on the Internet is called tail-drop.
Tail-drop works by queueing up to a certain amount, then dropping all traffic
that 'spills over'. This is very unfair, and also leads to retransmit 
synchronisation. When retransmit synchronisation occurs, the sudden burst
of drops from a router that has reached its fill will cause a delayed burst
of retransmits, which will over fill the congested router again. 

In order to cope with transient congestion on links, backbone routers will
often implement large queues. Unfortunately, while these queues are good for
throughput, they can substantially increase latency and cause TCP
connections to behave very bursty during congestion.

These issues with tail-drop are becoming increasingly troublesome on the
Internet because the use of network unfriendly applications is increasing.
The Linux kernel offers us RED, short for Random Early Detect.

RED isn't a cure-all for this, applications which inappropriately fail to 
implement exponential backoff still get an unfair share of the bandwidth,
however, with RED they do not cause as much harm to the throughput and
latency of other connections.

RED statistically drops packets from flows before it reaches its hard
limit. This causes a congested backbone link to slow more gracefully, and
prevents retransmit synchronisation. This also helps TCP find its 'fair'
speed faster by allowing some packets to get dropped sooner keeping queue
sizes low and latency under control. The probability of a packet being
dropped from a particular connection is proportional to its bandwidth usage
rather than the number of packets it transmits. 

RED is a good queue for backbones, where you can't afford the 
complexity of per-session state tracking needed by fairness queueing.

In order to use RED, you must decide on three parameters: Min, Max, and
burst. Min sets the minimum queue size in bytes before dropping will begin,
Max is a soft maximum that the algorithm will attempt to stay under, and
burst sets the maximum number of packets that can 'burst through'.

You should set the min by calculating that highest acceptable base queueing 
latency you wish, and multiply it by your bandwidth. For instance, on my 
64kbit/s ISDN link, I might want a base queueing latency of 200ms so I set
min to 1600 bytes. Setting min too small will degrade throughput and too
large will degrade latency. Setting a small min is not a replacement for
reducing the MTU on a slow link to improve interactive response.

You should make max at least twice min to prevent synchronisation. On slow
links with small min's it might be wise to make max perhaps four or
more times large then min.

Burst controls how the RED algorithm responds to bursts. Burst must be set
larger then min/avpkt. Experimentally, I've found (min+min+max)/(3*avpkt) to
work okay.

Additionally, you need to set limit and avpkt. Limit is a safety value, after
there are limit bytes in the queue, RED 'turns into' tail-drop. I typical set
limit to eight times max. Avpkt should be your average packet size. 1000
works okay on high speed Internet links with a 1500byte MTU. 

Read <url url="http://www.aciri.org/floyd/papers/red/red.html"
name="the paper on RED queueing"> by Sally Floyd and Van Jacobson for technical
information.

FIXME: more needed. This means *you* greg :-) - ahu


<sect>Cookbook
<p>
This section contains 'cookbook' entries which may help you solve problems.
A cookbook is no replacement for understanding however, so try and comprehend
what is going on. 
<!-- 
<sect1>Reserving bandwidth for your IRC server
<p>
Recently the IRC networks have been plagued by distributed denial of service
attacks. The aim of some of these attacks is to disrupt communication
between servers which split the network. You then join the splitted part
of the network. Because nobody else is there, the server assigns you
operator status. You then stop the disruption, the network rejoins and
voila, you can take over the channel.

This silly behaviour is seriously damaging IRC, and luckily, Linux is there
to protect it :-) 

We need to be smarter than your average scriptkid, so we'll use some
advanced netfilter features to help us.
-->

<sect1>Running multiple sites with different SLAs
<p>
You can do this in several ways. Apache has some support for this with a
module, but we'll show how Linux can do this for you, and do so for other
services as well. These commands are stolen from a presentation by Jamal
Hadi that's referenced below.

Let's say we have two customers, with http, ftp and streaming audio, and we
want to sell them a limited amount of bandwidth. We do so on the server itself.

Customer A should have at most 2 megabits, customer B has paid for 5
megabits. We separate our customers by creating virtual IP addresses on our
server.

<tscreen><verb>
# ip address add 188.177.166.1 dev eth0
# ip address add 188.177.166.2 dev eth0
</verb></tscreen>

It is up to you to attach the different servers to the right IP address. All
popular daemons have support for this.

We first attach a CBQ qdisc to eth0:
<tscreen><verb>
# tc qdisc add dev eth0 root handle 1: cbq bandwidth 10Mbit cell 8 avpkt 1000 \
  mpu 64
</verb></tscreen>

We then create classes for our customers:

<tscreen><verb>
# tc class add dev eth0 parent 1:0 classid 1:1 cbq bandwidth 10Mbit rate \
  2MBit avpkt 1000 prio 5 bounded isolated allot 1514 weight 1 maxburst 21
# tc class add dev eth0 parent 1:0 classid 1:2 cbq bandwidth 10Mbit rate \
  5Mbit avpkt 1000 prio 5 bounded isolated allot 1514 weight 1 maxburst 21
</verb></tscreen>

Then we add filters for our two classes:
<tscreen><verb>
##FIXME: Why this line, what does it do?, what is a divisor?:
##FIXME: A divisor has something to do with a hash table, and the number of
##       buckets - ahu
# tc filter add dev eth0 parent 1:0 protocol ip prio 5 handle 1: u32 divisor 1
# tc filter add dev eth0 parent 1:0 prio 5 u32 match ip src 188.177.166.1
  flowid 1:1
# tc filter add dev eth0 parent 1:0 prio 5 u32 match ip src 188.177.166.2
  flowid 1:2
</verb></tscreen>

And we're done.

FIXME: why no token bucket filter? is there a default pfifo_fast fallback
somewhere?

<sect1>Protecting your host from SYN floods 
<p>From Alexey's iproute documentation, adapted to netfilter and with more
plausible paths. If you use this, take care to adjust the numbers to
reasonable values for your system.

If you want to protect an entire network, skip this script, which is best
suited for a single host.

It appears that you need the very latest version of the iproute2 tools to
get this to work with 2.4.0.

<tscreen><verb>
#! /bin/sh -x
#
# sample script on using the ingress capabilities
# this script shows how one can rate limit incoming SYNs
# Useful for TCP-SYN attack protection. You can use
# IPchains to have more powerful additions to the SYN (eg 
# in addition the subnet)
#
#path to various utilities;
#change to reflect yours.
#
TC=/sbin/tc
IP=/sbin/ip
IPTABLES=/sbin/iptables
INDEV=eth2
#
# tag all incoming SYN packets through $INDEV as mark value 1
############################################################ 
$iptables -A PREROUTING -i $INDEV -t mangle -p tcp --syn \
  -j MARK --set-mark 1
############################################################ 
#
# install the ingress qdisc on the ingress interface
############################################################ 
$TC qdisc add dev $INDEV handle ffff: ingress
############################################################ 

#
# 
# SYN packets are 40 bytes (320 bits) so three SYNs equals
# 960 bits (approximately 1kbit); so we rate limit below
# the incoming SYNs to 3/sec (not very useful really; but
#serves to show the point - JHS
############################################################ 
$TC filter add dev $INDEV parent ffff: protocol ip prio 50 handle 1 fw \
police rate 1kbit burst 40 mtu 9k drop flowid :1
############################################################ 


#
echo "---- qdisc parameters Ingress  ----------"
$TC qdisc ls dev $INDEV
echo "---- Class parameters Ingress  ----------"
$TC class ls dev $INDEV
echo "---- filter parameters Ingress ----------"
$TC filter ls dev $INDEV parent ffff:

#deleting the ingress qdisc
#$TC qdisc del $INDEV ingress
</verb></tscreen>
<sect1>Ratelimit ICMP to prevent dDoS
<p>
Recently, distributed denial of service attacks have become a major nuisance
on the internet. By properly filtering and ratelimiting your network, you can
both prevent becoming a casualty or the cause of these attacks.

You should filter your networks so that you do not allow non-local IP source
addressed packets to leave your network. This stops people from anonymously
sending junk to the internet. 

<!-- FIXME: netfilter one liner.  Is there a netfilter one-liner? Martijn -->


Rate limiting goes much as shown earlier. To refresh your memory, our
ASCIIgram again:

<tscreen><verb>
[The Internet] ---<E3, T3, whatever>--- [Linux router] --- [Office+ISP]
                                      eth1          eth0
</verb></tscreen>

We first set up the prerequisite parts:

<tscreen><verb>
# tc qdisc add dev eth0 root handle 10: cbq bandwidth 10Mbit avpkt 1000
# tc class add dev eth0 parent 10:0 classid 10:1 cbq bandwidth 10Mbit rate \
  10Mbit allot 1514 prio 5 maxburst 20 avpkt 1000
</verb></tscreen>

If you have 100Mbit, or more, interfaces, adjust these numbers. Now you need
to determine how much ICMP traffic you want to allow. You can perform
measurements with tcpdump, by having it write to a file for a while, and
seeing how much ICMP passes your network. Do not forget to raise the
snapshot length!

If measurement is impractical, you might want to choose 5% of your available
bandwidth. Let's set up our class:
<tscreen><verb>
# tc class add dev eth0 parent 10:1 classid 10:100 cbq bandwidth 10Mbit rate \
  100Kbit allot 1514 weight 800Kbit prio 5 maxburst 20 avpkt 250 \
  bounded
</verb></tscreen>

This limits at 100Kbit. Now we need a filter to assign ICMP traffic to this
class:
<tscreen><verb>
# tc filter add dev eth0 parent 10:0 protocol ip prio 100 u32 match ip
  protocol 1 0xFF flowid 10:100

</verb></tscreen>

<sect1>Prioritizing interactive traffic
<p>
If lots of data is coming down your link, or going up for that matter, and
you are trying to do some maintenance via telnet or ssh, this may not go too
well. Other packets are blocking your keystrokes. Wouldn't it be great if
there were a way for your interactive packets to sneak past the bulk
traffic? Linux can do this for you!

As before, we need to handle traffic going both ways. Evidently, this works
best if there are Linux boxes on both ends of your link, although other
UNIX's are able to do this. Consult your local Solaris/BSD guru for this.

The standard pfifo_fast scheduler has 3 different 'bands'. Traffic in band 0
is transmitted first, after which traffic in band 1 and 2 gets considered.
It is vital that our interactive traffic be in band 0!

We blatantly adapt from the (soon to be obsolete) ipchains HOWTO:

There are four seldom-used bits in the IP header, called the Type of Service
(TOS) bits. They effect the way packets are treated; the four bits are
"Minimum Delay", "Maximum Throughput", "Maximum Reliability" and "Minimum
Cost". Only one of these bits is allowed to be set. Rob van Nieuwkerk, the
author of the ipchains TOS-mangling code, puts it as follows:

<tscreen>
Especially the "Minimum Delay" is important for me. I switch it on for
"interactive" packets in my upstream (Linux) router. I'm
behind a 33k6 modem link. Linux prioritizes packets in 3 queues. This
way I get acceptable interactive performance while doing bulk
downloads at the same time. 
</tscreen>

The most common use is to set telnet & ftp control connections to "Minimum
Delay" and FTP data to "Maximum Throughput". This would be
done as follows, on your upstream router:

<tscreen><verb>
# iptables -A PREROUTING -t mangle -p tcp --sport telnet \
  -j TOS --set-tos Minimize-Delay
# iptables -A PREROUTING -t mangle -p tcp --sport ftp \
  -j TOS --set-tos Minimize-Delay
# iptables -A PREROUTING -t mangle -p tcp --sport ftp-data \
  -j TOS --set-tos Maximize-Throughput
</verb></tscreen>

Now, this only works for data going from your telnet foreign host to your
local computer. The other way around appears to be done for you, ie, telnet,
ssh & friends all set the TOS field on outgoing packets automatically.

Should you have a client that does not do this, you can always do it with
netfilter. On your local box:

<tscreen><verb>
# iptables -A OUTPUT -t mangle -p tcp --dport telnet \
  -j TOS --set-tos Minimize-Delay
# iptables -A OUTPUT -t mangle -p tcp --dport ftp \
  -j TOS --set-tos Minimize-Delay
# iptables -A OUTPUT -t mangle -p tcp --dport ftp-data \
  -j TOS --set-tos Maximize-Throughput
</verb></tscreen>

<sect1>Transparent web-caching using netfilter, iproute2, ipchains and squid
<p>
<label id="SQUID">
This section was sent in by reader Ram Narula from Internet for Education
(Thailand).

The regular technique in accomplishing this in Linux
is probably with use of ipchains AFTER making sure
that the "outgoing" port 80(web) traffic gets routed through
the server running squid.

There are 3 common methods to make sure "outgoing"
port 80 traffic gets routed to the server running squid
and 4th one is being introduced here.

<descrip>
<tag>Making the gateway router do it.</tag>
If you can tell your gateway router to 
match packets that has outgoing destination port
of 80 to be sent to the IP address of squid server.
<p>
     BUT
<p>
This would put additional load on the router and
some commercial routers might not even support this.
<tag>Using a Layer 4 switch.</tag>
Layer 4 switches can handle this without any problem.
<p>
     BUT
<p>
The cost for this equipment is usually very high. Typical
layer 4 switch would normally cost more than
a typical router+good linux server.
<tag>Using cache server as network's gateway.</tag>
You can force ALL traffic through cache server.
<p>
    BUT
<p>
This is quite risky because Squid does
utilize lots of cpu power which might
result in slower over-all network performance
or the server itself might crash and no one on the
network will be able to access the internet if
that occurs.


<tag>Linux+NetFilter router.</tag>
By using NetFilter another technique can be implemented
which is using NetFilter for "mark"ing the packets
with destination port 80 and using iproute2 to
route the "mark"ed packets to the Squid server.
</descrip>
<tscreen><verb>
|----------------|
| Implementation |
|----------------|

 Addresses used
 10.0.0.1 naret (NetFilter server)
 10.0.0.2 silom (Squid server)
 10.0.0.3 donmuang (Router connected to the internet)
 10.0.0.4 kaosarn (other server on network)
 10.0.0.5 RAS
 10.0.0.0/24 main network
 10.0.0.0/19 total network

|---------------|
|Network diagram|
|---------------|

Internet
|
donmuang
|
------------hub/switch----------
|        |             |       |
naret   silom        kaosarn  RAS etc.
</verb></tscreen>
First, make all traffic pass through naret by making
sure it is the default gateway except for silom.
Silom's default gateway has to be donmuang (10.0.0.3) or
this would create web traffic loop.


<p>
(all servers on my network had 10.0.0.1 as the default gateway
which was the former IP address of donmuang router so what I did
was changed the IP address of donmuang to 10.0.0.3 and gave
naret ip address of 10.0.0.1)

<tscreen><verb>
Silom
-----
-setup squid and ipchains 
</verb></tscreen>

<p>
Setup Squid server on silom, make sure it does support
transparent caching/proxying, the default port is usually
3128, so all traffic for port 80 has to be redirected to port
3128 locally. This can be done by using ipchains with the following:

<tscreen><verb>
silom# ipchains -N allow1
silom# ipchains -A allow1 -p TCP -s 10.0.0.0/19 -d 0/0 80 -j REDIRECT 3128
silom# ipchains -I input -j allow1
</verb></tscreen>

<p> 

Or, in netfilter lingo:
<tscreen><verb>
silom# iptables -t nat -A PREROUTING -i eth0 -p tcp --dport 80 -j REDIRECT --to-port 3128
</verb></tscreen>

(note: you might have other entries as well)

<p>
For more information on setting Squid server please refer
to Squid faq page on <url
url="http://squid.nlanr.net" name="http://squid.nlanr.net">).


<p>
Make sure ip forwarding is enabled on this server and the default
gateway for this server is donmuang router (NOT naret).



<tscreen><verb>
Naret
-----
-setup iptables and iproute2
-disable icmp REDIRECT messages (if needed)
</verb></tscreen>

<enum>
<item>"Mark" packets of destination port 80 with value 2
<tscreen><verb> 
naret# iptables -A PREROUTING -i eth0 -t mangle -p tcp --dport 80 \
 -j MARK --set-mark 2
</verb></tscreen>
</item>
<item>Setup iproute2 so it will route packets with "mark" 2 to silom
<tscreen><verb>
naret# echo 202 www.out >> /etc/iproute2/rt_tables
naret# ip rule add fwmark 2 table www.out
naret# ip route add default via 10.0.0.2 dev eth0 table www.out
naret# ip route flush cache

</verb></tscreen>
<p>
If donmuang and naret is on the same subnet then
naret should not send out icmp REDIRECT messages.
In this case it is, so icmp REDIRECTs has to be
disabled by:
<tscreen><verb>
naret# echo 0 > /proc/sys/net/ipv4/conf/all/send_redirects
naret# echo 0 > /proc/sys/net/ipv4/conf/default/send_redirects
naret# echo 0 > /proc/sys/net/ipv4/conf/eth0/send_redirects

</verb></tscreen>
</item>
</enum>

The setup is complete, check the configuration

<tscreen><verb>
On naret:

naret# iptables -t mangle -L
Chain PREROUTING (policy ACCEPT)
target     prot opt source               destination         
MARK       tcp  --  anywhere             anywhere           tcp dpt:www MARK set 0x2 

Chain OUTPUT (policy ACCEPT)
target     prot opt source               destination         

naret# ip rule ls
0:      from all lookup local 
32765:  from all fwmark        2 lookup www.out 
32766:  from all lookup main 
32767:  from all lookup default 

naret# ip route list table www.out
default via 203.114.224.8 dev eth0 

naret# ip route   
10.0.0.1 dev eth0  scope link 
10.0.0.0/24 dev eth0  proto kernel  scope link  src 10.0.0.1
127.0.0.0/8 dev lo  scope link 
default via 10.0.0.3 dev eth0 

(make sure silom belongs to one of the above lines, in this case
it's the line with 10.0.0.0/24)

|------|
|-DONE-|
|------|


</verb></tscreen>
<sect2>Traffic flow diagram after implementation
<p>
<tscreen><verb>

|-----------------------------------------|
|Traffic flow diagram after implementation|
|-----------------------------------------|

INTERNET
/\
||
\/
-----------------donmuang router---------------------
/\                                      /\         ||
||                                      ||         ||
||                                      \/         ||
naret                                  silom       ||
*destination port 80 traffic=========>(cache)      ||
/\                                      ||         ||
||                                      \/         \/
\\===================================kaosarn, RAS, etc.

</verb></tscreen>

Note that the network is asymmetric as there is one extra hop on 
general outgoing path.

<tscreen><verb>
Here is run down for packet traversing the network from kaosarn
to and from the internet.

For web/http traffic:
kaosarn http request->naret->silom->donmuang->internet
http replies from internet->donmuang->silom->kaosarn

For non-web/http requests(eg. telnet):
kaosarn outgoing data->naret->donmuang->internet
incoming data from internet->donmuang->kaosarn
</verb></tscreen>

<sect1>Circumventing Path MTU Discovery issues with per route MTU settings
<p>
For sending bulk data, the internet generally works better when using larger
packets. Each packet implies a routing decision, when sending a 1 megabyte
file, this can either mean around 700 packets when using packets that are as
large as possible, or 4000 if using the smallest default.

However, not all parts of the internet support full 1460 bytes of payload
per packet. It is therefore necessary to try and find the largest packet
that will 'fit', in order to optimize a connection.

This process is called 'Path MTU Discovery', where MTU stands for 'Maximum
Transfer Unit.' 

When a router encounters a packet that's too big too send in one piece, AND
it has been flagged with the "Don't Fragment" bit, it returns an ICMP
message stating that it was forced to drop a packet because of this. The
sending host acts on this hint by sending smaller packets, and by iterating
it can find the optimum packet size for a connection over a certain path.

This used to work well until the internet was discovered by hooligans who do
their best to disrupt communications. This in turn lead administrators to
either block or shape ICMP traffic in a misguided attempt to improve
security or robustness of their internet service.

What has happened now is that Path MTU Discovery is working less and less
well and fails for certain routes, which leads to strange TCP/IP sessions
which die after a while.

Although I have no proof for this, two sites who I used to have this problem
with both run Alteon Acedirectors before the affected systems - perhaps
somebody more knowledgeable can provide clues as to why this happens.

<sect2>Solution
<p>
When you encounter sites that suffer from this problem, you can disable Path
MTU discovery by setting it manually. Koos van den Hout, slightly edited,
writes:

<tscreen>
<p>
The following problem: I set the mtu/mru of my leased line running ppp to
296 because it's only 33k6 and I cannot influence the queueing on the
other side. At 296, the response to a keypress is within a reasonable
timeframe.

And, on my side I have a masqrouter running (of course) Linux.

Recently I split 'server' and 'router' so most applications are run on a
different machine than the routing happens on.

I then had trouble logging into irc. Big panic! Some digging did find
out that I got connected to irc, even showed up as 'connected' on irc
but I did not receive the motd from irc. I checked what could be wrong
and noted that I already had some previous trouble reaching certain
websites related to the MTU, since I had no trouble reaching them when
the MTU was 1500, the problem just showed when the MTU was set to 296.
Since irc servers block about every kind of traffic not needed for their
immediate operation, they also block icmp. 

I managed to convince the operators of a webserver that this was the cause
of a problem, but the irc server operators were not going to fix this.

So, I had to make sure outgoing masqueraded traffic started with the lower
mtu of the outside link. But I want local ethernet traffic to have the
normal mtu (for things like nfs traffic).

Solution:
<tscreen><verb>
ip route add default via 10.0.0.1 mtu 296
</verb></tscreen>

(10.0.0.1 being the default gateway, the inside address of the
masquerading router)
</tscreen>

In general, it is possible to override PMTU Discovery by setting specific
routes. For example, if only a certain subnet is giving problems, this
should help:

<tscreen><verb>
ip route add 195.96.96.0/24 via 10.0.0.1 mtu 1000
</verb></tscreen>
<sect1>Circumventing Path MTU Discovery issues with MSS Clamping (for ADSL,
cable, PPPoE &amp; PPtP users)
<p>
As explained above, Path MTU Discovery doesn't work as well as it should
anymore. If you know for a fact that a hop somewhere in your network has a
limited (&lt;1500) MTU, you cannot rely on PMTU Discovery finding this out.

Besides MTU, there is yet another way to set the maximum packet size, the so
called Maximum Segment Size. This is a field in the TCP Options part of a
SYN packet.

Recent Linux kernels, and a few pppoe drivers (notably, the excellent
Roaring Penguin one), feature the possibility to 'clamp the MSS'. 

The good thing about this is that by setting the MSS value, you are telling
the remote side unequivocally 'do not ever try to send me packets bigger
than this value'. No ICMP traffic is needed to get this to work.

The bad thing is that it's an obvious hack - it breaks 'end to end' by
modifying packets. Having said that, we use this trick in many places and it
works like a charm.

In order for this to work you need at least iptables-1.2.1a and Linux 2.4.3
or higher. The basic commandline is:
<tscreen><verb>
# iptables -A FORWARD -p tcp --tcp-flags SYN,RST SYN -j TCPMSS  --clamp-mss-to-pmtu
</verb></tscreen>

This calculates the proper MSS for your link. If you are feeling brave, or
think that you know best, you can also do something like this:

<tscreen><verb>
# iptables -A FORWARD -p tcp --tcp-flags SYN,RST SYN -j TCPMSS --set-mss 128
</verb></tscreen>

This sets the MSS of passing SYN packets to 128. Use this if you have VoIP
with tiny packets, and huge http packets which are causing chopping in your
voice calls.
<sect>Building bridges, and pseudo-bridges with Proxy ARP
<p>
Bridges are devices which can be installed in a network without any
reconfiguration. A network switch is basically a many-port bridge. A bridge
is often a 2-port switch. Linux does however support multiple interfaces in
a bridge, making it a true switch.

Bridges are often deployed when confronted with a broken network that needs
to be fixed without any alterations. Because the bridge is a layer-2 device,
one layer below IP, routers and servers are not aware of its existence.
This means that you can transparently block or modify certain packets, or do
shaping.

Another good thing is that a bridge can often be replaced by a cross cable
or a hub, should it break down.

The bad news is that a bridge can cause great confusion unless it is very
well documented. It does not appear in traceroutes, but somehow packets
disappear or get changed from point A to point B. You should also wonder if
an organization that 'does not want to change anything' is doing the right
thing.

The Linux 2.4 bridge is documented on

<url url=" http://bridge.sourceforge.net/" name="this page">.

<sect1>State of bridging and iptables
<p>
As of Linux 2.4.14, bridging and iptables do not 'see' each other without
help. If you bridge packets from eth0 to eth1, they do not 'pass' by
iptables. This means that you cannot do filtering, or NAT or mangling or
whatever.

There are several projects going on to fix this, the truly right one is by
the author of the Linux 2.4 bridging code, Lennert Buytenhek. He recently
informed us that as of bridge-nf 0.0.2 (see the url above), the code is
stable and usable in production environments.  He is now asking the kernel
people if and how the patch can be merged, stay tuned!

<sect1>Bridging and shaping
<p>
This does work as advertised. Be sure to figure out which side each
interface is on, otherwise you might be shaping outbound traffic in your
internal interface, which won't work. Use tcpdump if needed.

<sect1>Pseudo-bridges with Proxy-ARP
<p>
If you just want to implement a Pseudo-bridge, skip down a few sections 
to 'Implementing it', but it is wise to read a bit about how it works in
practice.

A Pseudo-bridge works a bit differently. By default, a bridge passes packets
unaltered from one interface to the other. It only looks at the hardware
address of packets to determine what goes where. This in turn means that you
can bridge traffic that Linux does not understand, as long as it has an
hardware address it does.

A 'Pseudo-bridge' works differently and looks more like a hidden router than
a bridge, but like a bridge, it has little impact on network design.

An advantage of the fact that it is not a brige lies in the fact that
packets really pass through the kernel, and can be filtered, changed,
redirected or rerouted. 

A real bridge can also be made to perform these feats, but it needs special
code, like the Ethernet Frame Diverter, or the above mentioned patch.

Another advantage of a pseudo-bridge is that it does not pass packets it
does not understand - thus cleaning your network of a lot of cruft. In cases
where you need this cruft (like SAP packets, or Netbeui), use a real bridge.
<sect2>ARP &amp; Proxy-ARP
<p>
When a host wants to talk to another host on the same physical network
segment, it sends out an Address Resolution Protocol packet, which, somewhat
simplified, reads like this 'who has 10.0.0.1, tell 10.0.0.7'. In response
to this, 10.0.0.1 replies with a short 'here' packet.

10.0.0.7 then sends packets to the hardware address mentioned in the 'here' 
packet. It caches this hardware address for a relatively long time, and
after the cache expires, it reasks the question.

When building a Pseudo-bridge, we instruct the bridge to reply to these ARP
packets, which causes the hosts in the network to send its packets to the
bridge. The brige then processes these packets, and sends them to the
relevant interface.

So, in short, whenever a host on one side of the bridge asks for the
hardware address of a host on the other, the bridge replies with a packet
that says 'hand it to me'.

This way, all data traffic gets transmitted to the right place, and always
passes through the bridge.
<sect2>Implementing it
<p>
In the bad old days, it used to be possible to instruct the Linux Kernel to
perform 'proxy-ARP' for just any subnet. So, to configure a pseudo-bridge,
you would have to specify both the proper routes to both sides of the bridge
AND create matching proxy-ARP rules. This is bad in that it requires a lot
of typing, but also because it easily allows you to make mistakes which make
your bridge respond to ARP queries for networks it does not know how to
route.

With Linux 2.4 (and possibly 2.2), this possibility has been withdrawn and
has been replaced by a flag in the /proc directory, called 'proxy_arp'. The
procedure for building a pseudo-bridge is then:

<enum>
<item>Assign an IP address to both interfaces, the 'left' and the 'right'
one
<item>Create routes so your machine knows which hosts reside on the left,
and which on the right
<item>Turn on proxy-ARP on both interfaces, echo 1 >
/proc/sys/net/ipv4/conf/ethL/proxy_arp, echo 1 >
/proc/sys/net/ipv4/conf/ethR/proxy_arp, where L and R stand for the numbers
of your interfaces on the left and on the right side
</enum>

Also, do not forget to turn on the ip_forwarding flag! When converting from
a true bridge, you may find that this flag was turned off as it is not
needed when bridging.

Another thing you might note when converting is that you need to clear the
arp cache of computers in the network - the arp cache might contain old
pre-bridge hardware addresses which are no longer correct.

On a Cisco, this is done using the command 'clear arp-cache', under
Linux, use 'arp -d ip.address'. You can also wait for the cache to expire
manually, which can take rather long.
<sect>Advanced Linux Routing
<p>
This section is for all you people who either want to understand why the
whole system works or have a configuration that's so bizarre that you
need the low down to make it work.

This section is completely optional. It's quite possible that this section
will be quite complex and really not intended for normal users. You have
been warned.

FIXME: Decide what really needs to go in here.
<sect1>How does packet queueing really work?
<p>This is the low-down on how the packet queueing system really works.

Lists the steps the kernel takes to classify a packet, etc...

FIXME: Write this.

<sect1>Advanced uses of the packet queueing system
<p>Go through Alexey's extremely tricky example involving the unused bits
in the TOS field.

FIXME: Write this.

<sect1>Other packet shaping systems
<p>I'd like to include a brief description of other packet shaping systems
in other operating systems and how they compare to the Linux one. Since Linux
is one of the few OSes that has a completely original (non-BSD derived) TCP/IP
stack, I think it would be useful to see how other people do it.

Unfortunately I have no experience with other systems so cannot write this.

FIXME: Anyone? - Martijn

<sect>Dynamic routing - OSPF and BGP
<p>
Once your network starts to get really big, or you start to consider 'the
internet' as your network, you need tools which dynamically route your data.
Sites are often connected to each other with multiple links, and more are
popping up all the time. 

The Internet has mostly standardised on OSPF and BGP4 (rfc1771). Linux
supports both, by way of <tt>gated</tt> and <tt>zebra</tt>

While currently not within the scope of this document, we would like to
point you to the definitive works:
 
Overview:
 
  Cisco Systems
  <url
url="http://www.cisco.com/univercd/cc/td/doc/cisintwk/idg4/nd2003.htm"
name="Designing large-scale IP internetworks">

 
 
For OSPF:
 
  Moy, John T.
  "OSPF.  The anatomy of an Internet routing protocol"
  Addison Wesley. Reading, MA. 1998.
 
Halabi has also written a good guide to OSPF routing design, but this
appears to have been dropped from the Cisco web site.
 
 
For BGP:
 
  Halabi, Bassam
  "Internet routing architectures"
  Cisco Press (New Riders Publishing). Indianapolis, IN. 1997.
 
also
 
  Cisco Systems
  
  <url
url="http://www.cisco.com/univercd/cc/td/doc/cisintwk/ics/icsbgp4.htm"
name="Using the Border Gateway Protocol for interdomain routing">

 
Although the examples are Cisco-specific, they are remarkably similar
to the configuration language in Zebra :-)
<sect>Other possibilities 
<p>
This chapter is a list of projects having to do with advanced Linux routing
&amp; traffic shaping. Some of these links may deserve chapters of their
own, some are documented very well of themselves, and don't need more HOWTO.
<p>
<descrip>
<tag>802.1Q VLAN Implementation for Linux <url url="http://scry.wanfear.com/~greear/vlan.html"
name="(site)"></tag>
<p>
VLANs are a very cool way to segregate your
networks in a more virtual than physical way. Good information on VLANs can
be found <url
url="ftp://ftp.netlab.ohio-state.edu/pub/jain/courses/cis788-97/virtual_lans/index.htm"
name="here">. With this implementation, you can have your Linux box talk
VLANs with machines like Cisco Catalyst, 3Com: {Corebuilder, Netbuilder II,
SuperStack II switch 630}, Extreme Ntwks Summit 48, Foundry: {ServerIronXL,
FastIron}.

Update: has been included in the kernel as of 2.4.14 (perhaps 13).
<tag>Alternate 802.1Q VLAN Implementation for Linux <url 
url="http://vlan.sourceforge.net "
name="(site)"></tag>
Alternative VLAN implementation for linux. This project was started out of
disagreement with the 'established' VLAN project's architecture and coding
style, resulting in a cleaner overall design.

<tag>Linux Virtual Server <url url="http://www.LinuxVirtualServer.org/"  
name="(site)"></tag>
<p>
These people are brilliant. The Linux Virtual Server is a highly scalable and
highly available server built on a cluster of real servers, with the load
balancer running on the Linux operating system. The architecture of the
cluster is transparent to end users. End users only see a single virtual
server.

In short whatever you need to loadbalance, at whatever level of traffic, LVS
will have a way of doing it. Some of their techniques are positively evil!
For example, they let several machines have the same IP address on a
segment, but turn off ARP on them. Only the LVS machine does ARP - it then
decides which of the backend hosts should handle an incoming packet, and
sends it directly to the right MAC address of the backend server. Outgoing
traffic will flow directly to the router, and not via the LVS machine, which
does therefor not need to see your 5Gbit/s of content flowing to the world,
and cannot be a bottleneck.

The LVS is implemented as a kernel patch in Linux 2.0 and 2.2, but as a
Netfilter module in 2.4, so it does not need kernel patches! Their 2.4
support is still in early development, so beat on it and give feedback or
send patches.

<tag>CBQ.init <url url="ftp://ftp.equinox.gu.net/pub/linux/cbq/"
name="(site)"></tag>
Configuring CBQ can be a bit daunting, especially if all you want to do is
shape some computers behind a router. CBQ.init can help you configure Linux
with a simplified syntax.

For example, if you want all computers in your 192.168.1.0/24 subnet
(on 10mbit eth1) to be limited to 28kbit/s download speed, put
this in the CBQ.init configuration file:

<tscreen><verb>
DEVICE=eth1,10Mbit,1Mbit
RATE=28Kbit
WEIGHT=2Kbit
PRIO=5
RULE=192.168.1.0/24
</verb></tscreen>

By all means use this program if the 'how and why' don't interest you.
We're using CBQ.init in production and it works very well. It can even do
some more advanced things, like time dependent shaping. The documentation is
embedded in the script, which explains why you can't find a README.

<tag>Chronox easy shaping scripts <url url="http://www.chronox.de"
name="(site)"></tag>

Stephan Mueller (smueller@chronox.de) wrote two useful scripts, 'limit.conn'
and 'shaper'. The first one allows you to easily throttle a single download
session, like this:

<tscreen><verb>
# limit.conn -s SERVERIP -p SERVERPORT -l LIMIT
</verb></tscreen>

It works on Linux 2.2 and 2.4.

The second script is more complicated, and can be used to make lots of
different queues based on iptables rules, which are used to mark packets
which are then shaped.

<tag>Virtual Router
Redundancy Protocol implementation <url url="http://w3.arobas.net/~jetienne/vrrpd/index.html"
 name="(site)"></tag>
<p>
This is purely for redundancy. Two machines with their own IP address and
MAC Address together create a third IP Address and MAC Address, which is
virtual. Originally intended purely for routers, which need constant MAC
addresses, it also works for other servers.
<p>
The beauty of this approach is the incredibly easy configuration. No kernel
compiling or patching required, all userspace.

Just run this on all machines participating in a service:
<tscreen><verb>
# vrrpd -i eth0 -v 50 10.0.0.22
</verb></tscreen>

And you are in business! 10.0.0.22 is now carried by one of your servers,
probably the first one to run the vrrp daemon. Now disconnect that computer
from the network and very rapidly one of the other computers will assume the
10.0.0.22 address, as well as the MAC address.

I tried this over here and had it up and running in 1 minute. For some
strange reason it decided to drop my default gateway, but the -n flag
prevented that.

This is a 'live' failover:

<tscreen><verb>
64 bytes from 10.0.0.22: icmp_seq=3 ttl=255 time=0.2 ms
64 bytes from 10.0.0.22: icmp_seq=4 ttl=255 time=0.2 ms
64 bytes from 10.0.0.22: icmp_seq=5 ttl=255 time=16.8 ms
64 bytes from 10.0.0.22: icmp_seq=6 ttl=255 time=1.8 ms
64 bytes from 10.0.0.22: icmp_seq=7 ttl=255 time=1.7 ms
</verb></tscreen>

Not *one* ping packet was lost! Just after packet 4, I disconnected my P200
from the network, and my 486 took over, which you can see from the higher
latency.
</descrip>
<sect>Further reading
<p>
<descrip>
<tag><url url="http://snafu.freedom.org/linux2.2/iproute-notes.html"
name="http://snafu.freedom.org/linux2.2/iproute-notes.html"></tag>
Contains lots of technical information, comments from the kernel
<tag><url url="http://www.davin.ottawa.on.ca/ols/"
name="http://www.davin.ottawa.on.ca/ols/"></tag>
Slides by Jamal Hadi Salim, one of the authors of Linux traffic control
<tag><url url="http://defiant.coinet.com/iproute2/ip-cref/"
name="http://defiant.coinet.com/iproute2/ip-cref/"></tag>
HTML version of Alexeys LaTeX documentation - explains part of iproute2 in
great detail
<tag><url url="http://www.aciri.org/floyd/cbq.html"
name="http://www.aciri.org/floyd/cbq.html"></tag>
Sally Floyd has a good page on CBQ, including her original papers. None of
it is Linux specific, but it does a fair job discussing the theory and uses
of CBQ.
Very technical stuff, but good reading for those so inclined. 

<tag><url url="http://ceti.pl/~kravietz/cbq/NET4_tc.html"
name="http://ceti.pl/~kravietz/cbq/NET4_tc.html"></tag>
Yet another HOWTO, this time in Polish! You can copy/paste command lines
however, they work just the same in every language. The author is
cooperating with us and may soon author sections of this HOWTO.

<tag><url
url="http://snafu.freedom.org/linux2.2/docs/draft-almesberger-wajhak-diffserv-linux-00.txt"
name="Differentiated Services on Linux"></tag>
Discussion on how to use Linux in a diffserv compliant environment. Pretty
far removed from your everyday routing needs, but very interesting none the
less. We may include a section on this at a later date.

<tag><url
url="http://www.cisco.com/univercd/cc/td/doc/product/software/ios111/cc111/car.htm"
name="IOS Committed Access Rate"></tag>
<label id="CAR">
From the helpful folks of Cisco who have the laudable habit of putting
their documentation online. Cisco syntax is different but the concepts are
the same, except that we can do more and do it without routers the price of
cars :-)

<tag>Docum experimental site<url url="http://www.docum.org"
name="(site)"></tag>
Stef Coene is busy convincing his boss to sell Linux support, and so he is
experimenting a lot, especially with managing bandwidth. His site has a lot
of practical information, examples, tests and also points out some CBQ/tc bugs. 

<tag>TCP/IP Illustrated, volume 1, W. Richard Stevens, ISBN 0-201-63346-9</tag>
Required reading if you truly want to understand TCP/IP. Entertaining as
well.

</descrip> 
<sect>Acknowledgements 
<p> 
It is our goal to list everybody who has contributed to this HOWTO, or
helped us demystify how things work. While there are currently no plans
for a Netfilter type scoreboard, we do like to recognise the people who are
helping.

<itemize>
<item>Ron Brinker &lt;service%emcis.com&gt;
<item>Lennert Buytenhek &lt;buytenh@gnu.org&gt;
<item>Esteve Camps &lt;esteve@hades.udg.es&gt;
<item>Gerry Creager N5JXS &lt;gerry%cs.tamu.edu&gt;
<item>Marco Davids &lt;marco@sara.nl&gt;
<item>Jonathan Day &lt;jd9812@my-deja.com&gt;
<item>Martin Devera aka devik &lt;devik@cdi.cz&gt;
<item>Stephan "Kobold" Gehring &lt;Stephan.Gehring@bechtle.de&gt;
<item>Nadeem Hasan &lt;nhasan@usa.net&gt;
<item>Vik Heyndrickx &lt;vik.heyndrickx@edchq.com&gt;
<item>Koos van den Hout &lt;koos@kzdoos.xs4all.nl&gt;
<item>Martin Josefsson &lt;gandalf%wlug.westbo.se&gt;
<item>Pawel Krawczyk &lt;kravietz%alfa.ceti.pl&gt;
<item>Philippe Latu &lt;philippe.latu%linux-france.org&gt;
<item>Jason Lunz &lt;j@cc.gatech.edu&gt;
<item>Alexey Mahotkin &lt;alexm@formulabez.ru&gt;
<item>Andreas Mohr &lt;andi%lisas.de&gt;
<item>Wim van der Most 
<item>Stephan Mueller &lt;smueller@chronox.de&gt;
<item>Ram Narula &lt;ram@princess1.net&gt;
<item>Jorge Novo &lt;jnovo@educanet.net&gt;
<item>Patrik &lt;ph@kurd.nu&gt;
<item>Jason Pyeron &lt;jason%pyeron.com&gt;
<item>Rusty Russell (with apologies for always misspelling your name)
<item>Jamal Hadi Salim &lt;hadi%cyberus.ca&gt;
<item>David Sauer &lt;davids%penguin.cz&gt;
<item>Sheharyar Suleman Shaikh &lt;sss23@drexel.edu&gt;
<item>Nick Silberstein &lt;nhsilber%yahoo.com&gt;
<item>Konrads Smelkov &lt;konrads@interbaltika.com&gt;
<item>Jason Tackaberry &lt;tack@linux.com&gt;
<item>Charles Tassell &lt;ctassell%isn.net&gt;
<item>Glen Turner &lt;glen.turner%aarnet.edu.au&gt;
<item>Song Wang &lt;wsong@ece.uci.edu&gt;
<item>Jonathan Day &lt;jd9812@my-deja.com&gt;
</itemize>

</article>