1 <!DOCTYPE book PUBLIC "-//OASIS//DTD DocBook V4.1//EN">
4 <?dbhtml banner-text="Made possible by PowerDNS">
5 <?dbhtml banner-href="http://www.powerdns.com">
8 <Title>Linux Advanced Routing & Traffic Control HOWTO</Title>
11 <FirstName>Bert</FirstName><Surname>Hubert</Surname>
13 <orgname>Netherlabs BV</orgname>
14 <address><email>bert.hubert@netherlabs.nl</email></address>
19 <collabname>Gregory Maxwell (Section Author)</collabname>
21 <address><email>remco%virtu.nl</email></address>
27 <collabname>Remco van Mook (Section Author)</collabname>
29 <address><email>remco@virtu.nl</email></address>
34 <collabname>Martijn van Oosterhout (Section Author)</collabname>
36 <address><email>kleptog@cupid.suninternet.com</email></address>
41 <collabname>Paul B Schroeder (Section Author)</collabname>
43 <address><email>paulsch@us.ibm.com</email></address>
48 <collabname>Jasper Spaans (Section Author)</collabname>
50 <address><email>jasper@spaans.ds9a.nl</email></address>
54 <collabname>Pedro Larroy (Section Author)</collabname>
56 <address><email>piotr%omega.resa.ed</email></address>
64 <revnumber role="rcs">$Revision$</revnumber>
65 <date role="rcs">$Date$</date>
66 <revremark>DocBook Edition</revremark>
71 <Para>A very hands-on approach to <application>iproute2</application>,
72 traffic shaping and a bit of <application>netfilter</application>.
78 <chapter id="lartc.dedication">
79 <Title>Dedication</Title>
82 This document is dedicated to lots of people, and is my attempt to do
83 something back. To list but a few:
103 The good folks from Google
109 The staff of Casema Internet
119 <chapter id="lartc.intro">
120 <Title>Introduction</Title>
123 Welcome, gentle reader.
127 This document hopes to enlighten you on how to do more with Linux 2.2/2.4
128 routing. Unbeknownst to most users, you already run tools which allow you to
129 do spectacular things. Commands like <command>route</command> and
130 <command>ifconfig</command> are actually
131 very thin wrappers for the very powerful iproute2 infrastructure.
135 I hope that this HOWTO will become as readable as the ones by Rusty Russell
136 of (amongst other things) netfilter fame.
140 You can always reach us by writing to the <ULink
141 URL="mailto:HOWTO@ds9a.nl"
143 >. However, please consider posting to the mailing
144 list (see the relevant section) if you have questions which are not directly
145 related to this HOWTO. We are no free helpdesk, but we often will answer questions
150 Before losing your way in this HOWTO, if all you want to do is simple
151 traffic shaping, skip everything and head to the <citetitle><xref linkend="lartc.other"></citetitle> chapter, and read about CBQ.init.
154 <Sect1 id="lartc.intro.disclaimer">
155 <Title>Disclaimer & License</Title>
158 This document is distributed in the hope that it will be useful,
159 but WITHOUT ANY WARRANTY; without even the implied warranty of
160 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
164 In short, if your STM-64 backbone breaks down and distributes pornography to
165 your most esteemed customers - it's never our fault. Sorry.
169 Copyright (c) 2002 by bert hubert, Gregory Maxwell, Martijn van
170 Oosterhout, Remco van Mook, Paul B. Schroeder and others. This material may
171 be distributed only subject to the terms and conditions set forth in the
172 Open Publication License, v1.0 or later (the latest version is presently
173 available at http://www.opencontent.org/openpub/).
177 Please freely copy and distribute (sell or give away) this document in any
178 format. It's requested that corrections and/or comments be forwarded to the
183 It is also requested that if you publish this HOWTO in hardcopy that you
184 send the authors some samples for <quote>review purposes</quote> :-)
189 <Sect1 id="lartc.intro.prior">
190 <Title>Prior knowledge</Title>
193 As the title implies, this is the <quote>Advanced</quote> HOWTO.
194 While by no means rocket science, some prior knowledge is assumed.
198 Here are some other references which might help teach you more:
202 <ULink URL="http://netfilter.samba.org/unreliable-guides/networking-concepts-HOWTO/index.html">
203 Rusty Russell's networking-concepts-HOWTO</ULink>
206 <Para>Very nice introduction, explaining what a network is, and how it is
207 connected to other networks.
212 <Term>Linux Networking-HOWTO (Previously the Net-3 HOWTO)</Term>
214 <Para>Great stuff, although very verbose. It teaches you a lot of stuff
215 that's already configured if you are able to connect to the Internet.
216 Should be located in <filename>/usr/doc/HOWTO/NET3-4-HOWTO.txt</filename>
217 but can be also be found
218 <ULink URL="http://www.linuxports.com/howto/networking">online</ULink>.
227 <Sect1 id="lartc.intro.linux">
228 <Title>What Linux can do for you</Title>
231 A small list of things that are possible:
236 <Para>Throttle bandwidth for certain computers
240 <Para>Throttle bandwidth TO certain computers
244 <Para>Help you to fairly share your bandwidth
248 <Para>Protect your network from DoS attacks
252 <Para>Protect the Internet from your customers
256 <Para>Multiplex several servers as one, for load balancing or
257 enhanced availability
261 <Para>Restrict access to your computers
265 <Para>Limit access of your users to other hosts
269 <Para>Do routing based on user id (yes!), MAC address, source IP
270 address, port, type of service, time of day or content
276 Currently, not many people are using these advanced features. This is for
277 several reasons. While the provided documentation is verbose, it is not very
278 hands-on. Traffic control is almost undocumented.
283 <Sect1 id="lartc.intro.houskeeping">
284 <Title>Housekeeping notes</Title>
287 There are several things which should be noted about this document. While I
288 wrote most of it, I really don't want it to stay that way. I am a strong
289 believer in Open Source, so I encourage you to send feedback, updates,
290 patches etcetera. Do not hesitate to inform me of typos or plain old errors.
291 If my English sounds somewhat wooden, please realize that I'm not a native
292 speaker. Feel free to send suggestions.
296 If you feel to you are better qualified to maintain a section, or think that
297 you can author and maintain new sections, you are welcome to do so. The SGML
298 of this HOWTO is available via CVS, I very much envision more people
303 In aid of this, you will find lots of FIXME notices. Patches are always
304 welcome! Wherever you find a FIXME, you should know that you are treading in
305 unknown territory. This is not to say that there are no errors elsewhere,
306 but be extra careful. If you have validated something, please let us know so
307 we can remove the FIXME notice.
311 About this HOWTO, I will take some liberties along the road. For example, I
312 postulate a 10Mbit Internet connection, while I know full well that those
318 <Sect1 id="lartc.intro.cvs">
319 <Title>Access, CVS & submitting updates</Title>
322 The canonical location for the HOWTO is
323 <ULink URL="http://www.ds9a.nl/lartc">here</ULink>.
327 We now have anonymous CVS access available to the world at large. This is
328 good in a number of ways. You can easily upgrade to newer versions of this
329 HOWTO and submitting patches is no work at all.
333 Furthermore, it allows the authors to work on the source independently,
338 $ export CVSROOT=:pserver:anon@outpost.ds9a.nl:/var/cvsroot
340 CVS password: [enter 'cvs' (without 's)]
342 cvs server: Updating 2.4routing
343 U 2.4routing/lartc.db
347 If you made changes and want to contribute them, run <userinput>
348 cvs -z3 diff -uBb</userinput>,
349 and mail the output to <email>howto@ds9a.nl</email>, we
350 can then integrate it easily. Thanks! Please make sure that you edit the
351 .db file, by the way, the other files are generated from that one.
355 A Makefile is supplied which should help you create postscript, dvi, pdf,
356 html and plain text. You may need to install
357 <application>docbook</application>, <application>docbook-utils</application>,
358 <application>ghostscript</application> and <application>tetex</application>
363 Be careful not to edit 2.4routing.sgml! It contains an older version of the
364 HOWTO. The right file is lartc.db.
368 <Sect1 id="lartc.intro.mlist">
369 <Title>Mailing list</Title>
372 The authors receive an increasing amount of mail about this HOWTO. Because
373 of the clear interest of the community, it has been decided to start a
374 mailinglist where people can talk to each other about Advanced Routing and
375 Traffic Control. You can subscribe to the list
376 <ULink URL="http://mailman.ds9a.nl/mailman/listinfo/lartc">here</ULink>.
380 It should be pointed out that the authors are very hesitant of answering
381 questions not asked on the list. We would like the archive of the list to
382 become some kind of knowledge base. If you have a question, please search
383 the archive, and then post to the mailinglist.
388 <Sect1 id="lartc.intro.layout">
389 <Title>Layout of this document</Title>
392 We will be doing interesting stuff almost immediately, which also means that
393 there will initially be parts that are explained incompletely or are not
394 perfect. Please gloss over these parts and assume that all will become clear.
398 Routing and filtering are two distinct things. Filtering is documented very
399 well by Rusty's HOWTOs, available here:
404 <Para><ULink URL="http://netfilter.samba.org/unreliable-guides/">
405 Rusty's Remarkably Unreliable Guides</ULink>
410 <Para>We will be focusing mostly on what is possible by combining netfilter
418 <chapter id="lartc.iproute2">
419 <Title>Introduction to iproute2</Title>
421 <Sect1 id="lartc.iproute2.why">
422 <Title>Why iproute2?</Title>
425 Most Linux distributions, and most UNIX's, currently use the
426 venerable <command>arp</command>, <command>ifconfig</command> and
427 <command>route</command> commands.
428 While these tools work, they show some unexpected behaviour under Linux 2.2
430 For example, GRE tunnels are an integral part of routing these days, but
431 require completely different tools.
435 With <application>iproute2</application>, tunnels are an integral part of
440 The 2.2 and above Linux kernels include a completely redesigned network
441 subsystem. This new networking code brings Linux performance and a feature
442 set with little competition in the general OS arena. In fact, the new
443 routing, filtering, and classifying code is more featureful than the one
444 provided by many dedicated routers and firewalls and traffic shaping
449 As new networking concepts have been invented, people have found ways to
450 plaster them on top of the existing framework in existing OSes. This
451 constant layering of cruft has lead to networking code that is filled with
452 strange behaviour, much like most human languages. In the past, Linux
453 emulated SunOS's handling of many of these things, which was not ideal.
457 This new framework makes it possible to clearly express features
458 previously beyond Linux's reach.
463 <Sect1 id="lartc.iproute2.tour">
464 <Title>iproute2 tour</Title>
467 Linux has a sophisticated system for bandwidth provisioning called Traffic
468 Control. This system supports various method for classifying, prioritizing,
469 sharing, and limiting both inbound and outbound traffic.
473 We'll start off with a tiny tour of iproute2 possibilities.
478 <Sect1 id="lartc.iproute2.package">
479 <Title>Prerequisites</Title>
482 You should make sure that you have the userland tools installed. This
483 package is called 'iproute' on both RedHat and Debian, and may otherwise be
484 found at <filename>ftp://ftp.inr.ac.ru/ip-routing/iproute2-2.2.4-now-ss??????.tar.gz"</filename>.
489 <ULink URL="ftp://ftp.inr.ac.ru/ip-routing/iproute2-current.tar.gz">here</ULink>
490 for the latest version.
494 Some parts of iproute require you to have certain kernel options enabled. It
495 should also be noted that all releases of RedHat up to and including 6.2
496 come without most of the traffic control features in the default kernel.
500 RedHat 7.2 has everything in by default.
504 Also make sure that you have netlink support, should you choose to roll your
505 own kernel. Iproute2 needs it.
510 <Sect1 id="lartc.iproute2.explore">
511 <Title>Exploring your current configuration</Title>
514 This may come as a surprise, but iproute2 is already configured! The current
515 commands <command>ifconfig</command> and <command>route</command> are already using the advanced
516 syscalls, but mostly with very default (ie. boring) settings.
520 The <command>ip</command> tool is central, and we'll ask it to display our interfaces
525 <Title><command>ip</command> shows us our links</Title>
528 [ahu@home ahu]$ ip link list
529 1: lo: <LOOPBACK,UP> mtu 3924 qdisc noqueue
530 link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00
531 2: dummy: <BROADCAST,NOARP> mtu 1500 qdisc noop
532 link/ether 00:00:00:00:00:00 brd ff:ff:ff:ff:ff:ff
533 3: eth0: <BROADCAST,MULTICAST,PROMISC,UP> mtu 1400 qdisc pfifo_fast qlen 100
534 link/ether 48:54:e8:2a:47:16 brd ff:ff:ff:ff:ff:ff
535 4: eth1: <BROADCAST,MULTICAST,PROMISC,UP> mtu 1500 qdisc pfifo_fast qlen 100
536 link/ether 00:e0:4c:39:24:78 brd ff:ff:ff:ff:ff:ff
537 3764: ppp0: <POINTOPOINT,MULTICAST,NOARP,UP> mtu 1492 qdisc pfifo_fast qlen 10
543 Your mileage may vary, but this is what it shows on my NAT router at
544 home. I'll only explain part of the output as not everything is directly
549 We first see the loopback interface. While your computer may function
550 somewhat without one, I'd advise against it. The MTU size (Maximum Transfer
551 Unit) is 3924 octets, and it is not supposed to queue. Which makes sense
552 because the loopback interface is a figment of your kernel's imagination.
556 I'll skip the dummy interface for now, and it may not be present on your
557 computer. Then there are my two physical network interfaces, one at the side
558 of my cable modem, the other one serves my home ethernet segment.
559 Furthermore, we see a ppp0 interface.
563 Note the absence of IP addresses. iproute disconnects the concept of 'links'
564 and 'IP addresses'. With IP aliasing, the concept of 'the' IP address had
565 become quite irrelevant anyhow.
569 It does show us the MAC addresses though, the hardware identifier of our
576 <Title><command>ip</command> shows us our IP addresses</Title>
579 [ahu@home ahu]$ ip address show
580 1: lo: <LOOPBACK,UP> mtu 3924 qdisc noqueue
581 link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00
582 inet 127.0.0.1/8 brd 127.255.255.255 scope host lo
583 2: dummy: <BROADCAST,NOARP> mtu 1500 qdisc noop
584 link/ether 00:00:00:00:00:00 brd ff:ff:ff:ff:ff:ff
585 3: eth0: <BROADCAST,MULTICAST,PROMISC,UP> mtu 1400 qdisc pfifo_fast qlen 100
586 link/ether 48:54:e8:2a:47:16 brd ff:ff:ff:ff:ff:ff
587 inet 10.0.0.1/8 brd 10.255.255.255 scope global eth0
588 4: eth1: <BROADCAST,MULTICAST,PROMISC,UP> mtu 1500 qdisc pfifo_fast qlen 100
589 link/ether 00:e0:4c:39:24:78 brd ff:ff:ff:ff:ff:ff
590 3764: ppp0: <POINTOPOINT,MULTICAST,NOARP,UP> mtu 1492 qdisc pfifo_fast qlen 10
592 inet 212.64.94.251 peer 212.64.94.1/32 scope global ppp0
596 This contains more information. It shows all our addresses, and to which
597 cards they belong. 'inet' stands for Internet (IPv4). There are lots of other
598 address families, but these don't concern us right now.
602 Let's examine eth0 somewhat closer. It says that it is related to the inet
603 address '10.0.0.1/8'. What does this mean? The /8 stands for the number of
604 bits that are in the Network Address. There are 32 bits, so we have 24 bits
605 left that are part of our network. The first 8 bits of 10.0.0.1 correspond
606 to 10.0.0.0, our Network Address, and our netmask is 255.0.0.0.
610 The other bits are connected to this interface, so 10.250.3.13 is directly
611 available on eth0, as is 10.0.0.1 for example.
615 With ppp0, the same concept goes, though the numbers are different. Its
616 address is 212.64.94.251, without a subnet mask. This means that we have a
617 point-to-point connection and that every address, with the exception of
618 212.64.94.251, is remote. There is more information, however. It tells us
619 that on the other side of the link there is, yet again, only one address,
620 212.64.94.1. The /32 tells us that there are no 'network bits'.
624 It is absolutely vital that you grasp these concepts. Refer to the
625 documentation mentioned at the beginning of this HOWTO if you have trouble.
629 You may also note 'qdisc', which stands for Queueing Discipline. This will
630 become vital later on.
636 <Title><command>ip</command> shows us our routes</Title>
639 Well, we now know how to find 10.x.y.z addresses, and we are able to reach
640 212.64.94.1. This is not enough however, so we need instructions on how to
641 reach the world. The Internet is available via our ppp connection, and it
642 appears that 212.64.94.1 is willing to spread our packets around the
643 world, and deliver results back to us.
647 [ahu@home ahu]$ ip route show
648 212.64.94.1 dev ppp0 proto kernel scope link src 212.64.94.251
649 10.0.0.0/8 dev eth0 proto kernel scope link src 10.0.0.1
650 127.0.0.0/8 dev lo scope link
651 default via 212.64.94.1 dev ppp0
655 This is pretty much self explanatory. The first 4 lines of output explicitly
656 state what was already implied by <command>ip address show</command>, the last line
657 tells us that the rest of the world can be found via 212.64.94.1, our
658 default gateway. We can see that it is a gateway because of the word
659 via, which tells us that we need to send packets to 212.64.94.1, and that it
660 will take care of things.
664 For reference, this is what the old <command>route</command> utility shows us:
668 [ahu@home ahu]$ route -n
669 Kernel IP routing table
670 Destination Gateway Genmask Flags Metric Ref Use
672 212.64.94.1 0.0.0.0 255.255.255.255 UH 0 0 0 ppp0
673 10.0.0.0 0.0.0.0 255.0.0.0 U 0 0 0 eth0
674 127.0.0.0 0.0.0.0 255.0.0.0 U 0 0 0 lo
675 0.0.0.0 212.64.94.1 0.0.0.0 UG 0 0 0 ppp0
682 <Sect1 id="lartc.iproute2.arp">
686 ARP is the Address Resolution Protocol as described in
687 <ULink URL="http://www.faqs.org/rfcs/rfc826.html">RFC 826</ULink>.
688 ARP is used by a networked machine to resolve the hardware location/address of
689 another machine on the same
690 local network. Machines on the Internet are generally known by their names
692 addresses. This is how a machine on the foo.com network is able to communicate
693 with another machine which is on the bar.net network. An IP address, though,
694 cannot tell you the physical location of a machine. This is where ARP comes
699 Let's take a very simple example. Suppose I have a network composed of several
700 machines. Two of the machines which are currently on my network are foo
701 with an IP address of 10.0.0.1 and bar with an IP address of 10.0.0.2.
702 Now foo wants to ping bar to see that he is alive, but alas, foo has no idea
703 where bar is. So when foo decides to ping bar he will need to send
705 This ARP request is akin to foo shouting out on the network "Bar (10.0.0.2)!
706 Where are you?" As a result of this every machine on the network will hear
707 foo shouting, but only bar (10.0.0.2) will respond. Bar will then send an
708 ARP reply directly back to foo which is akin
710 "Foo (10.0.0.1) I am here at 00:60:94:E9:08:12." After this simple transaction
711 that's used to locate his friend on the network, foo is able to communicate
712 with bar until he (his arp cache) forgets where bar is (typically after
717 Now let's see how this works.
718 You can view your machines current arp/neighbor cache/table like so:
722 [root@espa041 /home/src/iputils]# ip neigh show
723 9.3.76.42 dev eth0 lladdr 00:60:08:3f:e9:f9 nud reachable
724 9.3.76.1 dev eth0 lladdr 00:06:29:21:73:c8 nud reachable
728 As you can see my machine espa041 (9.3.76.41) knows where to find espa042
730 espagate (9.3.76.1). Now let's add another machine to the arp cache.
734 [root@espa041 /home/paulsch/.gnome-desktop]# ping -c 1 espa043
735 PING espa043.austin.ibm.com (9.3.76.43) from 9.3.76.41 : 56(84) bytes of data.
736 64 bytes from 9.3.76.43: icmp_seq=0 ttl=255 time=0.9 ms
738 --- espa043.austin.ibm.com ping statistics ---
739 1 packets transmitted, 1 packets received, 0% packet loss
740 round-trip min/avg/max = 0.9/0.9/0.9 ms
742 [root@espa041 /home/src/iputils]# ip neigh show
743 9.3.76.43 dev eth0 lladdr 00:06:29:21:80:20 nud reachable
744 9.3.76.42 dev eth0 lladdr 00:60:08:3f:e9:f9 nud reachable
745 9.3.76.1 dev eth0 lladdr 00:06:29:21:73:c8 nud reachable
749 As a result of espa041 trying to contact espa043, espa043's hardware
750 address/location has now been added to the arp/neighbor cache.
751 So until the entry for
752 espa043 times out (as a result of no communication between the two) espa041
753 knows where to find espa043 and has no need to send an ARP request.
757 Now let's delete espa043 from our arp cache:
761 [root@espa041 /home/src/iputils]# ip neigh delete 9.3.76.43 dev eth0
762 [root@espa041 /home/src/iputils]# ip neigh show
763 9.3.76.43 dev eth0 nud failed
764 9.3.76.42 dev eth0 lladdr 00:60:08:3f:e9:f9 nud reachable
765 9.3.76.1 dev eth0 lladdr 00:06:29:21:73:c8 nud stale
769 Now espa041 has again forgotten where to find espa043 and will need to send
770 another ARP request the next time he needs to communicate with espa043.
771 You can also see from the above output that espagate (9.3.76.1) has been
772 changed to the "stale" state. This means that the location shown is still
773 valid, but it will have to be confirmed at the first transaction to that
781 <chapter id="lartc.rpdb">
782 <Title>Rules - routing policy database</Title>
785 If you have a large router, you may well cater for the needs of different
786 people, who should be served differently. The routing policy database allows
787 you to do this by having multiple sets of routing tables.
791 If you want to use this feature, make sure that your kernel is compiled with
792 the "IP: advanced router" and "IP: policy routing" features.
796 When the kernel needs to make a routing decision, it finds out which table
797 needs to be consulted. By default, there are three tables. The old 'route'
798 tool modifies the main and local tables, as does the ip tool (by default).
801 <Para>The default rules:
805 [ahu@home ahu]$ ip rule list
806 0: from all lookup local
807 32766: from all lookup main
808 32767: from all lookup default
812 This lists the priority of all rules. We see that all rules apply to all
813 packets ('from all'). We've seen the 'main' table before, it is output by
814 <userinput>ip route ls</userinput>, but the 'local' and 'default' table are new.
818 If we want to do fancy things, we generate rules which point to different
819 tables which allow us to override system wide routing rules.
823 For the exact semantics on what the kernel does when there are more matching
824 rules, see Alexey's ip-cref documentation.
827 <Sect1 id="lartc.rpdb.simple">
828 <Title>Simple source policy routing</Title>
831 Let's take a real example once again, I have 2 (actually 3, about time I
832 returned them) cable modems, connected to a Linux NAT ('masquerading')
833 router. People living here pay me to use the Internet. Suppose one of my
834 house mates only visits hotmail and wants to pay less. This is fine with me,
835 but they'll end up using the low-end cable modem.
839 The 'fast' cable modem is known as 212.64.94.251 and is a PPP link to
840 212.64.94.1. The 'slow' cable modem is known by various ip addresses,
841 212.64.78.148 in this example and is a link to 195.96.98.253.
844 <Para>The local table:
848 [ahu@home ahu]$ ip route list table local
849 broadcast 127.255.255.255 dev lo proto kernel scope link src 127.0.0.1
850 local 10.0.0.1 dev eth0 proto kernel scope host src 10.0.0.1
851 broadcast 10.0.0.0 dev eth0 proto kernel scope link src 10.0.0.1
852 local 212.64.94.251 dev ppp0 proto kernel scope host src 212.64.94.251
853 broadcast 10.255.255.255 dev eth0 proto kernel scope link src 10.0.0.1
854 broadcast 127.0.0.0 dev lo proto kernel scope link src 127.0.0.1
855 local 212.64.78.148 dev ppp2 proto kernel scope host src 212.64.78.148
856 local 127.0.0.1 dev lo proto kernel scope host src 127.0.0.1
857 local 127.0.0.0/8 dev lo proto kernel scope host src 127.0.0.1
861 Lots of obvious things, but things that need to be specified somewhere.
862 Well, here they are. The default table is empty.
865 <Para>Let's view the 'main' table:
869 [ahu@home ahu]$ ip route list table main
870 195.96.98.253 dev ppp2 proto kernel scope link src 212.64.78.148
871 212.64.94.1 dev ppp0 proto kernel scope link src 212.64.94.251
872 10.0.0.0/8 dev eth0 proto kernel scope link src 10.0.0.1
873 127.0.0.0/8 dev lo scope link
874 default via 212.64.94.1 dev ppp0
878 We now generate a new rule which we call 'John', for our hypothetical
879 house mate. Although we can work with pure numbers, it's far easier if we add
880 our tables to /etc/iproute2/rt_tables.
884 # echo 200 John >> /etc/iproute2/rt_tables
885 # ip rule add from 10.0.0.10 table John
887 0: from all lookup local
888 32765: from 10.0.0.10 lookup John
889 32766: from all lookup main
890 32767: from all lookup default
894 Now all that is left is to generate John's table, and flush the route cache:
898 # ip route add default via 195.96.98.253 dev ppp2 table John
899 # ip route flush cache
903 And we are done. It is left as an exercise for the reader to implement this
909 <sect1 id="lartc.rpdb.multiple-links">
910 <title>Routing for multiple uplinks/providers</title>
912 A common configuration is the following, in which there are two providers
913 that connect a local network (or even a single machine) to the big Internet.
919 +-------------+ Provider 1 +-------
921 ___/ \_ +------+-------+ +------------+ |
924 | Local network -----+ Linux router | | Internet
927 \___/ +------+-------+ +------------+ |
929 +-------------+ Provider 2 +-------
931 +------------+ \________
935 There are usually two questions given this setup.
937 <sect2><title>Split access</title>
939 The first is how to route answers to packets coming in over a
940 particular provider, say Provider 1, back out again over that same provider.
943 Let us first set some symbolical names. Let <command>$IF1</command> be the name of the
944 first interface (if1 in the picture above) and <command>$IF2</command> the name of the
945 second interface. Then let <command>$IP1</command> be the IP address associated with
946 <command>$IF1</command> and <command>$IP2</command> the IP address associated with
947 <command>$IF2</command>. Next, let <command>$P1</command> be the IP address of the gateway at
948 Provider 1, and <command>$P2</command> the IP address of the gateway at provider 2.
949 Finally, let <command>$P1_NET</command> be the IP network <command>$P1</command> is in,
950 and <command>$P2_NET</command> the IP network <command>$P2</command> is in.
953 One creates two additional routing tables, say <command>T1</command> and <command>T2</command>.
954 These are added in /etc/iproute2/rt_tables. Then you set up routing in
955 these tables as follows:
959 ip route add $P1_NET dev $IF1 src $IP1 table T1
960 ip route add default via $P1 table T1
961 ip route add $P2_NET dev $IF2 src $IP2 table T2
962 ip route add default via $P2 table T2
965 Nothing spectacular, just build a route to the gateway and build a
966 default route via that gateway, as you would do in the case of a single
967 upstream provider, but put the routes in a separate table per provider.
968 Note that the network route suffices, as it tells you how to find any host
969 in that network, which includes the gateway, as specified above.
972 Next you set up the main routing table. It is a good idea to route
973 things to the direct neighbour through the interface connected to that
974 neighbour. Note the `src' arguments, they make sure the right outgoing IP
978 ip route add $P1_NET dev $IF1 src $IP1
979 ip route add $P2_NET dev $IF2 src $IP2
982 Then, your preference for default route:
985 ip route add default via $P1
988 Next, you set up the routing rules. These actually choose what routing table
989 to route with. You want to make sure that you route out a given
990 interface if you already have the corresponding source address:
993 ip rule add from $IP1 table T1
994 ip rule add from $IP2 table T2
997 This set of commands makes sure all answers to traffic coming in on a
998 particular interface get answered from that interface.
1001 Now, this is just the very basic setup. It will work for all processes
1002 running on the router itself, and for the local network, if it is
1003 masqueraded. If it is not, then you either have IP space from both providers
1004 or you are going to want to masquerade to one of the two providers. In both
1005 cases you will want to add rules selecting which provider to route out from
1006 based on the IP address of the machine in the local network.
1009 <sect2><title>Load balancing</title>
1011 The second question is how to balance traffic going out over the two providers.
1012 This is actually not hard if you already have set up split access as above.
1015 Instead of choosing one of the two providers as your default route,
1016 you now set up the default route to be a multipath route. In the default
1017 kernel this will balance routes over the two providers. It is done
1018 as follows (once more building on the example in the section on
1022 ip route add default scope global nexthop via $P1 dev $IF1 weight 1 \
1023 nexthop via $P2 dev $IF2 weight 1
1026 This will balance the routes over both providers. The <command>weight</command>
1027 parameters can be tweaked to favor one provider over the other.
1030 Note that balancing will not be perfect, as it is route based, and routes
1031 are cached. This means that routes to often-used sites will always
1032 be over the same provider.
1035 Furthermore, if you really want to do this, you probably also want to look
1036 at Julian Anastasov's patches at <ulink
1037 url="http://www.linuxvirtualserver.org/~julian/#routes">http://www.linuxvirtualserver.org/~julian/#routes
1038 </ulink>, Julian's route patch page. They will make things nicer to work with.
1044 <chapter id="lartc.tunnel">
1045 <Title>GRE and other tunnels</Title>
1048 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).
1051 <Sect1 id="lartc.tunnel.remarks">
1052 <Title>A few general remarks about tunnels:</Title>
1055 Tunnels can be used to do some very unusual and very cool stuff. They can
1056 also make things go horribly wrong when you don't configure them right.
1057 Don't point your default route to a tunnel device unless you know
1058 <Emphasis>EXACTLY</Emphasis> what you are doing :-). Furthermore, tunneling increases
1059 overhead, because it needs an extra set of IP headers. Typically this is 20
1060 bytes per packet, so if the normal packet size (MTU) on a network is 1500
1061 bytes, a packet that is sent through a tunnel can only be 1480 bytes big.
1062 This is not necessarily a problem, but be sure to read up on IP packet
1063 fragmentation/reassembly when you plan to connect large networks with
1064 tunnels. Oh, and of course, the fastest way to dig a tunnel is to dig at
1070 <Sect1 id="lartc.tunnel.ip-ip">
1071 <Title>IP in IP tunneling</Title>
1074 This kind of tunneling has been available in Linux for a long time. It requires 2 kernel modules,
1075 ipip.o and new_tunnel.o.
1079 Let's say you have 3 networks: Internal networks A and B, and intermediate network C (or let's say, Internet).
1080 So we have network A:
1085 netmask 255.255.255.0
1089 <Para>The router has address 172.16.17.18 on network C.
1092 <Para>and network B:
1097 netmask 255.255.255.0
1101 <Para>The router has address 172.19.20.21 on network C.
1105 As far as network C is concerned, we assume that it will pass any packet sent
1106 from A to B and vice versa. You might even use the Internet for this.
1109 <Para>Here's what you do:
1112 <Para>First, make sure the modules are installed:
1120 <Para>Then, on the router of network A, you do the following:
1124 ifconfig tunl0 10.0.1.1 pointopoint 172.19.20.21
1125 route add -net 10.0.2.0 netmask 255.255.255.0 dev tunl0
1128 <Para>And on the router of network B:
1132 ifconfig tunl0 10.0.2.1 pointopoint 172.16.17.18
1133 route add -net 10.0.1.0 netmask 255.255.255.0 dev tunl0
1136 <Para>And if you're finished with your tunnel:
1143 <Para>Presto, you're done. You can't forward broadcast or IPv6 traffic through
1144 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.
1149 <Sect1 id="lartc.tunnel.gre">
1150 <Title>GRE tunneling</Title>
1153 GRE is a tunneling protocol that was originally developed by Cisco, and it
1154 can do a few more things than IP-in-IP tunneling. For example, you can also
1155 transport multicast traffic and IPv6 through a GRE tunnel.
1159 In Linux, you'll need the ip_gre.o module.
1163 <Title>IPv4 Tunneling</Title>
1166 Let's do IPv4 tunneling first:
1170 Let's say you have 3 networks: Internal networks A and B, and intermediate network C (or let's say, Internet).
1174 So we have network A:
1178 netmask 255.255.255.0
1182 The router has address 172.16.17.18 on network C.
1183 Let's call this network neta (ok, hardly original)
1191 netmask 255.255.255.0
1195 The router has address 172.19.20.21 on network C.
1196 Let's call this network netb (still not original)
1200 As far as network C is concerned, we assume that it will pass any packet sent
1201 from A to B and vice versa. How and why, we do not care.
1204 <Para>On the router of network A, you do the following:
1208 ip tunnel add netb mode gre remote 172.19.20.21 local 172.16.17.18 ttl 255
1210 ip addr add 10.0.1.1 dev netb
1211 ip route add 10.0.2.0/24 dev netb
1215 Let's discuss this for a bit. In line 1, we added a tunnel device, and
1216 called it netb (which is kind of obvious because that's where we want it to
1217 go). Furthermore we told it to use the GRE protocol (mode gre), that the
1218 remote address is 172.19.20.21 (the router at the other end), that our
1219 tunneling packets should originate from 172.16.17.18 (which allows your
1220 router to have several IP addresses on network C and let you decide which
1221 one to use for tunneling) and that the TTL field of the packet should be set
1226 The second line enables the device.
1230 In the third line we gave the newly born interface netb the address
1231 10.0.1.1. This is OK for smaller networks, but when you're starting up a
1232 mining expedition (LOTS of tunnels), you might want to consider using
1233 another IP range for tunneling interfaces (in this example, you could use
1238 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.
1242 But enough about this, let's go on with the router of network B.
1245 ip tunnel add neta mode gre remote 172.16.17.18 local 172.19.20.21 ttl 255
1247 ip addr add 10.0.2.1 dev neta
1248 ip route add 10.0.1.0/24 dev neta
1251 And when you want to remove the tunnel on router A:
1254 ip link set netb down
1258 Of course, you can replace netb with neta for router B.
1264 <Title>IPv6 Tunneling</Title>
1267 See Section 6 for a short bit about IPv6 Addresses.
1271 On with the tunnels.
1275 Let's assume that you have the following IPv6 network, and you want to connect it to 6bone, or a friend.
1281 Network 3ffe:406:5:1:5:a:2:1/96
1284 Your IPv4 address is 172.16.17.18, and the 6bone router has IPv4 address 172.22.23.24.
1290 ip tunnel add sixbone mode sit remote 172.22.23.24 local 172.16.17.18 ttl 255
1291 ip link set sixbone up
1292 ip addr add 3ffe:406:5:1:5:a:2:1/96 dev sixbone
1293 ip route add 3ffe::/15 dev sixbone
1299 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.
1303 GRE tunnels are currently the preferred type of tunneling. It's a standard that is also widely adopted outside the Linux community and therefore a Good Thing.
1310 <Sect1 id="lartc.tunnel.userland">
1311 <Title>Userland tunnels</Title>
1314 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.
1321 <chapter id="lartc.ipv6-tunnel">
1322 <Title>IPv6 tunneling with Cisco and/or 6bone</Title>
1325 By Marco Davids <marco@sara.nl>
1333 As far as I am concerned, this IPv6-IPv4 tunneling is not per definition
1334 GRE tunneling. You could tunnel IPv6 over IPv4 by means of GRE tunnel devices
1335 (GRE tunnels ANY to IPv4), but the device used here ("sit") only tunnels
1336 IPv6 over IPv4 and is therefore something different.
1339 <Sect1 id="lartc.tunnel-ipv6.addressing">
1340 <Title>IPv6 Tunneling</Title>
1343 This is another application of the tunneling capabilities of Linux. It is
1344 popular among the IPv6 early adopters, or pioneers if you like.
1345 The 'hands-on' example described below is certainly not the only way
1346 to do IPv6 tunneling. However, it is the method that is often used to tunnel
1347 between Linux and a Cisco IPv6 capable router and experience tells us that
1348 this is just the thing many people are after. Ten to one this applies to
1353 A short bit about IPv6 addresses:
1357 IPv6 addresses are, compared to IPv4 addresses, really big: 128 bits
1358 against 32 bits. And this provides us just with the thing we need: many, many
1359 IP-addresses: 340,282,266,920,938,463,463,374,607,431,768,211,465 to be
1360 precise. Apart from this, IPv6 (or IPng, for IP Next Generation) is supposed
1361 to provide for smaller routing tables on the Internet's backbone routers,
1362 simpler configuration of equipment, better security at the IP level and
1363 better support for QoS.
1367 An example: 2002:836b:9820:0000:0000:0000:836b:9886
1371 Writing down IPv6 addresses can be quite a burden. Therefore, to make
1372 life easier there are some rules:
1381 Don't use leading zeroes. Same as in IPv4.
1388 Use colons to separate every 16 bits or two bytes.
1395 When you have lots of consecutive zeroes,
1396 you can write this down as ::. You can only do this once in an
1397 address and only for quantities of 16 bits, though.
1406 The address 2002:836b:9820:0000:0000:0000:836b:9886 can be written down
1407 as 2002:836b:9820::836b:9886, which is somewhat friendlier.
1411 Another example, the address 3ffe:0000:0000:0000:0000:0020:34A1:F32C can be
1412 written down as 3ffe::20:34A1:F32C, which is a lot shorter.
1416 IPv6 is intended to be the successor of the current IPv4. Because it
1417 is relatively new technology, there is no worldwide native IPv6 network
1418 yet. To be able to move forward swiftly, the 6bone was introduced.
1422 Native IPv6 networks are connected to each other by encapsulating the IPv6
1423 protocol in IPv4 packets and sending them over the existing IPv4 infrastructure
1424 from one IPv6 site to another.
1428 That is precisely where the tunnel steps in.
1432 To be able to use IPv6, we should have a kernel that supports it. There
1433 are many good documents on how to achieve this. But it all comes down to
1440 Get yourself a recent Linux distribution, with suitable glibc.
1446 Then get yourself an up-to-date kernel source.
1452 If you are all set, then you can go ahead and compile an IPv6 capable
1459 Go to /usr/src/linux and type:
1471 Choose "Networking Options"
1477 Select "The IPv6 protocol", "IPv6: enable EUI-64 token format", "IPv6:
1478 disable provider based addresses"
1484 HINT: Don't go for the 'module' option. Often this won't work well.
1488 In other words, compile IPv6 as 'built-in' in your kernel.
1489 You can then save your config like usual and go ahead with compiling
1494 HINT: Before doing so, consider editing the Makefile:
1495 EXTRAVERSION = -x ; --> ; EXTRAVERSION = -x-IPv6
1499 There is a lot of good documentation about compiling and installing
1500 a kernel, however this document is about something else. If you run into
1501 problems at this stage, go and look for documentation about compiling a
1502 Linux kernel according to your own specifications.
1506 The file /usr/src/linux/README might be a good start.
1507 After you accomplished all this, and rebooted with your brand new kernel,
1508 you might want to issue an '/sbin/ifconfig -a' and notice the brand
1509 new 'sit0-device'. SIT stands for Simple Internet Transition. You may give
1510 yourself a compliment; you are now one major step closer to IP, the Next
1515 Now on to the next step. You want to connect your host, or maybe even
1516 your entire LAN to another IPv6 capable network. This might be the "6bone"
1517 that is setup especially for this particular purpose.
1521 Let's assume that you have the following IPv6 network: 3ffe:604:6:8::/64 and
1522 you want to connect it to 6bone, or a friend. Please note that the /64
1523 subnet notation works just like with regular IP addresses.
1527 Your IPv4 address is 145.100.24.181 and the 6bone router has IPv4 address
1532 # ip tunnel add sixbone mode sit remote 145.100.1.5 [local 145.100.24.181 ttl 255]
1533 # ip link set sixbone up
1534 # ip addr add 3FFE:604:6:7::2/126 dev sixbone
1535 # ip route add 3ffe::0/16 dev sixbone
1539 Let's discuss this. In the first line, we created a tunnel device called
1540 sixbone. We gave it mode sit (which is IPv6 in IPv4 tunneling) and told it
1541 where to go to (remote) and where to come from (local). TTL is set to
1546 Next, we made the device active (up). After that, we added our own network
1547 address, and set a route for 3ffe::/15 (which is currently all of 6bone)
1548 through the tunnel. If the particular machine you run this on is your IPv6
1549 gateway, then consider adding the following lines:
1553 # echo 1 >/proc/sys/net/ipv6/conf/all/forwarding
1554 # /usr/local/sbin/radvd
1558 The latter, radvd is -like zebra- a router advertisement daemon, to
1559 support IPv6's autoconfiguration features. Search for it with your favourite
1560 search-engine if you like.
1561 You can check things like this:
1565 # /sbin/ip -f inet6 addr
1569 If you happen to have radvd running on your IPv6 gateway and boot your
1570 IPv6 capable Linux on a machine on your local LAN, you would be able to
1571 enjoy the benefits of IPv6 autoconfiguration:
1575 # /sbin/ip -f inet6 addr
1576 1: lo: <LOOPBACK,UP> mtu 3924 qdisc noqueue inet6 ::1/128 scope host
1578 3: eth0: <BROADCAST,MULTICAST,UP> mtu 1500 qdisc pfifo_fast qlen 100
1579 inet6 3ffe:604:6:8:5054:4cff:fe01:e3d6/64 scope global dynamic
1580 valid_lft forever preferred_lft 604646sec inet6 fe80::5054:4cff:fe01:e3d6/10
1585 You could go ahead and configure your bind for IPv6 addresses. The A
1586 type has an equivalent for IPv6: AAAA. The in-addr.arpa's equivalent is:
1587 ip6.int. There's a lot of information available on this topic.
1591 There is an increasing number of IPv6-aware applications available,
1592 including secure shell, telnet, inetd, Mozilla the browser, Apache the
1593 webserver and a lot of others. But this is all outside the scope of this
1594 Routing document ;-)
1598 On the Cisco side the configuration would be something like this:
1603 description IPv6 tunnel
1605 no ip directed-broadcast
1606 ipv6 address 3FFE:604:6:7::1/126
1607 tunnel source Serial0
1608 tunnel destination 145.100.24.181
1611 ipv6 route 3FFE:604:6:8::/64 Tunnel1
1614 But if you don't have a Cisco at your disposal, try one of the many
1615 IPv6 tunnel brokers available on the Internet. They are willing to configure
1616 their Cisco with an extra tunnel for you. Mostly by means of a friendly
1617 web interface. Search for "ipv6 tunnel broker" on your favourite search engine.
1624 <chapter id="lartc.ipsec">
1625 <Title>IPSEC: secure IP over the Internet</Title>
1628 There are two kinds of IPSEC available for Linux these days. For 2.2
1629 and 2.4, there is FreeS/WAN, which was the first major implementation. They
1631 have <ULink URL="http://www.freeswan.org/">an official site</ulink> and <ulink url="http://www.freeswan.ca">
1632 an unofficial one</ulink> that is actually maintained. FreeS/WAN has traditionally not been merged with
1633 the mainline kernel for a number of reasons. Most often mentioned are 'political' issues with Americans
1634 working on crypto tainting its exportability. Furthermore, it does not integrate too well with the Linux kernel,
1635 leading it to be a bad candidate for actual merging.
1638 Additionally, <ulink
1639 url="http://www.edlug.ed.ac.uk/archive/Sep2002/msg00244.html">many</ulink> parties <ulink
1640 url="http://lists.freeswan.org/pipermail/design/2002-November/003901.html">have voiced
1641 worries</ulink> about the quality of the code. To setup FreeS/WAN, a lot of
1643 url="http://www.freeswan.ca/code/old/freeswan-Snapshot/doc/index.html">documentation</ulink>
1647 As of Linux 2.5.47, there is a native IPSEC implementation in the kernel. It was written by Alexey Kuznetsov and
1648 Dave Miller, inspired by the work of the USAGI IPv6 group. With its merge, James Morris' CrypoAPI also became
1649 part of the kernel - it does the actual crypting.
1652 This HOWTO will only document the 2.5+ version of IPSEC. FreeS/WAN is recommended for Linux 2.4 users for now, but be aware
1653 that its configuration wil differ from the native IPSEC.
1656 As of 2.5.49, IPSEC works without further patches.
1661 I've collected patches released by Alexey or Dave Miller <ulink
1662 url="http://ds9a.nl/ipsec">here</ulink>. Apply all of them to 2.5.48 before
1663 reporting problems! (as yet, there are none for 2.5.49). Crude userspace utilities are <ulink
1664 url="ftp://ftp.inr.ac.ru/ip-routing/iputils-ss021109-try.tar.bz2">
1665 here</ulink> (<ulink url="http://ds9a.nl/ipsec/setkey.tar.gz">pre-compiled
1666 binary & manpage</ulink>). Compiling these userspace utilities requires editing the Makefiles in there to point them at your
1667 2.5.x kernel. This situation is expected to improve rapidly however.
1670 When compiling your kernel, be sure to turn on 'PF_KEY', 'AH', 'ESP' and
1671 everything in the CryptoAPI! TCP_MSS netfilter target is currently broken,
1672 you need to turn it off.
1677 The author of this chapter is a complete IPSEC nitwit! If you find the inevitable mistakes, please email
1678 bert hubert <email>ahu@ds9a.nl</email>.
1683 First, we'll show how to manually setup secure communication between
1684 two hosts. A large part of this process can also be automated, but
1685 here we'll do it by hand so as to acquaint ourselves with what is going on
1689 Feel free to skip the following section if you are only interested
1690 in automatic keying but be aware that some understanding of manual keying is
1693 <sect1 id="lartc.ipsec.intro"><title>Intro with Manual Keying</title>
1695 IPSEC is a complicated subject. A lot of information is available online, this HOWTO will concentrate on getting you
1696 up and running and explaining the basic principles.
1701 Many iptables configurations drop IPSEC packets! To pass IPSEC, use: 'iptables -A xxx -p 50 -j ACCEPT' and 'iptables -A xxx -p 51 -j ACCEPT'
1706 IPSEC offers a secure version of the Internet Protocol. Security in this context means two different things: encryption and authentication.
1707 A naive vision of security offers only encryption but it can easily be shown that is insufficient - you may be communicating encyphered,
1708 but no guarantee is offered that the remote party is the one you expect it to be.
1711 IPSEC supports 'Encapsulated Security Payload' (ESP) for encryption and 'Authentication Header' (AH) for authenticating the remote partner.
1712 You can configure both of them, or decided to do only either.
1715 Both ESP and AH rely on security associations. A security association (SA) consists of a source, a destination and an instruction. A sample
1716 authentication SA may look like this:
1718 add 10.0.0.11 10.0.0.216 ah 15700 -A hmac-md5 "1234567890123456";
1720 This says 'traffic going from 10.0.0.11 to 10.0.0.216 that needs an AH can be signed using HMAC-MD5 using secret 1234567890123456'. This instruction
1721 is labelled with SPI ('Security Parameter Index') id '15700', more about that later.
1722 The interesting bit about SAs is that they are symmetrical. Both sides of a conversation share exactly the same SA, it is not mirrored on the
1723 other side. Do note however that there is no 'autoreverse' rule - this SA only describes a possible authentication from 10.0.0.11 to
1724 10.0.0.216. For two-way traffic, two SAs are needed.
1729 add 10.0.0.11 10.0.0.216 esp 15701 -E 3des-cbc "123456789012123456789012";
1731 This says 'traffic going from 10.0.0.11 to 10.0.0.216 that needs encryption can be encyphered using 3des-cbc with key 123456789012123456789012'. The
1735 So far, we've seen that SAs describe possible instructions, but do not in fact describe policy as to when these need to be used. In fact,
1736 there could be an arbitrary number of nearly identical SAs with only differing SPI ids. Incidentally, SPI stands for Security Parameter Index.
1737 To do actual crypto, we need to describe a policy. This policy can include things as 'use ipsec if available' or 'drop traffic unless we have ispec'.
1740 A typical simple Security Policy (SP) looks like this:
1742 spdadd 10.0.0.216 10.0.0.11 any -P out ipsec
1743 esp/transport//require
1744 ah/transport//require;
1746 If entered on host 10.0.0.216, this means that all traffic going out to 10.0.0.11 must be encrypted
1747 and be wrapped in an AH authenticating header. Note that this does not describe which SA is to be used,
1748 that is left as an exercise for the kernel to determine.
1751 In other words, a Security Policy specifies WHAT we want; a Security
1752 Association describes HOW we want it.
1755 Outgoing packets are labelled with the SA SPI ('the how') which the
1756 kernel used for encryption and authentication so the remote can
1757 lookup the corresponding verification and decryption instruction.
1761 What follows is a very simple configuration for talking from host 10.0.0.216 to 10.0.0.11 using
1762 encryption and authentication. Note that the reverse path is plaintext in this first version and that
1763 this configuration should not be deployed.
1769 add 10.0.0.216 10.0.0.11 ah 24500 -A hmac-md5 "1234567890123456";
1770 add 10.0.0.216 10.0.0.11 esp 24501 -E 3des-cbc "123456789012123456789012";
1772 spdadd 10.0.0.216 10.0.0.11 any -P out ipsec
1773 esp/transport//require
1774 ah/transport//require;
1778 On host 10.0.0.11, the same Security Associations, no Security Policy:
1781 add 10.0.0.216 10.0.0.11 ah 24500 -A hmac-md5 "1234567890123456";
1782 add 10.0.0.216 10.0.0.11 esp 24501 -E 3des-cbc "123456789012123456789012";
1786 With the above configuration in place (these files can be executed if 'setkey' is installed in /sbin),
1787 'ping 10.0.0.11' from 10.0.0.216 looks like this using tcpdump:
1789 22:37:52 10.0.0.216 > 10.0.0.11: AH(spi=0x00005fb4,seq=0xa): ESP(spi=0x00005fb5,seq=0xa) (DF)
1790 22:37:52 10.0.0.11 > 10.0.0.216: icmp: echo reply
1792 Note how the ping back from 10.0.0.11 is indeed plainly visible. The forward ping cannot be read by tcpdump
1793 of course, but it does show the Security Parameter Index of AH and ESP, which tells 10.0.0.11 how to
1794 verify the authenticity of our packet and how to decrypt it.
1797 A few things must be mentioned however. The configuration above is shown in a lot of IPSEC examples and it is very dangerous.
1798 The problem is that the above contains policy on how 10.0.0.216 should treat packets going to 10.0.0.11, and that it explains how 10.0.0.11
1799 should treat those packets but it does NOT instruct 10.0.0.11 to discard unauthenticated or unencrypted traffic!
1802 Anybody can now insert spoofed and completely unencrypted data and 10.0.0.11 will accept it. To remedy the above, we need an incoming
1803 Security Policy on 10.0.0.11, as follows:
1806 spdadd 10.0.0.216 10.0.0.11 any -P IN ipsec
1807 esp/transport//require
1808 ah/transport//require;
1810 This instructs 10.0.0.11 that any traffic coming to it from 10.0.0.216 is required to have valid ESP and AH.
1813 Now, to complete this configuration, we need return traffic to be encrypted and authenticated as well of course. The full configuration on
1821 add 10.0.0.11 10.0.0.216 ah 15700 -A hmac-md5 "1234567890123456";
1822 add 10.0.0.216 10.0.0.11 ah 24500 -A hmac-md5 "1234567890123456";
1825 add 10.0.0.11 10.0.0.216 esp 15701 -E 3des-cbc "123456789012123456789012";
1826 add 10.0.0.216 10.0.0.11 esp 24501 -E 3des-cbc "123456789012123456789012";
1828 spdadd 10.0.0.216 10.0.0.11 any -P out ipsec
1829 esp/transport//require
1830 ah/transport//require;
1832 spdadd 10.0.0.11 10.0.0.216 any -P in ipsec
1833 esp/transport//require
1834 ah/transport//require;
1846 add 10.0.0.11 10.0.0.216 ah 15700 -A hmac-md5 "1234567890123456";
1847 add 10.0.0.216 10.0.0.11 ah 24500 -A hmac-md5 "1234567890123456";
1850 add 10.0.0.11 10.0.0.216 esp 15701 -E 3des-cbc "123456789012123456789012";
1851 add 10.0.0.216 10.0.0.11 esp 24501 -E 3des-cbc "123456789012123456789012";
1854 spdadd 10.0.0.11 10.0.0.216 any -P out ipsec
1855 esp/transport//require
1856 ah/transport//require;
1858 spdadd 10.0.0.216 10.0.0.11 any -P in ipsec
1859 esp/transport//require
1860 ah/transport//require;
1865 Note that in this example we used identical keys for both directions of traffic. This is not in any way required however.
1868 To examine the configuration we just created, execute <command>setkey -D</command>, which shows the Security Associations or
1869 <command>setkey -DP</command> which shows the configured policies.
1872 <sect1 id="lartc.ipsec.automatic.keying"><title>Automatic keying</title>
1874 In the previous section, encryption was configured using simple shared secrets. In other words, to remain secure,
1875 we need to transfer our encryption configuration over a trusted channel. If we were to configure the remote host
1876 over telnet, any third party would know our shared secret and the setup would not be secure.
1879 Furthermore, because the secret is shared, it is not a secret. The remote can't do a lot with our secret, but we do
1880 need to make sure that we use a different secret for communicating with all our partners. This requires a large number of keys,
1881 if there are 10 parties, this needs at least 50 different secrets.
1884 Besides the symmetric key problem, there is also the need for key rollover. If a third party manages to sniff enough traffic,
1885 it may be in a position to reverse engineer the key. This is prevented by moving to a new key every once in a while but that is
1886 a process that needs to be automated.
1889 Another problem is that with manual keying as described above we exactly define the algorithms and key lengths used, something
1890 that requires a lot of coordination with the remote party. It is desireable to be able to have the ability to describe a
1891 broader key policy such as 'We can do 3DES and Blowfish with at least the following key lengths'.
1894 To solve these isses, IPSEC provides Internet Key Exchange to automatically exchange randomly generated keys which are
1895 transmitted using asymmetric encryption technology, according to negotiated algorithm details.
1898 The Linux 2.5 IPSEC implementation works with the KAME 'racoon' IKE
1899 daemon. As of 9 November, the racoon version in Alexey's iptools
1900 distribution can be compiled, although you may need to remove
1901 #include <net/route.h> in two files. Alternatively, I've supplied a
1902 <ulink url="http://ds9a.nl/ipsec/racoon.bz2">precompiled version</ulink>.
1907 IKE needs access to UDP port 500, be sure that iptables does
1912 <sect2 id="lartc.ipsec.keying.theory"><title>Theory</title>
1914 As explained before, automatic keying does a lot of the work
1915 for us. Specifically, it creates Security Associations on the fly. It does
1916 not however set policy for us, which is as it should be.
1919 So, to benefit from IKE, setup a policy, but do not supply any
1920 SAs. If the kernel discovers that there is an IPSEC policy, but no Security
1921 Association, it will notify the IKE daemon, which then goes to work on
1922 trying to negotiate one.
1925 Reiterating, a Security Policy specifies WHAT we want; a Security
1926 Association describes HOW we want it. Using automatic keying lets us get
1927 away with only specifying what we want.
1930 <sect2 id="lartc.ipsec.automatic.keying.example"><title>Example</title>
1932 Kame racoon comes with a grand host of options, most of which have
1933 very fine default values, so we don't need to touch them. As described
1934 above, the operator needs to define a Security Policy, but no Security
1935 Associations. We leave their negotiation to the IKE daemon.
1938 In this example, 10.0.0.11 and 10.0.0.216 are once again going to
1939 setup secure communications, but this time with help from racoon. For
1940 simplicity this configuration will be using pre-shared keys, the
1941 dreaded 'shared secrets'. X.509 certificates are discussed in a separate
1942 section, see <xref linkend="lartc.ipsec.x509">.
1945 going to stick to almost the default configuration, identical on both hosts:
1949 path pre_shared_key "/usr/local/etc/racoon/psk.txt";
1953 exchange_mode aggressive,main;
1955 situation identity_only;
1957 my_identifier address;
1959 lifetime time 2 min; # sec,min,hour
1961 proposal_check obey; # obey, strict or claim
1964 encryption_algorithm 3des;
1965 hash_algorithm sha1;
1966 authentication_method pre_shared_key;
1974 lifetime time 2 min;
1975 encryption_algorithm 3des ;
1976 authentication_algorithm hmac_sha1;
1977 compression_algorithm deflate ;
1982 Lots of settings - I think yet more can be removed to get closer to
1983 the default configuration. A few noteworthy things. We've configured two
1984 anonymous settings which hold for all remotes, making further configuration
1985 easy. There is no need for per-host stanzas here, unless we really want
1989 Furthermore, we've set it up such that we identify ourselves based
1990 on our IP address ('my_identifier address'), and declare that we can do
1991 3des, sha1, and that we will be using a pre-shared key, located in psk.txt.
1994 In psk.txt, we now setup two entries, which do differ on both hosts.
1997 10.0.0.216 password2
2003 Make sure these files are owned by root, and set to mode 0600,
2004 racoon will not trust their contents otherwise. Note that these files are
2005 mirrors from eachother.
2008 Now we are ready to setup our desired policy, which is simple
2009 enough. On host 10.0.0.216:
2015 spdadd 10.0.0.216 10.0.0.11 any -P out ipsec
2016 esp/transport//require;
2018 spdadd 10.0.0.11 10.0.0.216 any -P in ipsec
2019 esp/transport//require;
2027 spdadd 10.0.0.11 10.0.0.216 any -P out ipsec
2028 esp/transport//require;
2030 spdadd 10.0.0.216 10.0.0.11 any -P in ipsec
2031 esp/transport//require;
2033 Note how again these policies are mirrored.
2036 We are now ready to launch racoon! Once launched, the moment we try
2037 to telnet from 10.0.0.11 to 10.0.0.216, or the other way around, racoon
2038 will start negotiating:
2040 12:18:44: INFO: isakmp.c:1689:isakmp_post_acquire(): IPsec-SA
2041 request for 10.0.0.11 queued due to no phase1 found.
2042 12:18:44: INFO: isakmp.c:794:isakmp_ph1begin_i(): initiate new
2043 phase 1 negotiation: 10.0.0.216[500]<=>10.0.0.11[500]
2044 12:18:44: INFO: isakmp.c:799:isakmp_ph1begin_i(): begin Aggressive mode.
2045 12:18:44: INFO: vendorid.c:128:check_vendorid(): received Vendor ID:
2047 12:18:44: NOTIFY: oakley.c:2037:oakley_skeyid(): couldn't find
2048 the proper pskey, try to get one by the peer's address.
2049 12:18:44: INFO: isakmp.c:2417:log_ph1established(): ISAKMP-SA
2050 established 10.0.0.216[500]-10.0.0.11[500] spi:044d25dede78a4d1:ff01e5b4804f0680
2051 12:18:45: INFO: isakmp.c:938:isakmp_ph2begin_i(): initiate new phase 2
2052 negotiation: 10.0.0.216[0]<=>10.0.0.11[0]
2053 12:18:45: INFO: pfkey.c:1106:pk_recvupdate(): IPsec-SA established:
2054 ESP/Transport 10.0.0.11->10.0.0.216 spi=44556347(0x2a7e03b)
2055 12:18:45: INFO: pfkey.c:1318:pk_recvadd(): IPsec-SA established:
2056 ESP/Transport 10.0.0.216->10.0.0.11 spi=15863890(0xf21052)
2060 If we now run setkey -D, which shows the Security Associations, they
2063 10.0.0.216 10.0.0.11
2064 esp mode=transport spi=224162611(0x0d5c7333) reqid=0(0x00000000)
2065 E: 3des-cbc 5d421c1b d33b2a9f 4e9055e3 857db9fc 211d9c95 ebaead04
2066 A: hmac-sha1 c5537d66 f3c5d869 bd736ae2 08d22133 27f7aa99
2067 seq=0x00000000 replay=4 flags=0x00000000 state=mature
2068 created: Nov 11 12:28:45 2002 current: Nov 11 12:29:16 2002
2069 diff: 31(s) hard: 600(s) soft: 480(s)
2070 last: Nov 11 12:29:12 2002 hard: 0(s) soft: 0(s)
2071 current: 304(bytes) hard: 0(bytes) soft: 0(bytes)
2072 allocated: 3 hard: 0 soft: 0
2073 sadb_seq=1 pid=17112 refcnt=0
2074 10.0.0.11 10.0.0.216
2075 esp mode=transport spi=165123736(0x09d79698) reqid=0(0x00000000)
2076 E: 3des-cbc d7af8466 acd4f14c 872c5443 ec45a719 d4b3fde1 8d239d6a
2077 A: hmac-sha1 41ccc388 4568ac49 19e4e024 628e240c 141ffe2f
2078 seq=0x00000000 replay=4 flags=0x00000000 state=mature
2079 created: Nov 11 12:28:45 2002 current: Nov 11 12:29:16 2002
2080 diff: 31(s) hard: 600(s) soft: 480(s)
2081 last: hard: 0(s) soft: 0(s)
2082 current: 231(bytes) hard: 0(bytes) soft: 0(bytes)
2083 allocated: 2 hard: 0 soft: 0
2084 sadb_seq=0 pid=17112 refcnt=0
2086 As are the Security Policies we configured ourselves:
2088 10.0.0.11[any] 10.0.0.216[any] tcp
2090 esp/transport//require
2091 created:Nov 11 12:28:28 2002 lastused:Nov 11 12:29:12 2002
2092 lifetime:0(s) validtime:0(s)
2093 spid=3616 seq=5 pid=17134
2095 10.0.0.216[any] 10.0.0.11[any] tcp
2097 esp/transport//require
2098 created:Nov 11 12:28:28 2002 lastused:Nov 11 12:28:44 2002
2099 lifetime:0(s) validtime:0(s)
2100 spid=3609 seq=4 pid=17134
2103 <sect3><title>Problems and known defects</title>
2105 If this does not work, check that all configuration files
2106 are owned by root, and can only be read by root. To start racoon on the
2107 foreground, use '-F'. To force it to read a certain configuration file,
2108 instead of at the compiled location, use '-f'. For staggering amounts of
2109 detail, add a 'log debug;' statement to racoon.conf.
2112 <sect2 id="lartc.ipsec.x509"><title>Automatic keying using X.509 certificates</title>
2114 As mentioned before, the use of shared secrets is hard because they
2115 aren't easily shared and once shared, are no longer secret. Luckily, there
2116 is asymmetric encryption technology to help resolve this.
2119 If each IPSEC participant makes a public and a private key, secure
2120 communications can be setup by both parties publishing their public key, and
2124 Building a key is relatively easy, although it requires some work.
2125 The following is based on the 'openssl' tool.
2127 <sect3><title>Building an X.509 certificate for your host</title>
2129 OpenSSL has a lot of infrastructure for keys that may or may not be
2130 signed by certificate authorities. Right now, we need to circumvent all that
2131 infrastructure and practice some good old Snake Oil security, and do without
2132 a certificate authority.
2135 First we issue a 'certificate request' for our host, called
2138 $ openssl req -new -nodes -newkey rsa:1024 -sha1 -keyform PEM -keyout \
2139 laptop.private -outform PEM -out request.pem
2141 This asks us some questions:
2143 Country Name (2 letter code) [AU]:NL
2144 State or Province Name (full name) [Some-State]:.
2145 Locality Name (eg, city) []:Delft
2146 Organization Name (eg, company) [Internet Widgits Pty Ltd]:Linux Advanced
2147 Routing & Traffic Control
2148 Organizational Unit Name (eg, section) []:laptop
2149 Common Name (eg, YOUR name) []:bert hubert
2150 Email Address []:ahu@ds9a.nl
2152 Please enter the following 'extra' attributes
2153 to be sent with your certificate request
2154 A challenge password []:
2155 An optional company name []:
2157 It is left to your own discretion how completely you want to fill
2158 this out. You may or may not want to put your hostname in there, depending
2159 on your security needs. In this example, we have.
2162 We'll now 'self sign' this request:
2164 $ openssl x509 -req -in request.pem -signkey laptop.private -out \
2167 subject=/C=NL/L=Delft/O=Linux Advanced Routing & Traffic \
2168 Control/OU=laptop/CN=bert hubert/Email=ahu@ds9a.nl
2171 The 'request.pem' file can now be discarded.
2174 Repeat this procedure for all hosts you need a key for. You can
2175 distribute the '.public' file with impunity, but keep the '.private' one
2179 <sect3><title>Setting up and launching</title>
2181 Once we have a public and a private key for our hosts we can tell
2185 We return to our previous configuration and the two hosts, 10.0.0.11
2186 ('upstairs') and 10.0.0.216 ('laptop').
2189 To the <filename>racoon.conf</filename> file on 10.0.0.11, we add:
2191 path certificate "/usr/local/etc/racoon/certs";
2195 exchange_mode aggressive,main;
2196 my_identifier asn1dn;
2197 peers_identifier asn1dn;
2199 certificate_type x509 "upstairs.public" "upstairs.private";
2201 peers_certfile "laptop.public";
2203 encryption_algorithm 3des;
2204 hash_algorithm sha1;
2205 authentication_method rsasig;
2210 This tells racoon that certificates are to be found in
2211 <filename>/usr/local/etc/racoon/certs/</filename>. Furthermore, it contains
2212 configuration items specific for remote 10.0.0.216.
2215 The 'asn1dn' lines tell racoon that the identifier for both the
2216 local and remote ends are to be extracted from the public keys. This is the
2217 'subject=/C=NL/L=Delft/O=Linux Advanced Routing & Traffic
2218 Control/OU=laptop/CN=bert hubert/Email=ahu@ds9a.nl' output from above.
2221 The <command>certificate_type</command> line configures the local
2222 public and private key. The <command>peers_certfile</command> statement
2223 configures racoon to read the public key of the remote peer from the file
2224 <filename>laptop.public</filename>.
2227 The <command>proposal</command> stanza is unchanged from what we've
2228 seen earlier, with the exception that the
2229 <command>authentication_method</command> is now <command>rsasig</command>,
2230 indicating the use of RSA public/private keys for authentication.
2233 The addition to the configuration of 10.0.0.216 is nearly identical, except for the
2236 path certificate "/usr/local/etc/racoon/certs";
2240 exchange_mode aggressive,main;
2241 my_identifier asn1dn;
2242 peers_identifier asn1dn;
2244 certificate_type x509 "laptop.public" "laptop.private";
2246 peers_certfile "upstairs.public";
2249 encryption_algorithm 3des;
2250 hash_algorithm sha1;
2251 authentication_method rsasig;
2258 Now that we've added these statements to both hosts, we only need to
2259 move the key files in place. The 'upstairs' machine needs
2260 <filename>upstairs.private</filename>, <filename>upstairs.public</filename>,
2261 and <filename>laptop.public</filename> in
2262 <filename>/usr/local/etc/racoon/certs</filename>. Make sure that this
2263 directory is owned by root and has mode 0600 or racoon may refuse to read
2267 The 'laptop' machine needs
2268 <filename>laptop.private</filename>, <filename>laptop.public</filename>,
2269 and <filename>upstairs.public</filename> in
2270 <filename>/usr/local/etc/racoon/certs</filename>. In other words, each host
2271 needs its own public and private key and additionally, the public key of the
2275 Verify that a Security Policy is in place (execute the 'spdadd' lines in
2276 <xref linkend="lartc.ipsec.automatic.keying.example">). Then launch racoon and everything should
2279 <sect3><title>How to setup tunnels securely</title>
2281 To setup secure communications with a remote party, we must exchange
2282 public keys. While the public key does not need to be kept a secret, on the
2283 contrary, it is very important to be sure that it is in fact the unaltered
2284 key. In other words, you need to be certain there is no 'man in the middle'.
2287 To make this easy, OpenSSL provides the 'digest' command:
2289 $ openssl dgst upstairs.public
2290 MD5(upstairs.public)= 78a3bddafb4d681c1ca8ed4d23da4ff1
2294 Now all we need to do is verify if our remote partner sees the same
2295 digest. This might be done by meeting in real life or perhaps over the
2296 phone, making sure the number of the remote party was not in fact sent over
2297 the same email containing the key!
2300 Another way of doing this is the use of a Trusted Third Party which
2301 runs a Certificate Authority. This CA would then sign your key, which we've
2302 done ourselves above.
2309 <sect1 id="lartc.ipsec.tunnel"><title>IPSEC tunnels</title>
2311 So far, we've only seen IPSEC in so called 'transport' mode where both endpoints understand IPSEC directly. As this is often not
2312 the case, it may be necessary to have only routers understand IPSEC, and have them do the work for the hosts behind them.
2313 This is called 'tunnel mode'.
2316 Setting this up is a breeze. To tunnel all traffic to 130.161.0.0/16 from 10.0.0.216 via 10.0.0.11, we issue the following on
2323 add 10.0.0.216 10.0.0.11 esp 34501
2325 -E 3des-cbc "123456789012123456789012";
2327 spdadd 10.0.0.0/24 130.161.0.0/16 any -P out ipsec
2328 esp/tunnel/10.0.0.216-10.0.0.11/require;
2330 Note the '-m tunnel', it is vitally important! This first configures an ESP encryption SA between our tunnel endpoints,
2331 10.0.0.216 and 10.0.0.11.
2334 Next the actual tunnel is configured. It instructs the kernel to encrypt all traffic it has to route from 10.0.0.0/24 to
2335 130.161.0.0. Furthermore, this traffic then has to be shipped to 10.0.0.11.
2338 10.0.0.11 also needs some configuration:
2344 add 10.0.0.216 10.0.0.11 esp 34501
2346 -E 3des-cbc "123456789012123456789012";
2348 spdadd 10.0.0.0/24 130.161.0.0/16 any -P in ipsec
2349 esp/tunnel/10.0.0.216-10.0.0.11/require;
2351 Note that this is exactly identical, except for the change from '-P out' to '-P in'. As with earlier examples,
2352 we've now only configured traffic going one way. Completing the other half of the tunnel is left as an
2353 exercise for the reader.
2356 Another name for this setup is 'proxy ESP', which is somewhat clearer.
2361 The IPSEC tunnel needs to have IP Forwarding enabled in the kernel!
2366 <sect1 id="lartc.ipsec.interop"><title>IPSEC interoperation with other systems</title>
2370 <sect2 id="lartc.ipsec.interop.win32"><title>Windows</title>
2379 <chapter id="lartc.multicast">
2380 <Title>Multicast routing</Title>
2383 FIXME: Editor Vacancy!
2387 The Multicast-HOWTO is ancient (relatively-speaking) and may be inaccurate
2388 or misleading in places, for that reason.
2392 Before you can do any multicast routing, you need to configure the Linux
2393 kernel to support the type of multicast routing you want to do. This, in
2394 turn, requires you to decide what type of multicast routing you expect to
2395 be using. There are essentially four "common" types - DVMRP (the Multicast
2396 version of the RIP unicast protocol), MOSPF (the same, but for OSPF), PIM-SM
2397 ("Protocol Independent Multicasting - Sparse Mode", which assumes that users
2398 of any multicast group are spread out, rather than clumped) and PIM-DM (the
2399 same, but "Dense Mode", which assumes that there will be significant clumps
2400 of users of the same multicast group).
2404 In the Linux kernel, you will notice that these options don't appear. This is
2405 because the protocol itself is handled by a routing application, such as
2406 Zebra, mrouted, or pimd. However, you still have to have a good idea of which
2407 you're going to use, to select the right options in the kernel.
2411 For all multicast routing, you will definitely need to enable "multicasting"
2412 and "multicast routing". For DVMRP and MOSPF, this is sufficient. If you are
2413 going to use PIM, you must also enable PIMv1 or PIMv2, depending on whether
2414 the network you are connecting to uses version 1 or 2 of the PIM protocol.
2418 Once you have all that sorted out, and your new Linux kernel compiled, you
2419 will see that the IP protocols listed, at boot time, now include IGMP. This
2420 is a protocol for managing multicast groups. At the time of writing, Linux
2421 supports IGMP versions 1 and 2 only, although version 3 does exist and has
2422 been documented. This doesn't really affect us that much, as IGMPv3 is still
2423 new enough that the extra capabilities of IGMPv3 aren't going to be that
2424 much use. Because IGMP deals with groups, only the features present in the
2425 simplest version of IGMP over the entire group are going to be used. For the
2426 most part, that will be IGMPv2, although IGMPv1 is sill going to be
2431 So far, so good. We've enabled multicasting. Now, we have to tell the Linux
2432 kernel to actually do something with it, so we can start routing. This means
2433 adding the Multicast virtual network to the router table:
2437 ip route add 224.0.0.0/4 dev eth0
2441 (Assuming, of course, that you're multicasting over eth0! Substitute the
2442 device of your choice, for this.)
2446 Now, tell Linux to forward packets...
2450 echo 1 > /proc/sys/net/ipv4/ip_forward
2454 At this point, you may be wondering if this is ever going to do anything. So,
2455 to test our connection, we ping the default group, 224.0.0.1, to see if anyone
2456 is alive. All machines on your LAN with multicasting enabled <Emphasis>should</Emphasis>
2457 respond, but nothing else. You'll notice that none of the machines that
2458 respond have an IP address of 224.0.0.1. What a surprise! :) This is a group
2459 address (a "broadcast" to subscribers), and all members of the group will
2460 respond with their own address, not the group address.
2468 At this point, you're ready to do actual multicast routing. Well, assuming
2469 that you have two networks to route between.
2478 <chapter id="lartc.qdisc">
2479 <Title>Queueing Disciplines for Bandwidth Management</Title>
2482 Now, when I discovered this, it <Emphasis>really</Emphasis> blew me away. Linux 2.2/2.4
2483 comes with everything to manage bandwidth in ways comparable to high-end
2484 dedicated bandwidth management systems.
2488 Linux even goes far beyond what Frame and ATM provide.
2491 <Para>Just to prevent confusion, <command>tc</command> uses the following
2492 rules for bandwith specification:
2494 <literallayout class='monospaced'>
2495 mbps = 1024 kbps = 1024 * 1024 bps => byte/s
2496 mbit = 1024 kbit => kilo bit/s.
2497 mb = 1024 kb = 1024 * 1024 b => byte
2498 mbit = 1024 kbit => kilo bit.
2501 Internally, the number is stored in bps and b.
2504 <Para>But when <command>tc</command> prints the rate, it uses following :
2507 <literallayout class='monospaced'>
2508 1Mbit = 1024 Kbit = 1024 * 1024 bps => byte/s
2511 <Sect1 id="lartc.qdisc.explain">
2512 <Title>Queues and Queueing Disciplines explained</Title>
2515 With queueing we determine the way in which data is <Emphasis>SENT</Emphasis>.
2516 It is important to realise that we can only shape data that we transmit.
2520 With the way the Internet works, we have no direct control of what people
2521 send us. It's a bit like your (physical!) mailbox at home. There is no way
2522 you can influence the world to modify the amount of mail they send you,
2523 short of contacting everybody.
2527 However, the Internet is mostly based on TCP/IP which has a few features
2528 that help us. TCP/IP has no way of knowing the capacity of the network
2529 between two hosts, so it just starts sending data faster and faster ('slow
2530 start') and when packets start getting lost, because there is no room to
2531 send them, it will slow down. In fact it is a bit smarter than this, but
2532 more about that later.
2536 This is the equivalent of not reading half of your mail, and hoping that
2537 people will stop sending it to you. With the difference that it works for
2542 If you have a router and wish to prevent certain hosts within your network
2543 from downloading too fast, you need to do your shaping on the *inner* interface
2544 of your router, the one that sends data to your own computers.
2548 You also have to be sure you are controlling the bottleneck of the link.
2549 If you have a 100Mbit NIC and you have a router that has a 256kbit link,
2550 you have to make sure you are not sending more data than your router can
2551 handle. Otherwise, it will be the router who is controlling the link and
2552 shaping the available bandwith. We need to 'own the queue' so to speak, and
2553 be the slowest link in the chain. Luckily this is easily possible.
2558 <Sect1 id="lartc.qdisc.classless">
2559 <Title>Simple, classless Queueing Disciplines</Title>
2562 As said, with queueing disciplines, we change the way data is sent.
2563 Classless queueing disciplines are those that, by and large accept data and
2564 only reschedule, delay or drop it.
2568 These can be used to shape traffic for an entire interface, without any
2569 subdivisions. It is vital that you understand this part of queueing before
2570 we go on the the classful qdisc-containing-qdiscs!
2574 By far the most widely used discipline is the pfifo_fast qdisc - this is the
2575 default. This also explains why these advanced features are so robust. They
2576 are nothing more than 'just another queue'.
2580 Each of these queues has specific strengths and weaknesses. Not all of them
2581 may be as well tested.
2585 <Title>pfifo_fast</Title>
2588 This queue is, as the name says, First In, First Out, which means that no
2589 packet receives special treatment. At least, not quite. This queue has 3 so
2590 called 'bands'. Within each band, FIFO rules apply. However, as long as
2591 there are packets waiting in band 0, band 1 won't be processed. Same goes
2592 for band 1 and band 2.
2596 The kernel honors the so called Type of Service flag of packets, and takes
2597 care to insert 'minimum delay' packets in band 0.
2601 Do not confuse this classless simple qdisc with the classful PRIO one!
2602 Although they behave similarly, pfifo_fast is classless and you cannot add
2603 other qdiscs to it with the tc command.
2607 <Title>Parameters & usage</Title>
2610 You can't configure the pfifo_fast qdisc as it is the hardwired default.
2611 This is how it is configured by default:
2615 <Term>priomap</Term>
2618 Determines how packet priorities, as assigned by the kernel, map to bands.
2619 Mapping occurs based on the TOS octet of the packet, which looks like this:
2626 +-----+-----+-----+-----+-----+-----+-----+-----+
2628 | PRECEDENCE | TOS | MBZ |
2630 +-----+-----+-----+-----+-----+-----+-----+-----+
2636 The four TOS bits (the 'TOS field') are defined as:
2639 Binary Decimcal Meaning
2640 -----------------------------------------
2641 1000 8 Minimize delay (md)
2642 0100 4 Maximize throughput (mt)
2643 0010 2 Maximize reliability (mr)
2644 0001 1 Minimize monetary cost (mmc)
2645 0000 0 Normal Service
2651 As there is 1 bit to the right of these four bits, the actual value of the
2652 TOS field is double the value of the TOS bits. Tcpdump -v -v shows you the
2653 value of the entire TOS field, not just the four bits. It is the value you
2654 see in the first column of this table:
2660 TOS Bits Means Linux Priority Band
2661 ------------------------------------------------------------
2662 0x0 0 Normal Service 0 Best Effort 1
2663 0x2 1 Minimize Monetary Cost 1 Filler 2
2664 0x4 2 Maximize Reliability 0 Best Effort 1
2665 0x6 3 mmc+mr 0 Best Effort 1
2666 0x8 4 Maximize Throughput 2 Bulk 2
2667 0xa 5 mmc+mt 2 Bulk 2
2668 0xc 6 mr+mt 2 Bulk 2
2669 0xe 7 mmc+mr+mt 2 Bulk 2
2670 0x10 8 Minimize Delay 6 Interactive 0
2671 0x12 9 mmc+md 6 Interactive 0
2672 0x14 10 mr+md 6 Interactive 0
2673 0x16 11 mmc+mr+md 6 Interactive 0
2674 0x18 12 mt+md 4 Int. Bulk 1
2675 0x1a 13 mmc+mt+md 4 Int. Bulk 1
2676 0x1c 14 mr+mt+md 4 Int. Bulk 1
2677 0x1e 15 mmc+mr+mt+md 4 Int. Bulk 1
2683 Lots of numbers. The second column contains the value of the relevant four
2684 TOS bits, followed by their translated meaning. For example, 15 stands for a
2685 packet wanting Minimal Monetary Cost, Maximum Reliability, Maximum
2686 Throughput AND Minimum Delay. I would call this a 'Dutch Packet'.
2690 The fourth column lists the way the Linux kernel interprets the TOS bits, by
2691 showing to which Priority they are mapped.
2695 The last column shows the result of the default priomap. On the command line,
2696 the default priomap looks like this:
2699 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
2705 This means that priority 4, for example, gets mapped to band number 1. The
2706 priomap also allows you to list higher priorities (> 7) which do not
2707 correspond to TOS mappings, but which are set by other means.
2711 This table from RFC 1349 (read it for more details) tells you how
2712 applications might very well set their TOS bits:
2715 TELNET 1000 (minimize delay)
2717 Control 1000 (minimize delay)
2718 Data 0100 (maximize throughput)
2720 TFTP 1000 (minimize delay)
2723 Command phase 1000 (minimize delay)
2724 DATA phase 0100 (maximize throughput)
2727 UDP Query 1000 (minimize delay)
2729 Zone Transfer 0100 (maximize throughput)
2731 NNTP 0001 (minimize monetary cost)
2735 Requests 0000 (mostly)
2736 Responses <same as request> (mostly)
2742 <Term>txqueuelen</Term>
2745 The length of this queue is gleaned from the interface configuration, which
2746 you can see and set with ifconfig and ip. To set the queue length to 10,
2747 execute: ifconfig eth0 txqueuelen 10
2751 You can't set this parameter with tc!
2762 <Title>Token Bucket Filter</Title>
2765 The Token Bucket Filter (TBF) is a simple qdisc that only passes packets
2766 arriving at a rate which is not exceeding some administratively set rate, but
2767 with the possibility to allow short bursts in excess of this rate.
2771 TBF is very precise, network- and processor friendly. It should be your
2772 first choice if you simply want to slow an interface down!
2776 The TBF implementation consists of a buffer (bucket), constantly filled by
2777 some virtual pieces of information called tokens, at a specific rate (token
2778 rate). The most important parameter of the bucket is its size, that is the
2779 number of tokens it can store.
2783 Each arriving token collects one incoming data packet from the data queue
2784 and is then deleted from the bucket. Associating this algorithm
2785 with the two flows -- token and data, gives us three possible scenarios:
2794 The data arrives in TBF at a rate that's <Emphasis>equal</Emphasis> to the rate
2795 of incoming tokens. In this case each incoming packet has its matching token
2796 and passes the queue without delay.
2803 The data arrives in TBF at a rate that's <Emphasis>smaller</Emphasis> than the
2804 token rate. Only a part of the tokens are deleted at output of each data packet
2805 that's sent out the queue, so the tokens accumulate, up to the bucket size.
2806 The unused tokens can then be used to send data a a speed that's exceeding the
2807 standard token rate, in case short data bursts occur.
2814 The data arrives in TBF at a rate <Emphasis>bigger</Emphasis> than the token rate.
2815 This means that the bucket will soon be devoid of tokens, which causes the
2816 TBF to throttle itself for a while. This is called an 'overlimit situation'.
2817 If packets keep coming in, packets will start to get dropped.
2826 The last scenario is very important, because it allows to
2827 administratively shape the bandwidth available to data that's passing
2832 The accumulation of tokens allows a short burst of overlimit data to be
2833 still passed without loss, but any lasting overload will cause packets to be
2834 constantly delayed, and then dropped.
2838 Please note that in the actual implementation, tokens correspond to bytes,
2843 <Title>Parameters & usage</Title>
2846 Even though you will probably not need to change them, tbf has some knobs
2847 available. First the parameters that are always available:
2851 <Term>limit or latency</Term>
2854 Limit is the number of bytes that can be queued waiting for tokens to become
2855 available. You can also specify this the other way around by setting the
2856 latency parameter, which specifies the maximum amount of time a packet can
2857 sit in the TBF. The latter calculation takes into account the size of the
2858 bucket, the rate and possibly the peakrate (if set).
2862 <Term>burst/buffer/maxburst</Term>
2865 Size of the bucket, in bytes. This is the maximum amount of bytes that
2866 tokens can be available for instantaneously. In general, larger shaping
2867 rates require a larger buffer. For 10mbit/s on Intel, you need at least
2868 10kbyte buffer if you want to reach your configured rate!
2872 If your buffer is too small, packets may be dropped because more tokens
2873 arrive per timer tick than fit in your bucket.
2880 A zero-sized packet does not use zero bandwidth. For ethernet, no packet
2881 uses less than 64 bytes. The Minimum Packet Unit determines the minimal
2882 token usage for a packet.
2889 The speedknob. See remarks above about limits!
2896 If the bucket contains tokens and is allowed to empty, by default it does so
2897 at infinite speed. If this is unacceptable, use the following parameters:
2904 <Term>peakrate</Term>
2907 If tokens are available, and packets arrive, they are sent out immediately
2908 by default, at 'lightspeed' so to speak. That may not be what you want,
2909 especially if you have a large bucket.
2913 The peakrate can be used to specify how quickly the bucket is allowed to be
2914 depleted. If doing everything by the book, this is achieved by releasing a
2915 packet, and then wait just long enough, and release the next. We calculated
2916 our waits so we send just at peakrate.
2920 However, due to de default 10ms timer resolution of Unix, with 10.000 bits
2921 average packets, we are limited to 1mbit/s of peakrate!
2925 <Term>mtu/minburst</Term>
2928 The 1mbit/s peakrate is not very useful if your regular rate is more than
2929 that. A higher peakrate is possible by sending out more packets per
2930 timertick, which effectively means that we create a second bucket!
2934 This second bucket defaults to a single packet, which is not a bucket at
2939 To calculate the maximum possible peakrate, multiply the configured mtu by
2940 100 (or more correctly, HZ, which is 100 on Intel, 1024 on Alpha).
2949 <Title>Sample configuration</Title>
2952 A simple but *very* useful configuration is this:
2955 # tc qdisc add dev ppp0 root tbf rate 220kbit latency 50ms burst 1540
2961 Ok, why is this useful? If you have a networking device with a large queue,
2962 like a DSL modem or a cable modem, and you talk to it over a fast device,
2963 like over an ethernet interface, you will find that uploading absolutely
2964 destroys interactivity.
2968 This is because uploading will fill the queue in the modem, which is
2969 probably *huge* because this helps actually achieving good data throughput
2970 uploading. But this is not what you want, you want to have the queue not too
2971 big so interactivity remains and you can still do other stuff while sending
2976 The line above slows down sending to a rate that does not lead to a queue in
2977 the modem - the queue will be in Linux, where we can control it to a limited
2982 Change 220kbit to your uplink's *actual* speed, minus a few percent. If you
2983 have a really fast modem, raise 'burst' a bit.
2990 <Sect2 id="lartc.sfq">
2991 <Title>Stochastic Fairness Queueing</Title>
2994 Stochastic Fairness Queueing (SFQ) is a simple implementation of the fair
2995 queueing algorithms family. It's less accurate than others, but it also
2996 requires less calculations while being almost perfectly fair.
3000 The key word in SFQ is conversation (or flow), which mostly corresponds to a
3001 TCP session or a UDP stream. Traffic is divided into a pretty large number
3002 of FIFO queues, one for each conversation. Traffic is then sent in a round
3003 robin fashion, giving each session the chance to send data in turn.
3007 This leads to very fair behaviour and disallows any single conversation from
3008 drowning out the rest. SFQ is called 'Stochastic' because it doesn't really
3009 allocate a queue for each session, it has an algorithm which divides traffic
3010 over a limited number of queues using a hashing algorithm.
3014 Because of the hash, multiple sessions might end up in the same bucket, which
3015 would halve each session's chance of sending a packet, thus halving the
3016 effective speed available. To prevent this situation from becoming
3017 noticeable, SFQ changes its hashing algorithm quite often so that any two
3018 colliding sessions will only do so for a small number of seconds.
3022 It is important to note that SFQ is only useful in case your actual outgoing
3023 interface is really full! If it isn't then there will be no queue on your
3024 linux machine and hence no effect. Later on we will describe how to combine
3025 SFQ with other qdiscs to get a best-of-both worlds situation.
3029 Specifically, setting SFQ on the ethernet interface heading to your
3030 cable modem or DSL router is pointless without further shaping!
3034 <Title>Parameters & usage</Title>
3037 The SFQ is pretty much self tuning:
3041 <Term>perturb</Term>
3044 Reconfigure hashing once this many seconds. If unset, hash will never be
3045 reconfigured. Not recommended. 10 seconds is probably a good value.
3049 <Term>quantum</Term>
3052 Amount of bytes a stream is allowed to dequeue before the next queue gets a
3053 turn. Defaults to 1 maximum sized packet (MTU-sized). Do not set below the
3063 <Title>Sample configuration</Title>
3066 If you have a device which has identical link speed and actual available
3067 rate, like a phone modem, this configuration will help promote fairness:
3070 # tc qdisc add dev ppp0 root sfq perturb 10
3072 qdisc sfq 800c: dev ppp0 quantum 1514b limit 128p flows 128/1024 perturb 10sec
3073 Sent 4812 bytes 62 pkts (dropped 0, overlimits 0)
3079 The number 800c: is the automatically assigned handle number, limit means
3080 that 128 packets can wait in this queue. There are 1024 hashbuckets
3081 available for accounting, of which 128 can be active at a time (no more
3082 packets fit in the queue!) Once every 10 seconds, the hashes are
3092 <Sect1 id="lartc.qdisc.advice">
3093 <Title>Advice for when to use which queue</Title>
3096 Summarizing, these are the simple queues that actually manage traffic by
3097 reordering, slowing or dropping packets.
3101 The following tips may help in choosing which queue to use. It mentions some
3102 qdiscs described in the
3103 <citetitle><xref linkend="lartc.adv-qdisc"></citetitle> chapter.
3109 To purely slow down outgoing traffic, use the Token Bucket Filter. Works up
3110 to huge bandwidths, if you scale the bucket.
3116 If your link is truly full and you want to make sure that no single session
3117 can dominate your outgoing bandwidth, use Stochastical Fairness Queueing.
3123 If you have a big backbone and know what you are doing, consider Random
3124 Early Drop (see Advanced chapter).
3130 To 'shape' incoming traffic which you are not forwarding, use the Ingress
3131 Policer. Incoming shaping is called 'policing', by the way, not 'shaping'.
3137 If you *are* forwarding it, use a TBF on the interface you are forwarding
3138 the data to. Unless you want to shape traffic that may go out over several
3139 interfaces, in which case the only common factor is the incoming interface.
3140 In that case use the Ingress Policer.
3146 If you don't want to shape, but only want to see if your interface is so
3147 loaded that it has to queue, use the pfifo queue (not pfifo_fast). It lacks
3148 internal bands but does account the size of its backlog.
3153 Finally - you can also do <quote>social shaping</quote>.
3154 You may not always be able to use technology to achieve what you want.
3155 Users experience technical constraints as hostile.
3156 A kind word may also help with getting your bandwidth to be divided right!
3163 <Sect1 id="lartc.qdisc.terminology">
3164 <Title>Terminology</Title>
3167 To properly understand more complicated configurations it is necessary to
3168 explain a few concepts first. Because of the complexity and he relative
3169 youth of the subject, a lot of different words are used when people in fact
3170 mean the same thing.
3174 The following is loosely based on
3175 <filename>draft-ietf-diffserv-model-06.txt</filename>,
3176 <citetitle>An Informal Management Model for Diffserv Routers</citetitle>.
3177 It can currently be found at
3178 <ulink url="http://www.ietf.org/internet-drafts/draft-ietf-diffserv-model-06.txt">
3179 http://www.ietf.org/internet-drafts/draft-ietf-diffserv-model-06.txt
3184 Read it for the strict definitions of the terms used.
3188 <Term>Queueing Discipline</Term>
3191 An algorithm that manages the queue of a device, either incoming (ingress)
3192 or outgoing (egress).
3196 <Term>Classless qdisc</Term>
3199 A qdisc with no configurable internal subdivisions.
3203 <Term>Classful qdisc</Term>
3206 A classful qdisc contains multiple classes. Each of these classes contains a
3207 further qdisc, which may again be classful, but need not be. According to
3208 the strict definition, pfifo_fast *is* classful, because it contains three
3209 bands which are, in fact, classes. However, from the user's configuration
3210 perspective, it is classless as the classes can't be touched with the tc
3215 <Term>Classes</Term>
3218 A classful qdisc may have many classes, which each are internal to the
3219 qdisc. Each of these classes may contain a real qdisc.
3223 <Term>Classifier</Term>
3226 Each classful qdisc needs to determine to which class it needs to send a
3227 packet. This is done using the classifier.
3234 Classification can be performed using filters. A filter contains a number of
3235 conditions which if matched, make the filter match.
3239 <Term>Scheduling</Term>
3242 A qdisc may, with the help of a classifier, decide that some packets need to
3243 go out earlier than others. This process is called Scheduling, and is
3244 performed for example by the pfifo_fast qdisc mentioned earlier. Scheduling
3245 is also called 'reordering', but this is confusing.
3249 <Term>Shaping</Term>
3252 The process of delaying packets before they go out to make traffic confirm
3253 to a configured maximum rate. Shaping is performed on egress. Colloquially,
3254 dropping packets to slow traffic down is also often called Shaping.
3258 <Term>Policing</Term>
3261 Delaying or dropping packets in order to make traffic stay below a
3262 configured bandwidth. In Linux, policing can only drop a packet and not
3263 delay it - there is no 'ingress queue'.
3267 <Term>Work-Conserving</Term>
3270 A work-conserving qdisc always delivers a packet if one is available. In
3271 other words, it never delays a packet if the network adaptor is ready to
3272 send one (in the case of an egress qdisc).
3276 <Term>non-Work-Conserving</Term>
3279 Some queues, like for example the Token Bucket Filter, may need to hold on
3280 to a packet for a certain time in order to limit the bandwidth. This means
3281 that they sometimes refuse to give up a packet, even though they have one
3289 Now that we have our terminology straight, let's see where all these things
3299 +---------------+-----------------------------------------+
3301 | -------> IP Stack |
3306 | | / ----------> Forwarding -> |
3311 | | Egress /--qdisc2--\ |
3312 --->->Ingress Classifier ---qdisc3---- | ->
3313 | Qdisc \__qdisc4__/ |
3316 +----------------------------------------------------------+
3319 Thanks to Jamal Hadi Salim for this ASCII representation.
3323 The big block represents the kernel. The leftmost arrow represents traffic
3324 entering your machine from the network. It is then fed to the Ingress
3325 Qdisc which may apply Filters to a packet, and decide to drop it. This
3326 is called 'Policing'.
3330 This happens at a very early stage, before it has seen a lot of the kernel.
3331 It is therefore a very good place to drop traffic very early, without
3332 consuming a lot of CPU power.
3336 If the packet is allowed to continue, it may be destined for a local
3337 application, in which case it enters the IP stack in order to be processed,
3338 and handed over to a userspace program. The packet may also be forwarded
3339 without entering an application, in which case it is destined for egress.
3340 Userspace programs may also deliver data, which is then examined and
3341 forwarded to the Egress Classifier.
3345 There it is investigated and enqueued to any of a number of qdiscs. In the
3346 unconfigured default case, there is only one egress qdisc installed, the
3347 pfifo_fast, which always receives the packet. This is called 'enqueueing'.
3351 The packet now sits in the qdisc, waiting for the kernel to ask for
3352 it for transmission over the network interface. This is called 'dequeueing'.
3356 This picture also holds in case there is only one network adaptor - the
3357 arrows entering and leaving the kernel should not be taken too literally.
3358 Each network adaptor has both ingress and egress hooks.
3363 <Sect1 id="lartc.qdisc.classful">
3364 <Title>Classful Queueing Disciplines</Title>
3367 Classful qdiscs are very useful if you have different kinds of traffic which
3368 should have differing treatment. One of the classful qdiscs is called 'CBQ'
3369 , 'Class Based Queueing' and it is so widely mentioned that people identify
3370 queueing with classes solely with CBQ, but this is not the case.
3374 CBQ is merely the oldest kid on the block - and also the most complex one.
3375 It may not always do what you want. This may come as something of a shock
3376 to many who fell for the 'sendmail effect', which teaches us that any
3377 complex technology which doesn't come with documentation must be the best
3382 More about CBQ and its alternatives shortly.
3386 <Title>Flow within classful qdiscs & classes</Title>
3389 When traffic enters a classful qdisc, it needs to be sent to any of the
3390 classes within - it needs to be 'classified'. To determine what to do with a
3391 packet, the so called 'filters' are consulted. It is important to know that
3392 the filters are called from within a qdisc, and not the other way around!
3396 The filters attached to that qdisc then return with a decision, and the
3397 qdisc uses this to enqueue the packet into one of the classes. Each subclass
3398 may try other filters to see if further instructions apply. If not, the
3399 class enqueues the packet to the qdisc it contains.
3403 Besides containing other qdiscs, most classful qdiscs also perform shaping.
3404 This is useful to perform both packet scheduling (with SFQ, for example) and
3405 rate control. You need this in cases where you have a high speed
3406 interface (for example, ethernet) to a slower device (a cable modem).
3410 If you were only to run SFQ, nothing would happen, as packets enter &
3411 leave your router without delay: the output interface is far faster than
3412 your actual link speed. There is no queue to schedule then.
3418 <Title>The qdisc family: roots, handles, siblings and parents</Title>
3421 Each interface has one egress 'root qdisc', by default the earlier mentioned
3422 classless pfifo_fast queueing discipline. Each qdisc can be assigned a
3423 handle, which can be used by later configuration statements to refer to that
3424 qdisc. Besides an egress qdisc, an interface may also have an ingress, which
3425 polices traffic coming in.
3429 The handles of these qdiscs consist of two parts, a major number and a minor
3430 number. It is habitual to name the root qdisc '1:', which is equal to '1:0'.
3431 The minor number of a qdisc is always 0.
3435 Classes need to have the same major number as their parent.
3439 <Title>How filters are used to classify traffic </Title>
3442 Recapping, a typical hierarchy might look like this:
3459 But don't let this tree fool you! You should *not* imagine the kernel to be
3460 at the apex of the tree and the network below, that is just not the case.
3461 Packets get enqueued and dequeued at the root qdisc, which is the only thing
3462 the kernel talks to.
3466 A packet might get classified in a chain like this:
3470 1: -> 1:1 -> 12: -> 12:2
3474 The packet now resides in a queue in a qdisc attached to class 12:2. In this
3475 example, a filter was attached to each 'node' in the tree, each choosing a
3476 branch to take next. This can make sense. However, this is also possible:
3484 In this case, a filter attached to the root decided to send the packet
3491 <Title>How packets are dequeued to the hardware</Title>
3494 When the kernel decides that it needs to extract packets to send to the
3495 interface, the root qdisc 1: gets a dequeue request, which is passed to
3496 1:1, which is in turn passed to 10:, 11: and 12:, which each query their
3497 siblings, and try to dequeue() from them. In this case, the kernel needs to
3498 walk the entire tree, because only 12:2 contains a packet.
3502 In short, nested classes ONLY talk to their parent qdiscs, never to an
3503 interface. Only the root qdisc gets dequeued by the kernel!
3507 The upshot of this is that classes never get dequeued faster than their
3508 parents allow. And this is exactly what we want: this way we can have SFQ in
3509 an inner class, which doesn't do any shaping, only scheduling, and have a
3510 shaping outer qdisc, which does the shaping.
3518 <Title>The PRIO qdisc</Title>
3521 The PRIO qdisc doesn't actually shape, it only subdivides traffic based on
3522 how you configured your filters. You can consider the PRIO qdisc a kind
3523 of pfifo_fast on steroids, whereby each band is a separate class instead of
3528 When a packet is enqueued to the PRIO qdisc, a class is chosen based on the
3529 filter commands you gave. By default, three classes are created. These
3530 classes by default contain pure FIFO qdiscs with no internal
3531 structure, but you can replace these by any qdisc you have available.
3535 Whenever a packet needs to be dequeued, class :1 is tried first. Higher
3536 classes are only used if lower bands all did not give up a packet.
3540 This qdisc is very useful in case you want to prioritize certain kinds of
3541 traffic without using only TOS-flags but using all the power of the tc
3542 filters. It can also contain more all qdiscs, whereas pfifo_fast is limited
3543 to simple fifo qdiscs.
3547 Because it doesn't actually shape, the same warning as for SFQ holds: either
3548 use it only if your physical link is really full or wrap it inside a
3549 classful qdisc that does shape. The last holds for almost all cable modems
3554 In formal words, the PRIO qdisc is a Work-Conserving scheduler.
3558 <Title>PRIO parameters & usage</Title>
3561 The following parameters are recognized by tc:
3568 Number of bands to create. Each band is in fact a class. If you change this
3569 number, you must also change:
3573 <Term>priomap</Term>
3576 If you do not provide tc filters to classify traffic, the PRIO qdisc looks
3577 at the TC_PRIO priority to decide how to enqueue traffic.
3581 This works just like with the pfifo_fast qdisc mentioned earlier, see there
3586 The bands are classes, and are called major:1 to major:3 by default, so if
3587 your PRIO qdisc is called 12:, tc filter traffic to 12:1 to grant it more
3592 Reiterating, band 0 goes to minor number 1! Band 1 to minor number 2, etc.
3598 <Title>Sample configuration</Title>
3601 We will create this tree:
3616 Bulk traffic will go to 30:, interactive traffic to 20: or 10:.
3623 # tc qdisc add dev eth0 root handle 1: prio
3624 ## This *instantly* creates classes 1:1, 1:2, 1:3
3626 # tc qdisc add dev eth0 parent 1:1 handle 10: sfq
3627 # tc qdisc add dev eth0 parent 1:2 handle 20: tbf rate 20kbit buffer 1600 limit 3000
3628 # tc qdisc add dev eth0 parent 1:3 handle 30: sfq
3634 Now let's see what we created:
3637 # tc -s qdisc ls dev eth0
3638 qdisc sfq 30: quantum 1514b
3639 Sent 0 bytes 0 pkts (dropped 0, overlimits 0)
3641 qdisc tbf 20: rate 20Kbit burst 1599b lat 667.6ms
3642 Sent 0 bytes 0 pkts (dropped 0, overlimits 0)
3644 qdisc sfq 10: quantum 1514b
3645 Sent 132 bytes 2 pkts (dropped 0, overlimits 0)
3647 qdisc prio 1: bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1
3648 Sent 174 bytes 3 pkts (dropped 0, overlimits 0)
3651 As you can see, band 0 has already had some traffic, and one packet was sent
3652 while running this command!
3656 We now do some bulk data transfer with a tool that properly sets TOS flags,
3657 and take another look:
3660 # scp tc ahu@10.0.0.11:./
3661 ahu@10.0.0.11's password:
3662 tc 100% |*****************************| 353 KB 00:00
3663 # tc -s qdisc ls dev eth0
3664 qdisc sfq 30: quantum 1514b
3665 Sent 384228 bytes 274 pkts (dropped 0, overlimits 0)
3667 qdisc tbf 20: rate 20Kbit burst 1599b lat 667.6ms
3668 Sent 2640 bytes 20 pkts (dropped 0, overlimits 0)
3670 qdisc sfq 10: quantum 1514b
3671 Sent 2230 bytes 31 pkts (dropped 0, overlimits 0)
3673 qdisc prio 1: bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1
3674 Sent 389140 bytes 326 pkts (dropped 0, overlimits 0)
3677 As you can see, all traffic went to handle 30:, which is the lowest priority
3678 band, just as intended. Now to verify that interactive traffic goes to
3679 higher bands, we create some interactive traffic:
3685 # tc -s qdisc ls dev eth0
3686 qdisc sfq 30: quantum 1514b
3687 Sent 384228 bytes 274 pkts (dropped 0, overlimits 0)
3689 qdisc tbf 20: rate 20Kbit burst 1599b lat 667.6ms
3690 Sent 2640 bytes 20 pkts (dropped 0, overlimits 0)
3692 qdisc sfq 10: quantum 1514b
3693 Sent 14926 bytes 193 pkts (dropped 0, overlimits 0)
3695 qdisc prio 1: bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1
3696 Sent 401836 bytes 488 pkts (dropped 0, overlimits 0)
3702 It worked - all additional traffic has gone to 10:, which is our highest
3703 priority qdisc. No traffic was sent to the lowest priority, which previously
3704 received our entire scp.
3712 <Title>The famous CBQ qdisc</Title>
3715 As said before, CBQ is the most complex qdisc available, the most hyped, the
3716 least understood, and probably the trickiest one to get right. This is not
3717 because the authors are evil or incompetent, far from it, it's just that the
3718 CBQ algorithm isn't all that precise and doesn't really match the way Linux
3723 Besides being classful, CBQ is also a shaper and it is in that aspect that
3724 it really doesn't work very well. It should work like this. If you try to
3725 shape a 10mbit/s connection to 1mbit/s, the link should be idle 90% of the
3726 time. If it isn't, we need to throttle so that it IS idle 90% of the time.
3730 This is pretty hard to measure, so CBQ instead derives the idle time from
3731 the number of microseconds that elapse between requests from the hardware
3732 layer for more data. Combined, this can be used to approximate how full or
3737 This is rather circumspect and doesn't always arrive at proper results. For
3738 example, what if the actual link speed of an interface that is not really
3739 able to transmit the full 100mbit/s of data, perhaps because of a badly
3740 implemented driver? A PCMCIA network card will also never achieve 100mbit/s
3741 because of the way the bus is designed - again, how do we calculate the idle
3746 It gets even worse if we consider not-quite-real network devices like PPP
3747 over Ethernet or PPTP over TCP/IP. The effective bandwidth in that case is
3748 probably determined by the efficiency of pipes to userspace - which is huge.
3752 People who have done measurements discover that CBQ is not always very
3753 accurate and sometimes completely misses the mark.
3757 In many circumstances however it works well. With the documentation provided
3758 here, you should be able to configure it to work well in most cases.
3762 <Title>CBQ shaping in detail</Title>
3765 As said before, CBQ works by making sure that the link is idle just long
3766 enough to bring down the real bandwidth to the configured rate. To do so, it
3767 calculates the time that should pass between average packets.
3771 During operations, the effective idletime is measured using an exponential
3772 weighted moving average (EWMA), which considers recent packets to be
3773 exponentially more important than past ones. The UNIX loadaverage is
3774 calculated in the same way.
3778 The calculated idle time is subtracted from the EWMA measured one, the
3779 resulting number is called 'avgidle'. A perfectly loaded link has an avgidle
3780 of zero: packets arrive exactly once every calculated interval.
3784 An overloaded link has a negative avgidle and if it gets too negative, CBQ
3785 shuts down for a while and is then 'overlimit'.
3789 Conversely, an idle link might amass a huge avgidle, which would then allow
3790 infinite bandwidths after a few hours of silence. To prevent this, avgidle is
3795 If overlimit, in theory, the CBQ could throttle itself for exactly the
3796 amount of time that was calculated to pass between packets, and then pass
3797 one packet, and throttle again. But see the 'minburst' parameter below.
3801 These are parameters you can specify in order to configure shaping:
3808 Average size of a packet, measured in bytes. Needed for calculating maxidle,
3809 which is derived from maxburst, which is specified in packets.
3813 <Term>bandwidth</Term>
3816 The physical bandwidth of your device, needed for idle time
3824 The time a packet takes to be transmitted over a device may grow in steps,
3825 based on the packet size. An 800 and an 806 size packet may take just as long
3826 to send, for example - this sets the granularity. Most often set to '8'.
3827 Must be an integral power of two.
3831 <Term>maxburst</Term>
3834 This number of packets is used to calculate maxidle so that when avgidle is
3835 at maxidle, this number of average packets can be burst before avgidle drops
3836 to 0. Set it higher to be more tolerant of bursts. You can't set maxidle
3837 directly, only via this parameter.
3841 <Term>minburst</Term>
3844 As mentioned before, CBQ needs to throttle in case of overlimit. The ideal
3845 solution is to do so for exactly the calculated idle time, and pass 1
3846 packet. However, Unix kernels generally have a hard time scheduling events
3847 shorter than 10ms, so it is better to throttle for a longer period, and then
3848 pass minburst packets in one go, and then sleep minburst times longer.
3852 The time to wait is called the offtime. Higher values of minburst lead to
3853 more accurate shaping in the long term, but to bigger bursts at millisecond
3858 <Term>minidle</Term>
3861 If avgidle is below 0, we are overlimits and need to wait until avgidle will
3862 be big enough to send one packet. To prevent a sudden burst from shutting
3863 down the link for a prolonged period of time, avgidle is reset to minidle if
3868 Minidle is specified in negative microseconds, so 10 means that avgidle is
3876 Minimum packet size - needed because even a zero size packet is padded
3877 to 64 bytes on ethernet, and so takes a certain time to transmit. CBQ needs
3878 to know this to accurately calculate the idle time.
3885 Desired rate of traffic leaving this qdisc - this is the 'speed knob'!
3892 Internally, CBQ has a lot of fine tuning. For example, classes which are
3893 known not to have data enqueued to them aren't queried. Overlimit classes
3894 are penalized by lowering their effective priority. All very smart &
3901 <Title>CBQ classful behaviour</Title>
3904 Besides shaping, using the aforementioned idletime approximations, CBQ also
3905 acts like the PRIO queue in the sense that classes can have differing
3906 priorities and that lower priority numbers will be polled before the higher
3911 Each time a packet is requested by the hardware layer to be sent out to the
3912 network, a weighted round robin process ('WRR') starts, beginning with the
3913 lower priority classes.
3917 These are then grouped and queried if they have data available. If so, it is
3918 returned. After a class has been allowed to dequeue a number of bytes, the
3919 next class within that priority is tried.
3923 The following parameters control the WRR process:
3930 When the outer CBQ is asked for a packet to send out on the interface, it
3931 will try all inner qdiscs (in the classes) in turn, in order of
3932 the 'priority' parameter. Each time a class gets its turn, it can only send out
3933 a limited amount of data. 'Allot' is the base unit of this amount. See
3934 the 'weight' parameter for more information.
3941 The CBQ can also act like the PRIO device. Inner classes with lower priority
3942 are tried first and as long as they have traffic, other classes are not
3950 Weight helps in the Weighted Round Robin process. Each class gets a chance
3951 to send in turn. If you have classes with significantly more bandwidth than
3952 other classes, it makes sense to allow them to send more data in one round
3957 A CBQ adds up all weights under a class, and normalizes them, so you can use
3958 arbitrary numbers: only the ratios are important. People have been
3959 using 'rate/10' as a rule of thumb and it appears to work well. The renormalized
3960 weight is multiplied by the 'allot' parameter to determine how much data can
3961 be sent in one round.
3968 Please note that all classes within an CBQ hierarchy need to share the same
3975 <Title>CBQ parameters that determine link sharing & borrowing</Title>
3978 Besides purely limiting certain kinds of traffic, it is also possible to
3979 specify which classes can borrow capacity from other classes or, conversely,
3987 <Term>Isolated/sharing</Term>
3990 A class that is configured with 'isolated' will not lend out bandwidth to
3991 sibling classes. Use this if you have competing or mutually-unfriendly
3992 agencies on your link who do want to give each other freebies.
3996 The control program tc also knows about 'sharing', which is the reverse
4001 <Term>bounded/borrow</Term>
4004 A class can also be 'bounded', which means that it will not try to borrow
4005 bandwidth from sibling classes. tc also knows about 'borrow', which is the
4006 reverse of 'bounded'.
4010 A typical situation might be where you have two agencies on your link which
4011 are both 'isolated' and 'bounded', which means that they are really limited
4012 to their assigned rate, and also won't allow each other to borrow.
4016 Within such an agency class, there might be other classes which are allowed
4023 <Title>Sample configuration</Title>
4026 This configuration limits webserver traffic to 5mbit and SMTP traffic to 3
4027 mbit. Together, they may not get more than 6mbit. We have a 100mbit NIC and
4028 the classes may borrow bandwidth from each other.
4031 # tc qdisc add dev eth0 root handle 1:0 cbq bandwidth 100Mbit \
4033 # tc class add dev eth0 parent 1:0 classid 1:1 cbq bandwidth 100Mbit \
4034 rate 6Mbit weight 0.6Mbit prio 8 allot 1514 cell 8 maxburst 20 \
4038 This part installs the root and the customary 1:0 class. The 1:1 class is
4039 bounded, so the total bandwidth can't exceed 6mbit.
4043 As said before, CBQ requires a *lot* of knobs. All parameters are explained
4044 above, however. The corresponding HTB configuration is lots simpler.
4050 # tc class add dev eth0 parent 1:1 classid 1:3 cbq bandwidth 100Mbit \
4051 rate 5Mbit weight 0.5Mbit prio 5 allot 1514 cell 8 maxburst 20 \
4053 # tc class add dev eth0 parent 1:1 classid 1:4 cbq bandwidth 100Mbit \
4054 rate 3Mbit weight 0.3Mbit prio 5 allot 1514 cell 8 maxburst 20 \
4061 These are our two classes. Note how we scale the weight with the configured
4062 rate. Both classes are not bounded, but they are connected to class 1:1
4063 which is bounded. So the sum of bandwith of the 2 classes will never be
4064 more than 6mbit. The classids need to be within the same major number as
4065 the parent CBQ, by the way!
4071 # tc qdisc add dev eth0 parent 1:3 handle 30: sfq
4072 # tc qdisc add dev eth0 parent 1:4 handle 40: sfq
4078 Both classes have a FIFO qdisc by default. But we replaced these with an SFQ
4079 queue so each flow of data is treated equally.
4082 # tc filter add dev eth0 parent 1:0 protocol ip prio 1 u32 match ip \
4083 sport 80 0xffff flowid 1:3
4084 # tc filter add dev eth0 parent 1:0 protocol ip prio 1 u32 match ip \
4085 sport 25 0xffff flowid 1:4
4091 These commands, attached directly to the root, send traffic to the right
4096 Note that we use 'tc class add' to CREATE classes within a qdisc, but that
4097 we use 'tc qdisc add' to actually add qdiscs to these classes.
4101 You may wonder what happens to traffic that is not classified by any of the
4102 two rules. It appears that in this case, data will then be processed within
4103 1:0, and be unlimited.
4107 If SMTP+web together try to exceed the set limit of 6mbit/s, bandwidth will
4108 be divided according to the weight parameter, giving 5/8 of traffic to the
4109 webserver and 3/8 to the mail server.
4113 With this configuration you can also say that webserver traffic will always
4114 get at minimum 5/8 * 6 mbit = 3.75 mbit.
4120 <Title>Other CBQ parameters: split & defmap</Title>
4123 As said before, a classful qdisc needs to call filters to determine
4124 which class a packet will be enqueued to.
4128 Besides calling the filter, CBQ offers other options, defmap & split.
4129 This is pretty complicated to understand, and it is not vital. But as this
4130 is the only known place where defmap & split are properly explained, I'm
4135 As you will often want to filter on the Type of Service field only, a special
4136 syntax is provided. Whenever the CBQ needs to figure out where a packet
4137 needs to be enqueued, it checks if this node is a 'split node'. If so, one
4138 of the sub-qdiscs has indicated that it wishes to receive all packets with
4139 a certain configured priority, as might be derived from the TOS field, or
4140 socket options set by applications.
4144 The packets' priority bits are or-ed with the defmap field to see if a match
4145 exists. In other words, this is a short-hand way of creating a very fast
4146 filter, which only matches certain priorities. A defmap of ff (hex) will
4147 match everything, a map of 0 nothing. A sample configuration may help make
4154 # tc qdisc add dev eth1 root handle 1: cbq bandwidth 10Mbit allot 1514 \
4155 cell 8 avpkt 1000 mpu 64
4157 # tc class add dev eth1 parent 1:0 classid 1:1 cbq bandwidth 10Mbit \
4158 rate 10Mbit allot 1514 cell 8 weight 1Mbit prio 8 maxburst 20 \
4162 Standard CBQ preamble. I never get used to the sheer amount of numbers
4167 Defmap refers to TC_PRIO bits, which are defined as follows:
4173 TC_PRIO.. Num Corresponds to TOS
4174 -------------------------------------------------
4175 BESTEFFORT 0 Maximize Reliablity
4176 FILLER 1 Minimize Cost
4177 BULK 2 Maximize Throughput (0x8)
4179 INTERACTIVE 6 Minimize Delay (0x10)
4186 The TC_PRIO.. number corresponds to bits, counted from the right. See the
4187 pfifo_fast section for more details how TOS bits are converted to
4192 Now the interactive and the bulk classes:
4198 # tc class add dev eth1 parent 1:1 classid 1:2 cbq bandwidth 10Mbit \
4199 rate 1Mbit allot 1514 cell 8 weight 100Kbit prio 3 maxburst 20 \
4200 avpkt 1000 split 1:0 defmap c0
4202 # tc class add dev eth1 parent 1:1 classid 1:3 cbq bandwidth 10Mbit \
4203 rate 8Mbit allot 1514 cell 8 weight 800Kbit prio 7 maxburst 20 \
4204 avpkt 1000 split 1:0 defmap 3f
4210 The 'split qdisc' is 1:0, which is where the choice will be made. C0 is
4211 binary for 11000000, 3F for 00111111, so these two together will match
4212 everything. The first class matches bits 7 & 6, and thus corresponds
4213 to 'interactive' and 'control' traffic. The second class matches the rest.
4217 Node 1:0 now has a table like this:
4234 For additional fun, you can also pass a 'change mask', which indicates
4235 exactly which priorities you wish to change. You only need to use this if you
4236 are running 'tc class change'. For example, to add best effort traffic to
4237 1:2, we could run this:
4243 # tc class change dev eth1 classid 1:2 cbq defmap 01/01
4249 The priority map over at 1:0 now looks like this:
4269 FIXME: did not test 'tc class change', only looked at the source.
4277 <Title>Hierarchical Token Bucket </Title>
4280 Martin Devera (<devik>) rightly realised that CBQ is complex and does
4281 not seem optimized for many typical situations. His Hierarchical approach is
4282 well suited for setups where you have a fixed amount of bandwidth which you
4283 want to divide for different purposes, giving each purpose a guaranteed
4284 bandwidth, with the possibility of specifying how much bandwidth can be
4289 HTB works just like CBQ but does not resort to idle time calculations to
4290 shape. Instead, it is a classful Token Bucket Filter - hence the name. It
4291 has only a few parameters, which are well documented on his
4293 URL="http://luxik.cdi.cz/~devik/qos/htb/"
4299 As your HTB configuration gets more complex, your configuration scales
4300 well. With CBQ it is already complex even in simple cases! HTB3 (check
4301 <ulink url="http://luxik.cdi.cz/~devik/qos/htb/">its homepage</ulink> for
4302 details on HTB versions) is now part of the official kernel sources
4303 (from 2.4.20-pre1 and 2.5.31 onwards). However, maybe you still need to
4304 get a HTB3 patched version of 'tc': HTB kernel and userspace parts must
4305 be the same major version, or 'tc' will not work with HTB.
4310 If you already have a modern kernel, or are in a position to patch your
4311 kernel, by all means consider HTB.
4316 <Title>Sample configuration</Title>
4319 Functionally almost identical to the CBQ sample configuration above:
4325 # tc qdisc add dev eth0 root handle 1: htb default 30
4327 # tc class add dev eth0 parent 1: classid 1:1 htb rate 6mbit burst 15k
4329 # tc class add dev eth0 parent 1:1 classid 1:10 htb rate 5mbit burst 15k
4330 # tc class add dev eth0 parent 1:1 classid 1:20 htb rate 3mbit ceil 6mbit burst 15k
4331 # tc class add dev eth0 parent 1:1 classid 1:30 htb rate 1kbit ceil 6mbit burst 15k
4337 The author then recommends SFQ for beneath these classes:
4340 # tc qdisc add dev eth0 parent 1:10 handle 10: sfq perturb 10
4341 # tc qdisc add dev eth0 parent 1:20 handle 20: sfq perturb 10
4342 # tc qdisc add dev eth0 parent 1:30 handle 30: sfq perturb 10
4348 Add the filters which direct traffic to the right classes:
4351 # U32="tc filter add dev eth0 protocol ip parent 1:0 prio 1 u32"
4352 # $U32 match ip dport 80 0xffff flowid 1:10
4353 # $U32 match ip sport 25 0xffff flowid 1:20
4356 And that's it - no unsightly unexplained numbers, no undocumented
4361 HTB certainly looks wonderful - if 10: and 20: both have their guaranteed
4362 bandwidth, and more is left to divide, they borrow in a 5:3 ratio, just as
4367 Unclassified traffic gets routed to 30:, which has little bandwidth of its
4368 own but can borrow everything that is left over. Because we chose SFQ
4369 internally, we get fairness thrown in for free!
4378 <Sect1 id="lartc.qdisc.filters">
4379 <Title>Classifying packets with filters</Title>
4382 To determine which class shall process a packet, the so-called 'classifier
4383 chain' is called each time a choice needs to be made. This chain consists of
4384 all filters attached to the classful qdisc that needs to decide.
4387 <Para>To reiterate the tree, which is not a tree:
4403 When enqueueing a packet, at each branch the filter chain is consulted for a
4404 relevant instruction. A typical setup might be to have a filter in 1:1 that
4405 directs a packet to 12: and a filter on 12: that sends the packet to 12:2.
4409 You might also attach this latter rule to 1:1, but you can make efficiency
4410 gains by having more specific tests lower in the chain.
4414 You can't filter a packet 'upwards', by the way. Also, with HTB, you should
4415 attach all filters to the root!
4419 And again - packets are only enqueued downwards! When they are dequeued,
4420 they go up again, where the interface lives. They do NOT fall off the end of
4421 the tree to the network adaptor!
4425 <Title>Some simple filtering examples</Title>
4428 As explained in the Classifier chapter, you can match on literally anything,
4429 using a very complicated syntax. To start, we will show how to do the
4430 obvious things, which luckily are quite easy.
4434 Let's say we have a PRIO qdisc called '10:' which contains three classes, and
4435 we want to assign all traffic from and to port 22 to the highest priority
4436 band, the filters would be:
4442 # tc filter add dev eth0 protocol ip parent 10: prio 1 u32 match \
4443 ip dport 22 0xffff flowid 10:1
4444 # tc filter add dev eth0 protocol ip parent 10: prio 1 u32 match \
4445 ip sport 80 0xffff flowid 10:1
4446 # tc filter add dev eth0 protocol ip parent 10: prio 2 flowid 10:2
4452 What does this say? It says: attach to eth0, node 10: a priority 1 u32
4453 filter that matches on IP destination port 22 *exactly* and send it to band
4454 10:1. And it then repeats the same for source port 80. The last command says
4455 that anything unmatched so far should go to band 10:2, the next-highest
4460 You need to add 'eth0', or whatever your interface is called, because each
4461 interface has a unique namespace of handles.
4465 To select on an IP address, use this:
4468 # tc filter add dev eth0 parent 10:0 protocol ip prio 1 u32 \
4469 match ip dst 4.3.2.1/32 flowid 10:1
4470 # tc filter add dev eth0 parent 10:0 protocol ip prio 1 u32 \
4471 match ip src 1.2.3.4/32 flowid 10:1
4472 # tc filter add dev eth0 protocol ip parent 10: prio 2 \
4479 This assigns traffic to 4.3.2.1 and traffic from 1.2.3.4 to the highest
4480 priority queue, and the rest to the next-highest one.
4484 You can concatenate matches, to match on traffic from 1.2.3.4 and from port
4488 # tc filter add dev eth0 parent 10:0 protocol ip prio 1 u32 match ip src 4.3.2.1/32
4489 match ip sport 80 0xffff flowid 10:1
4496 <Sect2 id="lartc.filtering.simple">
4497 <Title>All the filtering commands you will normally need</Title>
4500 Most shaping commands presented here start with this preamble:
4503 # tc filter add dev eth0 parent 1:0 protocol ip prio 1 u32 ..
4506 These are the so called 'u32' matches, which can match on ANY part of a
4511 <Term>On source/destination address</Term>
4514 Source mask 'match ip src 1.2.3.0/24', destination mask 'match ip dst
4515 4.3.2.0/24'. To match a single host, use /32, or omit the mask.
4519 <Term>On source/destination port, all IP protocols</Term>
4522 Source: 'match ip sport 80 0xffff', 'match ip dport 0xffff'
4526 <Term>On ip protocol (tcp, udp, icmp, gre, ipsec)</Term>
4529 Use the numbers from /etc/protocols, for example, icmp is 1: 'match ip
4534 <Term>On fwmark</Term>
4537 You can mark packets with either ipchains and have that mark survive routing
4538 across interfaces. This is really useful to for example only shape traffic on
4539 eth1 that came in on eth0. Syntax:
4540 # tc filter add dev eth1 protocol ip parent 1:0 prio 1 handle 6 fw flowid 1:1
4541 Note that this is not a u32 match!
4545 You can place a mark like this:
4548 # iptables -A PREROUTING -t mangle -i eth0 -j MARK --set-mark 6
4551 The number 6 is arbitrary.
4555 If you don't want to understand the full tc filter syntax, just use
4556 iptables, and only learn to select on fwmark.
4560 <Term>On the TOS field</Term>
4563 To select interactive, minimum delay traffic:
4566 # tc filter add dev ppp0 parent 1:0 protocol ip prio 10 u32 \
4567 match ip tos 0x10 0xff \
4571 Use 0x08 0xff for bulk traffic.
4578 For more filtering commands, see the Advanced Filters chapter.
4584 <Sect1 id="lartc.imq">
4585 <Title>The Intermediate queueing device (IMQ)</Title>
4588 The Intermediate queueing device is not a qdisc but its usage is tightly bound
4589 to qdiscs. Within linux, qdiscs are attached to network devices and everything
4590 that is queued to the device is first queued to the qdisc. From this concept,
4591 two limitations arise:
4595 1. Only egress shaping is possible (an ingress qdisc exists, but its
4596 possibilities are very limited compared to classful qdiscs).
4600 2. A qdisc can only see traffic of one interface, global limitations can't be
4605 IMQ is there to help solve those two limitations. In short, you can put
4606 everything you choose in a qdisc. Specially marked packets get intercepted
4607 in netfilter NF_IP_PRE_ROUTING and NF_IP_POST_ROUTING hooks and pass through
4608 the qdisc attached to an imq device. An iptables target is used for marking
4613 This enables you to do ingress shaping as you can just mark packets coming in from somewhere and/or treat interfaces as classes to set global limits.
4614 You can also do lots of other stuff like just putting your http traffic in a
4615 qdisc, put new connection requests in a qdisc, ...
4619 <Title>Sample configuration</Title>
4622 The first thing that might come to mind is use ingress shaping to give yourself
4623 a high guaranteed bandwidth. ;)
4624 Configuration is just like with any other interface:
4627 tc qdisc add dev imq0 root handle 1: htb default 20
4629 tc class add dev imq0 parent 1: classid 1:1 htb rate 2mbit burst 15k
4631 tc class add dev imq0 parent 1:1 classid 1:10 htb rate 1mbit
4632 tc class add dev imq0 parent 1:1 classid 1:20 htb rate 1mbit
4634 tc qdisc add dev imq0 parent 1:10 handle 10: pfifo
4635 tc qdisc add dev imq0 parent 1:20 handle 20: sfq
4637 tc filter add dev imq0 parent 10:0 protocol ip prio 1 u32 match \
4638 ip dst 10.0.0.230/32 flowid 1:10
4641 In this example u32 is used for classification. Other classifiers should work as
4643 Next traffic has to be selected and marked to be enqueued to imq0.
4646 iptables -t mangle -A PREROUTING -i eth0 -j IMQ --todev 0
4654 The IMQ iptables targets is valid in the PREROUTING and POSTROUTING chains of
4655 the mangle table. It's syntax is
4658 IMQ [ --todev n ] n : number of imq device
4661 An ip6tables target is also provided.
4665 Please note traffic is not enqueued when the target is hit but afterwards.
4666 The exact location where traffic enters the imq device depends on the
4667 direction of the traffic (in/out).
4668 These are the predefined netfilter hooks used by iptables:
4671 enum nf_ip_hook_priorities {
4672 NF_IP_PRI_FIRST = INT_MIN,
4673 NF_IP_PRI_CONNTRACK = -200,
4674 NF_IP_PRI_MANGLE = -150,
4675 NF_IP_PRI_NAT_DST = -100,
4676 NF_IP_PRI_FILTER = 0,
4677 NF_IP_PRI_NAT_SRC = 100,
4678 NF_IP_PRI_LAST = INT_MAX,
4685 For ingress traffic, imq registers itself with NF_IP_PRI_MANGLE + 1 priority
4686 which means packets enter the imq device directly after the mangle PREROUTING
4687 chain has been passed.
4691 For egress imq uses NF_IP_PRI_LAST which honours the fact that packets dropped
4692 by the filter table won't occupy bandwidth.
4696 The patches and some more information can be found at the
4698 URL="http://luxik.cdi.cz/~patrick/imq/"
4708 <chapter id="lartc.loadshare">
4709 <Title>Load sharing over multiple interfaces</Title>
4712 There are several ways of doing this. One of the easiest and straightforward
4713 ways is 'TEQL' - "True" (or "trivial") link equalizer. Like most things
4714 having to do with queueing, load sharing goes both ways. Both ends of a link
4715 may need to participate for full effect.
4719 Imagine this situation:
4725 +-------+ eth1 +-------+
4727 'network 1' ----| A | | B |---- 'network 2'
4729 +-------+ eth2 +-------+
4735 A and B are routers, and for the moment we'll assume both run Linux. If
4736 traffic is going from network 1 to network 2, router A needs to distribute
4737 the packets over both links to B. Router B needs to be configured to accept
4738 this. Same goes the other way around, when packets go from network 2 to
4739 network 1, router B needs to send the packets over both eth1 and eth2.
4743 The distributing part is done by a 'TEQL' device, like this (it couldn't be
4750 # tc qdisc add dev eth1 root teql0
4751 # tc qdisc add dev eth2 root teql0
4752 # ip link set dev teql0 up
4758 Don't forget the 'ip link set up' command!
4762 This needs to be done on both hosts. The device teql0 is basically a
4763 roundrobbin distributor over eth1 and eth2, for sending packets. No data
4764 ever comes in over an teql device, that just appears on the 'raw' eth1 and
4769 But now we just have devices, we also need proper routing. One way to do
4770 this is to assign a /31 network to both links, and a /31 to the teql0 device
4778 # ip addr add dev eth1 10.0.0.0/31
4779 # ip addr add dev eth2 10.0.0.2/31
4780 # ip addr add dev teql0 10.0.0.4/31
4789 # ip addr add dev eth1 10.0.0.1/31
4790 # ip addr add dev eth2 10.0.0.3/31
4791 # ip addr add dev teql0 10.0.0.5/31
4797 Router A should now be able to ping 10.0.0.1, 10.0.0.3 and 10.0.0.5 over the
4798 2 real links and the 1 equalized device. Router B should be able to ping
4799 10.0.0.0, 10.0.0.2 and 10.0.0.4 over the links.
4803 If this works, Router A should make 10.0.0.5 its route for reaching network
4804 2, and Router B should make 10.0.0.4 its route for reaching network 1. For
4805 the special case where network 1 is your network at home, and network 2 is
4806 the Internet, Router A should make 10.0.0.5 its default gateway.
4809 <Sect1 id="lartc.loadshare.caveats">
4810 <Title>Caveats</Title>
4813 Nothing is as easy as it seems. eth1 and eth2 on both router A and B need to
4814 have return path filtering turned off, because they will otherwise drop
4815 packets destined for ip addresses other than their own:
4821 # echo 0 > /proc/sys/net/ipv4/conf/eth1/rp_filter
4822 # echo 0 > /proc/sys/net/ipv4/conf/eth2/rp_filter
4828 Then there is the nasty problem of packet reordering. Let's say 6 packets
4829 need to be sent from A to B - eth1 might get 1, 3 and 5. eth2 would then do
4830 2, 4 and 6. In an ideal world, router B would receive this in order, 1, 2,
4831 3, 4, 5, 6. But the possibility is very real that the kernel gets it like
4832 this: 2, 1, 4, 3, 6, 5. The problem is that this confuses TCP/IP. While not
4833 a problem for links carrying many different TCP/IP sessions, you won't be
4834 able to to a bundle multiple links and get to ftp a single file lots faster,
4835 except when your receiving or sending OS is Linux, which is not easily
4836 shaken by some simple reordering.
4840 However, for lots of applications, link load balancing is a great idea.
4844 <Sect1 id="lartc.loadshare.other">
4845 <Title>Other possibilities</Title>
4847 William Stearns has used an advanced tunneling setup to achieve good use of
4848 multiple, unrelated, internet connections together. It can be found on
4850 URL="http://www.stearns.org/tunnel/">his tunneling page</ULink>.
4853 The HOWTO may feature more about this in the future.
4858 <chapter id="lartc.netfilter">
4859 <Title>Netfilter & iproute - marking packets</Title>
4862 So far we've seen how iproute works, and netfilter was mentioned a few
4863 times. This would be a good time to browse through <ULink
4864 URL="http://netfilter.samba.org/unreliable-guides/"
4865 >Rusty's Remarkably Unreliable Guides</ULink
4868 URL="http://netfilter.filewatcher.org/"
4874 Netfilter allows us to filter packets, or mangle their headers. One special
4875 feature is that we can mark a packet with a number. This is done with the
4876 --set-mark facility.
4880 As an example, this command marks all packets destined for port 25, outgoing
4887 # iptables -A PREROUTING -i eth0 -t mangle -p tcp --dport 25 \
4888 -j MARK --set-mark 1
4894 Let's say that we have multiple connections, one that is fast (and
4895 expensive, per megabyte) and one that is slower, but flat fee. We would most
4896 certainly like outgoing mail to go via the cheap route.
4900 We've already marked the packets with a '1', we now instruct the routing
4901 policy database to act on this:
4907 # echo 201 mail.out >> /etc/iproute2/rt_tables
4908 # ip rule add fwmark 1 table mail.out
4910 0: from all lookup local
4911 32764: from all fwmark 1 lookup mail.out
4912 32766: from all lookup main
4913 32767: from all lookup default
4919 Now we generate the mail.out table with a route to the slow but cheap link:
4922 # /sbin/ip route add default via 195.96.98.253 dev ppp0 table mail.out
4928 And we are done. Should we want to make exceptions, there are lots of ways
4929 to achieve this. We can modify the netfilter statement to exclude certain
4930 hosts, or we can insert a rule with a lower priority that points to the main
4931 table for our excepted hosts.
4935 We can also use this feature to honour TOS bits by marking packets with a
4936 different type of service with different numbers, and creating rules to act
4937 on that. This way you can even dedicate, say, an ISDN line to interactive
4942 Needless to say, this also works fine on a host that's doing NAT
4947 IMPORTANT: We received a report that MASQ and SNAT at least collide
4948 with marking packets. Rusty Russell explains it in
4950 URL="http://lists.samba.org/pipermail/netfilter/2000-November/006089.html"
4951 >this posting</ULink
4952 >. Turn off the reverse path filter to make it work
4957 Note: to mark packets, you need to have some options enabled in your
4964 IP: advanced router (CONFIG_IP_ADVANCED_ROUTER) [Y/n/?]
4965 IP: policy routing (CONFIG_IP_MULTIPLE_TABLES) [Y/n/?]
4966 IP: use netfilter MARK value as routing key (CONFIG_IP_ROUTE_FWMARK) [Y/n/?]
4972 See also the <xref linkend="lartc.cookbook.squid"> in the
4973 <citetitle><xref linkend="lartc.cookbook"></citetitle>.
4978 <chapter id="lartc.adv-filter"
4979 xreflabel="Advanced filters for (re-)classifying packets">
4980 <Title>Advanced filters for (re-)classifying packets</Title>
4983 As explained in the section on classful queueing disciplines, filters are
4984 needed to classify packets into any of the sub-queues. These filters are
4985 called from within the classful qdisc.
4989 Here is an incomplete list of classifiers available:
4996 Bases the decision on how the firewall has marked the packet. This can be
4997 the easy way out if you don't want to learn tc filter syntax. See the
4998 Queueing chapter for details.
5005 Bases the decision on fields within the packet (i.e. source IP address, etc)
5012 Bases the decision on which route the packet will be routed by
5016 <Term>rsvp, rsvp6</Term>
5019 Routes packets based on <ULink
5020 URL="http://www.isi.edu/div7/rsvp/overview.html"
5023 on networks you control - the Internet does not respect RSVP.
5027 <Term>tcindex</Term>
5030 Used in the DSMARK qdisc, see the relevant section.
5037 Note that in general there are many ways in which you can classify packet
5038 and that it generally comes down to preference as to which system you wish
5043 Classifiers in general accept a few arguments in common. They are listed
5044 here for convenience:
5051 <Term>protocol</Term>
5054 The protocol this classifier will accept. Generally you will only be
5055 accepting only IP traffic. Required.
5062 The handle this classifier is to be attached to. This handle must be
5063 an already existing class. Required.
5070 The priority of this classifier. Lower numbers get tested first.
5077 This handle means different things to different filters.
5084 All the following sections will assume you are trying to shape the traffic
5085 going to <Literal remap="tt">HostA</Literal>. They will assume that the root class has been
5086 configured on 1: and that the class you want to send the selected traffic to
5090 <Sect1 id="lartc.adv-filter.u32">
5091 <Title>The <option>u32</option> classifier</Title>
5094 The U32 filter is the most advanced filter available in the current
5095 implementation. It entirely based on hashing tables, which make it
5096 robust when there are many filter rules.
5100 In its simplest form the U32 filter is a list of records, each
5101 consisting of two fields: a selector and an action. The selectors,
5102 described below, are compared with the currently processed IP packet
5103 until the first match occurs, and then the associated action is performed.
5104 The simplest type of action would be directing the packet into defined
5109 The command line of <Literal remap="tt">tc filter</Literal> program, used to configure the filter,
5110 consists of three parts: filter specification, a selector and an action.
5111 The filter specification can be defined as:
5117 tc filter add dev IF [ protocol PROTO ]
5118 [ (preference|priority) PRIO ]
5125 The <Literal remap="tt">protocol</Literal> field describes protocol that the filter will be
5126 applied to. We will only discuss case of <Literal remap="tt">ip</Literal> protocol. The
5127 <Literal remap="tt">preference</Literal> field (<Literal remap="tt">priority</Literal> can be used alternatively)
5128 sets the priority of currently defined filter. This is important, since
5129 you can have several filters (lists of rules) with different priorities.
5130 Each list will be passed in the order the rules were added, then list with
5131 lower priority (higher preference number) will be processed. The <Literal remap="tt">parent</Literal>
5132 field defines the CBQ tree top (e.g. 1:0), the filter should be attached
5137 The options described above apply to all filters, not only U32.
5141 <Title>U32 selector </Title>
5144 The U32 selector contains definition of the pattern, that will be matched
5145 to the currently processed packet. Precisely, it defines which bits are
5146 to be matched in the packet header and nothing more, but this simple
5147 method is very powerful. Let's take a look at the following examples,
5148 taken directly from a pretty complex, real-world filter:
5154 # tc filter add dev eth0 protocol ip parent 1:0 pref 10 u32 \
5155 match u32 00100000 00ff0000 at 0 flowid 1:10
5161 For now, leave the first line alone - all these parameters describe
5162 the filter's hash tables. Focus on the selector line, containing
5163 <Literal remap="tt">match</Literal> keyword. This selector will match to IP headers, whose
5164 second byte will be 0x10 (0010). As you can guess, the 00ff number is
5165 the match mask, telling the filter exactly which bits to match. Here
5166 it's 0xff, so the byte will match if it's exactly 0x10. The <Literal remap="tt">at</Literal>
5167 keyword means that the match is to be started at specified offset (in
5168 bytes) -- in this case it's beginning of the packet. Translating all
5169 that to human language, the packet will match if its Type of Service
5170 field will have `low delay' bits set. Let's analyze another rule:
5176 # tc filter add dev eth0 protocol ip parent 1:0 pref 10 u32 \
5177 match u32 00000016 0000ffff at nexthdr+0 flowid 1:10
5183 The <Literal remap="tt">nexthdr</Literal> option means next header encapsulated in the IP packet,
5184 i.e. header of upper-layer protocol. The match will also start here
5185 at the beginning of the next header. The match should occur in the
5186 second, 32-bit word of the header. In TCP and UDP protocols this field
5187 contains packet's destination port. The number is given in big-endian
5188 format, i.e. older bits first, so we simply read 0x0016 as 22 decimal,
5189 which stands for SSH service if this was TCP. As you guess, this match
5190 is ambiguous without a context, and we will discuss this later.
5194 Having understood all the above, we will find the following selector
5195 quite easy to read: <Literal remap="tt">match c0a80100 ffffff00 at 16</Literal>. What we
5196 got here is a three byte match at 17-th byte, counting from the IP
5197 header start. This will match for packets with destination address
5198 anywhere in 192.168.1/24 network. After analyzing the examples, we
5199 can summarize what we have learned.
5205 <Title>General selectors</Title>
5208 General selectors define the pattern, mask and offset the pattern
5209 will be matched to the packet contents. Using the general selectors
5210 you can match virtually any single bit in the IP (or upper layer)
5211 header. They are more difficult to write and read, though, than
5212 specific selectors that described below. The general selector syntax
5219 match [ u32 | u16 | u8 ] PATTERN MASK [ at OFFSET | nexthdr+OFFSET]
5225 One of the keywords <Literal remap="tt">u32</Literal>, <Literal remap="tt">u16</Literal> or <Literal remap="tt">u8</Literal> specifies
5226 length of the pattern in bits. PATTERN and MASK should follow, of length
5227 defined by the previous keyword. The OFFSET parameter is the offset,
5228 in bytes, to start matching. If <Literal remap="tt">nexthdr+</Literal> keyword is given,
5229 the offset is relative to start of the upper layer header.
5239 # tc filter add dev ppp14 parent 1:0 prio 10 u32 \
5240 match u8 64 0xff at 8 \
5247 Packet will match to this rule, if its time to live (TTL) is 64.
5248 TTL is the field starting just after 8-th byte of the IP header.
5253 Matches all TCP packets which have the ACK bit set:
5259 # tc filter add dev ppp14 parent 1:0 prio 10 u32 \
5260 match ip protocol 6 0xff \
5261 match u8 0x10 0xff at nexthdr+13 \
5268 Use this to match ACKs on packets smaller than 64 bytes:
5274 ## match acks the hard way,
5276 ## IP header length 0x5(32 bit words),
5277 ## IP Total length 0x34 (ACK + 12 bytes of TCP options)
5278 ## TCP ack set (bit 5, offset 33)
5279 # tc filter add dev ppp14 parent 1:0 protocol ip prio 10 u32 \
5280 match ip protocol 6 0xff \
5281 match u8 0x05 0x0f at 0 \
5282 match u16 0x0000 0xffc0 at 2 \
5283 match u8 0x10 0xff at 33 \
5290 This rule will only match TCP packets with ACK bit set, and no further
5291 payload. Here we can see an example of using two selectors, the final result
5292 will be logical AND of their results. If we take a look at TCP header
5293 diagram, we can see that the ACK bit is second older bit (0x10) in the 14-th
5294 byte of the TCP header (<Literal remap="tt">at nexthdr+13</Literal>). As for the second
5295 selector, if we'd like to make our life harder, we could write <Literal remap="tt">match u8
5296 0x06 0xff at 9</Literal> instead of using the specific selector <Literal remap="tt">protocol
5297 tcp</Literal>, because 6 is the number of TCP protocol, present in 10-th byte of
5298 the IP header. On the other hand, in this example we couldn't use any
5299 specific selector for the first match - simply because there's no specific
5300 selector to match TCP ACK bits.
5304 The filter below is a modified version of the filter above. The difference is, that it
5305 doesn't check the ip header length. Why? Because the filter above does only work on 32
5312 tc filter add dev ppp14 parent 1:0 protocol ip prio 10 u32 \
5313 match ip protocol 6 0xff \
5314 match u8 0x10 0xff at nexthdr+13 \
5315 match u16 0x0000 0xffc0 at 2 \
5325 <Title>Specific selectors</Title>
5328 The following table contains a list of all specific selectors
5329 the author of this section has found in the <Literal remap="tt">tc</Literal> program
5330 source code. They simply make your life easier and increase readability
5331 of your filter's configuration.
5335 FIXME: table placeholder - the table is in separate file ,,selector.html''
5339 FIXME: it's also still in Polish :-(
5343 FIXME: must be sgml'ized
5353 # tc filter add dev ppp0 parent 1:0 prio 10 u32 \
5354 match ip tos 0x10 0xff \
5361 FIXME: tcp dport match does not work as described below:
5365 The above rule will match packets which have the TOS field set to 0x10.
5366 The TOS field starts at second byte of the packet and is one byte big,
5367 so we could write an equivalent general selector: <Literal remap="tt">match u8 0x10 0xff
5368 at 1</Literal>. This gives us hint to the internals of U32 filter -- the
5369 specific rules are always translated to general ones, and in this
5370 form they are stored in the kernel memory. This leads to another conclusion
5371 -- the <Literal remap="tt">tcp</Literal> and <Literal remap="tt">udp</Literal> selectors are exactly the same
5372 and this is why you can't use single <Literal remap="tt">match tcp dport 53 0xffff</Literal>
5373 selector to match TCP packets sent to given port -- they will also
5374 match UDP packets sent to this port. You must remember to also specify
5375 the protocol and end up with the following rule:
5381 # tc filter add dev ppp0 parent 1:0 prio 10 u32 \
5382 match tcp dport 53 0xffff \
5383 match ip protocol 0x6 0xff \
5393 <Sect1 id="lartc.adv-filter.route">
5394 <Title>The <option>route</option> classifier</Title>
5397 This classifier filters based on the results of the routing tables. When a
5398 packet that is traversing through the classes reaches one that is marked
5399 with the "route" filter, it splits the packets up based on information in
5406 # tc filter add dev eth1 parent 1:0 protocol ip prio 100 route
5412 Here we add a route classifier onto the parent node 1:0 with priority 100.
5413 When a packet reaches this node (which, since it is the root, will happen
5414 immediately) it will consult the routing table and if one matches will
5415 send it to the given class and give it a priority of 100. Then, to finally
5416 kick it into action, you add the appropriate routing entry:
5420 The trick here is to define 'realm' based on either destination or source.
5421 The way to do it is like this:
5427 # ip route add Host/Network via Gateway dev Device realm RealmNumber
5433 For instance, we can define our destination network 192.168.10.0 with a realm
5440 # ip route add 192.168.10.0/24 via 192.168.10.1 dev eth1 realm 10
5446 When adding route filters, we can use realm numbers to represent the
5447 networks or hosts and specify how the routes match the filters.
5453 # tc filter add dev eth1 parent 1:0 protocol ip prio 100 \
5454 route to 10 classid 1:10
5460 The above rule says packets going to the network 192.168.10.0 match class id
5465 Route filter can also be used to match source routes. For example, there is
5466 a subnetwork attached to the Linux router on eth2.
5472 # ip route add 192.168.2.0/24 dev eth2 realm 2
5473 # tc filter add dev eth1 parent 1:0 protocol ip prio 100 \
5474 route from 2 classid 1:2
5480 Here the filter specifies that packets from the subnetwork 192.168.2.0
5481 (realm 2) will match class id 1:2.
5486 <Sect1 id="lartc.adv-filter.policing">
5487 <Title>Policing filters</Title>
5490 To make even more complicated setups possible, you can have filters that
5491 only match up to a certain bandwidth. You can declare a filter to entirely
5492 cease matching above a certain rate, or only to not match only the bandwidth
5493 exceeding a certain rate.
5497 So if you decided to police at 4mbit/s, but 5mbit/s of traffic is present,
5498 you can stop matching either the entire 5mbit/s, or only not match 1mbit/s,
5499 and do send 4mbit/s to the configured class.
5503 If bandwidth exceeds the configured rate, you can drop a packet, reclassify
5504 it, or see if another filter will match it.
5508 <Title>Ways to police</Title>
5511 There are basically two ways to police. If you compiled the kernel
5512 with 'Estimators', the kernel can measure for each filter how much traffic
5513 it is passing, more or less. These estimators are very easy on the CPU, as
5514 they simply count 25 times per second how many data has been passed, and
5515 calculate the bitrate from that.
5519 The other way works again via a Token Bucket Filter, this time living within
5520 your filter. The TBF only matches traffic UP TO your configured bandwidth,
5521 if more is offered, only the excess is subject to the configured overlimit
5526 <Title>With the kernel estimator</Title>
5529 This is very simple and has only one parameter: avrate. Either the flow
5530 remains below avrate, and the filter classifies the traffic to the classid
5531 configured, or your rate exceeds it in which case the specified action is
5532 taken, which is 'reclassify' by default.
5536 The kernel uses an Exponential Weighted Moving Average for your bandwidth
5537 which makes it less sensitive to short bursts.
5543 <Title>With Token Bucket Filter</Title>
5546 Uses the following parameters:
5579 Which behave mostly identical to those described in the Token Bucket Filter
5580 section. Please note however that if you set the mtu of a TBF policer too
5581 low, *no* packets will pass, whereas the egress TBF qdisc will just pass
5586 Another difference is that a policer can only let a packet pass, or drop it.
5587 It cannot delay hold on to it in order to delay it.
5595 <Title>Overlimit actions</Title>
5598 If your filter decides that it is overlimit, it can take 'actions'.
5599 Currently, three actions are available:
5603 <Term>continue</Term>
5606 Causes this filter not to match, but perhaps other filters will.
5613 This is a very fierce option which simply discards traffic exceeding a
5614 certain rate. It is often used in the ingress policer and has limited uses.
5615 For example, you may have a name server that falls over if offered more than
5616 5mbit/s of packets, in which case an ingress filter could be used to make
5617 sure no more is ever offered.
5621 <Term>Pass/OK</Term>
5624 Pass on traffic ok. Might be used to disable a complicated filter, but leave
5629 <Term>reclassify</Term>
5632 Most often comes down to reclassification to Best Effort. This is the
5642 <Title>Examples</Title>
5645 The only real example known is mentioned in the 'Protecting your host
5646 from SYN floods' section.
5650 FIXME: if you have used this, please share your experience with us
5657 <Sect1 id="lartc.adv-filter.hashing">
5658 <Title>Hashing filters for very fast massive filtering</Title>
5661 If you have a need for thousands of rules, for example if you have a lot of
5662 clients or computers, all with different QoS specifications, you may find
5663 that the kernel spends a lot of time matching all those rules.
5667 By default, all filters reside in one big chain which is matched in
5668 descending order of priority. If you have 1000 rules, 1000 checks may be
5669 needed to determine what to do with a packet.
5673 Matching would go much quicker if you would have 256 chains with each four
5674 rules - if you could divide packets over those 256 chains, so that the right
5679 Hashing makes this possible. Let's say you have 1024 cable modem customers in
5680 your network, with IP addresses ranging from 1.2.0.0 to 1.2.3.255, and each
5681 has to go in another bin, for example 'lite', 'regular' and 'premium'. You
5682 would then have 1024 rules like this:
5688 # tc filter add dev eth1 parent 1:0 protocol ip prio 100 match ip src \
5690 # tc filter add dev eth1 parent 1:0 protocol ip prio 100 match ip src \
5693 # tc filter add dev eth1 parent 1:0 protocol ip prio 100 match ip src \
5694 1.2.3.254 classid 1:3
5695 # tc filter add dev eth1 parent 1:0 protocol ip prio 100 match ip src \
5696 1.2.3.255 classid 1:2
5702 To speed this up, we can use the last part of the IP address as a 'hash
5703 key'. We then get 256 tables, the first of which looks like this:
5706 # tc filter add dev eth1 parent 1:0 protocol ip prio 100 match ip src \
5708 # tc filter add dev eth1 parent 1:0 protocol ip prio 100 match ip src \
5710 # tc filter add dev eth1 parent 1:0 protocol ip prio 100 match ip src \
5712 # tc filter add dev eth1 parent 1:0 protocol ip prio 100 match ip src \
5719 The next one starts like this:
5722 # tc filter add dev eth1 parent 1:0 protocol ip prio 100 match ip src \
5730 This way, only four checks are needed at most, two on average.
5734 Configuration is pretty complicated, but very worth it by the time you have
5735 this many rules. First we make a filter root, then we create a table with
5739 # tc filter add dev eth1 parent 1:0 prio 5 protocol ip u32
5740 # tc filter add dev eth1 parent 1:0 prio 5 handle 2: protocol ip u32 divisor 256
5746 Now we add some rules to entries in the created table:
5752 # tc filter add dev eth1 protocol ip parent 1:0 prio 5 u32 ht 2:7b: \
5753 match ip src 1.2.0.123 flowid 1:1
5754 # tc filter add dev eth1 protocol ip parent 1:0 prio 5 u32 ht 2:7b: \
5755 match ip src 1.2.1.123 flowid 1:2
5756 # tc filter add dev eth1 protocol ip parent 1:0 prio 5 u32 ht 2:7b: \
5757 match ip src 1.2.3.123 flowid 1:3
5758 # tc filter add dev eth1 protocol ip parent 1:0 prio 5 u32 ht 2:7b: \
5759 match ip src 1.2.4.123 flowid 1:2
5762 This is entry 123, which contains matches for 1.2.0.123, 1.2.1.123,
5763 1.2.2.123, 1.2.3.123, and sends them to 1:1, 1:2, 1:3 and 1:2 respectively.
5764 Note that we need to specify our hash bucket in hex, 0x7b is 123.
5768 Next create a 'hashing filter' that directs traffic to the right entry in
5772 # tc filter add dev eth1 protocol ip parent 1:0 prio 5 u32 ht 800:: \
5773 match ip src 1.2.0.0/16 \
5774 hashkey mask 0x000000ff at 12 \
5778 Ok, some numbers need explaining. The default hash table is called 800:: and
5779 all filtering starts there. Then we select the source address, which lives
5780 as position 12, 13, 14 and 15 in the IP header, and indicate that we are
5781 only interested in the last part. This we send to hash table 2:, which we
5786 It is quite complicated, but it does work in practice and performance will
5787 be staggering. Note that this example could be improved to the ideal case
5788 where each chain contains 1 filter!
5795 <chapter id="lartc.kernel">
5796 <Title>Kernel network parameters </Title>
5800 The kernel has lots of parameters which
5801 can be tuned for different circumstances. While, as usual, the default
5802 parameters serve 99% of installations very well, we don't call this the
5803 Advanced HOWTO for the fun of it!
5807 The interesting bits are in /proc/sys/net, take a look there. Not everything
5808 will be documented here initially, but we're working on it.
5812 In the meantime you may want to have a look at the Linux-Kernel sources;
5813 read the file Documentation/filesystems/proc.txt. Most of the
5814 features are explained there.
5821 <Sect1 id="lartc.kernel.rpf"
5822 xreflabel="Reverse Path Filtering">
5823 <Title>Reverse Path Filtering</Title>
5826 By default, routers route everything, even packets which 'obviously' don't
5827 belong on your network. A common example is private IP space escaping onto
5828 the Internet. If you have an interface with a route of 195.96.96.0/24 to it,
5829 you do not expect packets from 212.64.94.1 to arrive there.
5833 Lots of people will want to turn this feature off, so the kernel hackers
5834 have made it easy. There are files in /proc where you can tell
5835 the kernel to do this for you. The method is called "Reverse Path
5836 Filtering". Basically, if the reply to this packet wouldn't go out the
5837 interface this packet came in, then this is a bogus packet and should be
5842 The following fragment will turn this on for all current and future
5849 # for i in /proc/sys/net/ipv4/conf/*/rp_filter ; do
5850 > echo 2 > $i
5857 Going by the example above, if a packet arrived on the Linux router on eth1
5858 claiming to come from the Office+ISP subnet, it would be dropped. Similarly,
5859 if a packet came from the Office subnet, claiming to be from somewhere
5860 outside your firewall, it would be dropped also.
5864 The above is full reverse path filtering. The default is to only filter
5865 based on IPs that are on directly connected networks. This is because the
5866 full filtering breaks in the case of asymmetric routing (where packets come
5867 in one way and go out another, like satellite traffic, or if you have
5868 dynamic (bgp, ospf, rip) routes in your network. The data comes down
5869 through the satellite dish and replies go back through normal land-lines).
5873 If this exception applies to you (and you'll probably know if it does) you
5874 can simply turn off the rp_filter on the interface where the
5875 satellite data comes in. If you want to see if any packets are being
5876 dropped, the log_martians file in the same directory will tell
5877 the kernel to log them to your syslog.
5883 # echo 1 >/proc/sys/net/ipv4/conf/<interfacename>/log_martians
5889 FIXME: is setting the conf/{default,all}/* files enough? - martijn
5894 <Sect1 id="lartc.kernel.obscure">
5895 <Title>Obscure settings</Title>
5898 Ok, there are a lot of parameters which can be modified. We try to list them
5899 all. Also documented (partly) in Documentation/ip-sysctl.txt.
5903 Some of these settings have different defaults based on whether you
5904 answered 'Yes' to 'Configure as router and not host' while compiling your
5909 Oskar Andreasson also has a page on all these flags and it appears to be
5910 better than ours, so also check
5911 <ulink url="http://ipsysctl-tutorial.frozentux.net/">
5912 http://ipsysctl-tutorial.frozentux.net/</ulink>.
5916 <Title>Generic ipv4</Title>
5919 As a generic note, most rate limiting features don't work on loopback, so
5920 don't test them locally. The limits are supplied in 'jiffies', and are
5921 enforced using the earlier mentioned token bucket filter.
5925 The kernel has an internal clock which runs at 'HZ' ticks (or 'jiffies') per
5926 second. On Intel, 'HZ' is mostly 100. So setting a *_rate file to, say 50,
5927 would allow for 2 packets per second. The token bucket filter is also
5928 configured to allow for a burst of at most 6 packets, if enough tokens have
5933 Several entries in the following list have been copied from
5934 /usr/src/linux/Documentation/networking/ip-sysctl.txt, written by Alexey
5935 Kuznetsov <kuznet@ms2.inr.ac.ru> and Andi Kleen <ak@muc.de>
5939 <Term>/proc/sys/net/ipv4/icmp_destunreach_rate</Term>
5942 If the kernel decides that it can't deliver a packet, it will drop it, and
5943 send the source of the packet an ICMP notice to this effect.
5947 <Term>/proc/sys/net/ipv4/icmp_echo_ignore_all</Term>
5950 Don't act on echo packets at all. Please don't set this by default, but if
5951 you are used as a relay in a DoS attack, it may be useful.
5955 <Term>/proc/sys/net/ipv4/icmp_echo_ignore_broadcasts [Useful]</Term>
5958 If you ping the broadcast address of a network, all hosts are supposed to
5959 respond. This makes for a dandy denial-of-service tool. Set this to 1 to
5960 ignore these broadcast messages.
5964 <Term>/proc/sys/net/ipv4/icmp_echoreply_rate</Term>
5967 The rate at which echo replies are sent to any one destination.
5971 <Term>/proc/sys/net/ipv4/icmp_ignore_bogus_error_responses</Term>
5974 Set this to ignore ICMP errors caused by hosts in the network reacting badly
5975 to frames sent to what they perceive to be the broadcast address.
5979 <Term>/proc/sys/net/ipv4/icmp_paramprob_rate</Term>
5982 A relatively unknown ICMP message, which is sent in response to incorrect
5983 packets with broken IP or TCP headers. With this file you can control the
5984 rate at which it is sent.
5988 <Term>/proc/sys/net/ipv4/icmp_timeexceed_rate</Term>
5991 This the famous cause of the 'Solaris middle star' in traceroutes. Limits
5992 number of ICMP Time Exceeded messages sent.
5996 <Term>/proc/sys/net/ipv4/igmp_max_memberships</Term>
5999 Maximum number of listening igmp (multicast) sockets on the host.
6000 FIXME: Is this true?
6004 <Term>/proc/sys/net/ipv4/inet_peer_gc_maxtime</Term>
6007 FIXME: Add a little explanation about the inet peer storage?
6009 Minimum interval between garbage collection passes. This interval is in
6010 effect under low (or absent) memory pressure on the pool. Measured in
6015 <Term>/proc/sys/net/ipv4/inet_peer_gc_mintime</Term>
6018 Minimum interval between garbage collection passes. This interval is in
6019 effect under high memory pressure on the pool. Measured in jiffies.
6023 <Term>/proc/sys/net/ipv4/inet_peer_maxttl</Term>
6026 Maximum time-to-live of entries. Unused entries will expire after this
6027 period of time if there is no memory pressure on the pool (i.e. when the
6028 number of entries in the pool is very small). Measured in jiffies.
6032 <Term>/proc/sys/net/ipv4/inet_peer_minttl</Term>
6035 Minimum time-to-live of entries. Should be enough to cover fragment
6036 time-to-live on the reassembling side. This minimum time-to-live
6037 is guaranteed if the pool size is less than inet_peer_threshold.
6038 Measured in jiffies.
6042 <Term>/proc/sys/net/ipv4/inet_peer_threshold</Term>
6045 The approximate size of the INET peer storage. Starting from this threshold
6046 entries will be thrown aggressively. This threshold also determines
6047 entries' time-to-live and time intervals between garbage collection passes.
6048 More entries, less time-to-live, less GC interval.
6052 <Term>/proc/sys/net/ipv4/ip_autoconfig</Term>
6055 This file contains the number one if the host received its IP configuration by
6056 RARP, BOOTP, DHCP or a similar mechanism. Otherwise it is zero.
6060 <Term>/proc/sys/net/ipv4/ip_default_ttl</Term>
6063 Time To Live of packets. Set to a safe 64. Raise it if you have a huge
6064 network. Don't do so for fun - routing loops cause much more damage that
6065 way. You might even consider lowering it in some circumstances.
6069 <Term>/proc/sys/net/ipv4/ip_dynaddr</Term>
6072 You need to set this if you use dial-on-demand with a dynamic interface
6073 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
6074 connection that brings up your interface itself does not work, but the
6079 <Term>/proc/sys/net/ipv4/ip_forward</Term>
6082 If the kernel should attempt to forward packets. Off by default.
6086 <Term>/proc/sys/net/ipv4/ip_local_port_range</Term>
6089 Range of local ports for outgoing connections. Actually quite small by
6090 default, 1024 to 4999.
6094 <Term>/proc/sys/net/ipv4/ip_no_pmtu_disc</Term>
6097 Set this if you want to disable Path MTU discovery - a technique to
6098 determine the largest Maximum Transfer Unit possible on your path. See also
6099 the section on Path MTU discovery in the
6100 <citetitle><xref linkend="lartc.cookbook"></citetitle> chapter.
6104 <Term>/proc/sys/net/ipv4/ipfrag_high_thresh</Term>
6107 Maximum memory used to reassemble IP fragments. When
6108 ipfrag_high_thresh bytes of memory is allocated for this purpose,
6109 the fragment handler will toss packets until ipfrag_low_thresh
6114 <Term>/proc/sys/net/ipv4/ip_nonlocal_bind</Term>
6117 Set this if you want your applications to be able to bind to an address
6118 which doesn't belong to a device on your system. This can be useful when
6119 your machine is on a non-permanent (or even dynamic) link, so your services
6120 are able to start up and bind to a specific address when your link is down.
6124 <Term>/proc/sys/net/ipv4/ipfrag_low_thresh</Term>
6127 Minimum memory used to reassemble IP fragments.
6131 <Term>/proc/sys/net/ipv4/ipfrag_time</Term>
6134 Time in seconds to keep an IP fragment in memory.
6138 <Term>/proc/sys/net/ipv4/tcp_abort_on_overflow</Term>
6141 A boolean flag controlling the behaviour under lots of incoming connections.
6142 When enabled, this causes the kernel to actively send RST packets when a
6143 service is overloaded.
6147 <Term>/proc/sys/net/ipv4/tcp_fin_timeout</Term>
6150 Time to hold socket in state FIN-WAIT-2, if it was closed by our side. Peer
6151 can be broken and never close its side, or even died unexpectedly. Default
6152 value is 60sec. Usual value used in 2.2 was 180 seconds, you may restore it,
6153 but remember that if your machine is even underloaded WEB server, you risk
6154 to overflow memory with kilotons of dead sockets, FIN-WAIT-2 sockets are
6155 less dangerous than FIN-WAIT-1, because they eat maximum 1.5K of memory, but
6156 they tend to live longer. Cf. tcp_max_orphans.
6160 <Term>/proc/sys/net/ipv4/tcp_keepalive_time</Term>
6163 How often TCP sends out keepalive messages when keepalive is enabled.
6169 <Term>/proc/sys/net/ipv4/tcp_keepalive_intvl</Term>
6172 How frequent probes are retransmitted, when a probe isn't acknowledged.
6174 Default: 75 seconds.
6178 <Term>/proc/sys/net/ipv4/tcp_keepalive_probes</Term>
6181 How many keepalive probes TCP will send, until it decides that the
6182 connection is broken.
6186 Multiplied with tcp_keepalive_intvl, this gives the time a link can be
6187 non-responsive after a keepalive has been sent.
6191 <Term>/proc/sys/net/ipv4/tcp_max_orphans</Term>
6194 Maximal number of TCP sockets not attached to any user file handle, held by
6195 system. If this number is exceeded orphaned connections are reset
6196 immediately and warning is printed. This limit exists only to prevent simple
6197 DoS attacks, you _must_ not rely on this or lower the limit artificially,
6198 but rather increase it (probably, after increasing installed memory), if
6199 network conditions require more than default value, and tune network
6200 services to linger and kill such states more aggressively. Let me remind you
6201 again: each orphan eats up to 64K of unswappable memory.
6205 <Term>/proc/sys/net/ipv4/tcp_orphan_retries</Term>
6208 How may times to retry before killing TCP connection, closed by our side.
6209 Default value 7 corresponds to 50sec-16min depending on RTO. If your machine
6210 is a loaded WEB server, you should think about lowering this value, such
6211 sockets may consume significant resources. Cf. tcp_max_orphans.
6215 <Term>/proc/sys/net/ipv4/tcp_max_syn_backlog</Term>
6218 Maximal number of remembered connection requests, which still did not
6219 receive an acknowledgment from connecting client. Default value is 1024 for
6220 systems with more than 128Mb of memory, and 128 for low memory machines. If
6221 server suffers of overload, try to increase this number. Warning! If you
6222 make it greater than 1024, it would be better to change TCP_SYNQ_HSIZE in
6223 include/net/tcp.h to keep TCP_SYNQ_HSIZE*16<=tcp_max_syn_backlog and to
6228 <Term>/proc/sys/net/ipv4/tcp_max_tw_buckets</Term>
6231 Maximal number of timewait sockets held by system simultaneously. If this
6232 number is exceeded time-wait socket is immediately destroyed and warning is
6233 printed. This limit exists only to prevent simple DoS attacks, you _must_
6234 not lower the limit artificially, but rather increase it (probably, after
6235 increasing installed memory), if network conditions require more than
6240 <Term>/proc/sys/net/ipv4/tcp_retrans_collapse</Term>
6243 Bug-to-bug compatibility with some broken printers.
6244 On retransmit try to send bigger packets to work around bugs in
6249 <Term>/proc/sys/net/ipv4/tcp_retries1</Term>
6252 How many times to retry before deciding that something is wrong
6253 and it is necessary to report this suspicion to network layer.
6254 Minimal RFC value is 3, it is default, which corresponds
6255 to 3sec-8min depending on RTO.
6259 <Term>/proc/sys/net/ipv4/tcp_retries2</Term>
6262 How may times to retry before killing alive TCP connection.
6264 URL="http://www.ietf.org/rfc/rfc1122.txt"
6267 says that the limit should be longer than 100 sec.
6268 It is too small number. Default value 15 corresponds to 13-30min
6273 <Term>/proc/sys/net/ipv4/tcp_rfc1337</Term>
6276 This boolean enables a fix for 'time-wait assassination hazards in tcp', described
6277 in RFC 1337. If enabled, this causes the kernel to drop RST packets for
6278 sockets in the time-wait state.
6284 <Term>/proc/sys/net/ipv4/tcp_sack</Term>
6287 Use Selective ACK which can be used to signify that specific packets are
6288 missing - therefore helping fast recovery.
6292 <Term>/proc/sys/net/ipv4/tcp_stdurg</Term>
6295 Use the Host requirements interpretation of the TCP urg pointer
6298 Most hosts use the older BSD interpretation, so if you turn this on
6299 Linux might not communicate correctly with them.
6305 <Term>/proc/sys/net/ipv4/tcp_syn_retries</Term>
6308 Number of SYN packets the kernel will send before giving up on the new
6313 <Term>/proc/sys/net/ipv4/tcp_synack_retries</Term>
6316 To open the other side of the connection, the kernel sends a SYN with a
6317 piggybacked ACK on it, to acknowledge the earlier received SYN. This is part
6318 2 of the threeway handshake. This setting determines the number of SYN+ACK
6319 packets sent before the kernel gives up on the connection.
6323 <Term>/proc/sys/net/ipv4/tcp_timestamps</Term>
6326 Timestamps are used, amongst other things, to protect against wrapping
6327 sequence numbers. A 1 gigabit link might conceivably re-encounter a previous
6328 sequence number with an out-of-line value, because it was of a previous
6329 generation. The timestamp will let it recognize this 'ancient packet'.
6333 <Term>/proc/sys/net/ipv4/tcp_tw_recycle</Term>
6336 Enable fast recycling TIME-WAIT sockets. Default value is 1.
6337 It should not be changed without advice/request of technical experts.
6341 <Term>/proc/sys/net/ipv4/tcp_window_scaling</Term>
6344 TCP/IP normally allows windows up to 65535 bytes big. For really fast
6345 networks, this may not be enough. The window scaling options allows for
6346 almost gigabyte windows, which is good for high bandwidth*delay products.
6355 <Title>Per device settings</Title>
6358 DEV can either stand for a real interface, or for 'all' or 'default'.
6359 Default also changes settings for interfaces yet to be created.
6363 <Term>/proc/sys/net/ipv4/conf/DEV/accept_redirects</Term>
6366 If a router decides that you are using it for a wrong purpose (ie, it needs
6367 to resend your packet on the same interface), it will send us a ICMP
6368 Redirect. This is a slight security risk however, so you may want to turn it
6369 off, or use secure redirects.
6373 <Term>/proc/sys/net/ipv4/conf/DEV/accept_source_route</Term>
6376 Not used very much anymore. You used to be able to give a packet a list of
6377 IP addresses it should visit on its way. Linux can be made to honor this IP
6382 <Term>/proc/sys/net/ipv4/conf/DEV/bootp_relay</Term>
6385 Accept packets with source address 0.b.c.d with destinations not to this host
6386 as local ones. It is supposed that a BOOTP relay daemon will catch and forward
6391 The default is 0, since this feature is not implemented yet (kernel version
6396 <Term>/proc/sys/net/ipv4/conf/DEV/forwarding</Term>
6399 Enable or disable IP forwarding on this interface.
6403 <Term>/proc/sys/net/ipv4/conf/DEV/log_martians</Term>
6407 <citetitle><xref linkend="lartc.kernel.rpf"></citetitle>.
6411 <Term>/proc/sys/net/ipv4/conf/DEV/mc_forwarding</Term>
6414 If we do multicast forwarding on this interface
6418 <Term>/proc/sys/net/ipv4/conf/DEV/proxy_arp</Term>
6421 If you set this to 1, this interface will respond to ARP requests for
6422 addresses the kernel has routes to. Can be very useful when building 'ip
6423 pseudo bridges'. Do take care that your netmasks are very correct before
6424 enabling this! Also be aware that the rp_filter, mentioned elsewhere, also
6425 operates on ARP queries!
6429 <Term>/proc/sys/net/ipv4/conf/DEV/rp_filter</Term>
6433 <citetitle><xref linkend="lartc.kernel.rpf"></citetitle>.
6437 <Term>/proc/sys/net/ipv4/conf/DEV/secure_redirects</Term>
6440 Accept ICMP redirect messages only for gateways, listed in default gateway
6441 list. Enabled by default.
6445 <Term>/proc/sys/net/ipv4/conf/DEV/send_redirects</Term>
6448 If we send the above mentioned redirects.
6452 <Term>/proc/sys/net/ipv4/conf/DEV/shared_media</Term>
6455 If it is not set the kernel does not assume that different subnets on this
6456 device can communicate directly. Default setting is 'yes'.
6460 <Term>/proc/sys/net/ipv4/conf/DEV/tag</Term>
6472 <Title>Neighbor policy</Title>
6475 Dev can either stand for a real interface, or for 'all' or 'default'.
6476 Default also changes settings for interfaces yet to be created.
6480 <Term>/proc/sys/net/ipv4/neigh/DEV/anycast_delay</Term>
6483 Maximum for random delay of answers to neighbor solicitation messages in
6484 jiffies (1/100 sec). Not yet implemented (Linux does not have anycast support
6489 <Term>/proc/sys/net/ipv4/neigh/DEV/app_solicit</Term>
6492 Determines the number of requests to send to the user level ARP daemon. Use 0
6497 <Term>/proc/sys/net/ipv4/neigh/DEV/base_reachable_time</Term>
6500 A base value used for computing the random reachable time value as specified
6505 <Term>/proc/sys/net/ipv4/neigh/DEV/delay_first_probe_time</Term>
6508 Delay for the first time probe if the neighbor is reachable. (see
6513 <Term>/proc/sys/net/ipv4/neigh/DEV/gc_stale_time</Term>
6516 Determines how often to check for stale ARP entries. After an ARP entry is
6517 stale it will be resolved again (which is useful when an IP address migrates
6518 to another machine). When ucast_solicit is greater than 0 it first tries to
6519 send an ARP packet directly to the known host When that fails and
6520 mcast_solicit is greater than 0, an ARP request is broadcast.
6524 <Term>/proc/sys/net/ipv4/neigh/DEV/locktime</Term>
6527 An ARP/neighbor entry is only replaced with a new one if the old is at least
6528 locktime old. This prevents ARP cache thrashing.
6532 <Term>/proc/sys/net/ipv4/neigh/DEV/mcast_solicit</Term>
6535 Maximum number of retries for multicast solicitation.
6539 <Term>/proc/sys/net/ipv4/neigh/DEV/proxy_delay</Term>
6542 Maximum time (real time is random [0..proxytime]) before answering to an ARP
6543 request for which we have an proxy ARP entry. In some cases, this is used to
6544 prevent network flooding.
6548 <Term>/proc/sys/net/ipv4/neigh/DEV/proxy_qlen</Term>
6551 Maximum queue length of the delayed proxy arp timer. (see proxy_delay).
6555 <Term>/proc/sys/net/ipv4/neigh/DEV/retrans_time</Term>
6558 The time, expressed in jiffies (1/100 sec), between retransmitted Neighbor
6559 Solicitation messages. Used for address resolution and to determine if a
6560 neighbor is unreachable.
6564 <Term>/proc/sys/net/ipv4/neigh/DEV/ucast_solicit</Term>
6567 Maximum number of retries for unicast solicitation.
6571 <Term>/proc/sys/net/ipv4/neigh/DEV/unres_qlen</Term>
6574 Maximum queue length for a pending arp request - the number of packets which
6575 are accepted from other layers while the ARP address is still resolved.
6579 <Term>Internet QoS: Architectures and Mechanisms for Quality of Service,
6580 Zheng Wang, ISBN 1-55860-608-4</Term>
6583 Hardcover textbook covering topics
6584 related to Quality of Service. Good for understanding basic concepts.
6593 <Title>Routing settings</Title>
6599 <Term>/proc/sys/net/ipv4/route/error_burst</Term>
6602 These parameters are used to limit the warning messages written to the kernel
6603 log from the routing code. The higher the error_cost factor is, the fewer
6604 messages will be written. Error_burst controls when messages will be dropped.
6605 The default settings limit warning messages to one every five seconds.
6609 <Term>/proc/sys/net/ipv4/route/error_cost</Term>
6612 These parameters are used to limit the warning messages written to the kernel
6613 log from the routing code. The higher the error_cost factor is, the fewer
6614 messages will be written. Error_burst controls when messages will be dropped.
6615 The default settings limit warning messages to one every five seconds.
6619 <Term>/proc/sys/net/ipv4/route/flush</Term>
6622 Writing to this file results in a flush of the routing cache.
6626 <Term>/proc/sys/net/ipv4/route/gc_elasticity</Term>
6629 Values to control the frequency and behavior of the garbage collection
6630 algorithm for the routing cache. This can be important for when doing
6631 fail over. At least gc_timeout seconds will elapse before Linux will skip
6632 to another route because the previous one has died. By default set to 300,
6633 you may want to lower it if you want to have a speedy fail over.
6638 URL="http://mailman.ds9a.nl/pipermail/lartc/2002q1/002667.html"
6640 > by Ard van Breemen.
6644 <Term>/proc/sys/net/ipv4/route/gc_interval</Term>
6647 See /proc/sys/net/ipv4/route/gc_elasticity.
6651 <Term>/proc/sys/net/ipv4/route/gc_min_interval</Term>
6654 See /proc/sys/net/ipv4/route/gc_elasticity.
6658 <Term>/proc/sys/net/ipv4/route/gc_thresh</Term>
6661 See /proc/sys/net/ipv4/route/gc_elasticity.
6665 <Term>/proc/sys/net/ipv4/route/gc_timeout</Term>
6668 See /proc/sys/net/ipv4/route/gc_elasticity.
6672 <Term>/proc/sys/net/ipv4/route/max_delay</Term>
6675 Delays for flushing the routing cache.
6679 <Term>/proc/sys/net/ipv4/route/max_size</Term>
6682 Maximum size of the routing cache. Old entries will be purged once the cache
6683 reached has this size.
6687 <Term>/proc/sys/net/ipv4/route/min_adv_mss</Term>
6694 <Term>/proc/sys/net/ipv4/route/min_delay</Term>
6697 Delays for flushing the routing cache.
6701 <Term>/proc/sys/net/ipv4/route/min_pmtu</Term>
6708 <Term>/proc/sys/net/ipv4/route/mtu_expires</Term>
6715 <Term>/proc/sys/net/ipv4/route/redirect_load</Term>
6718 Factors which determine if more ICMP redirects should be sent to a specific
6719 host. No redirects will be sent once the load limit or the maximum number of
6720 redirects has been reached.
6724 <Term>/proc/sys/net/ipv4/route/redirect_number</Term>
6727 See /proc/sys/net/ipv4/route/redirect_load.
6731 <Term>/proc/sys/net/ipv4/route/redirect_silence</Term>
6734 Timeout for redirects. After this period redirects will be sent again, even if
6735 this has been stopped, because the load or number limit has been reached.
6747 <chapter id="lartc.adv-qdisc">
6748 <Title>Advanced & less common queueing disciplines</Title>
6751 Should you find that you have needs not addressed by the queues mentioned
6752 earlier, the kernel contains some other more specialized queues mentioned here.
6755 <Sect1 id="lartc.adv-qdisc.bfifo-pfifo">
6756 <Title><literal>bfifo</literal>/<literal>pfifo</literal></Title>
6759 These classless queues are even simpler than pfifo_fast in that they lack
6760 the internal bands - all traffic is really equal. They have one important
6761 benefit though, they have some statistics. So even if you don't need shaping
6762 or prioritizing, you can use this qdisc to determine the backlog on your
6767 pfifo has a length measured in packets, bfifo in bytes.
6771 <Title>Parameters & usage</Title>
6780 Specifies the length of the queue. Measured in bytes for bfifo, in packets
6781 for pfifo. Defaults to the interface txqueuelen (see pfifo_fast chapter)
6782 packets long or txqueuelen*mtu bytes for bfifo.
6792 <Sect1 id="lartc.adv-qdisc.csz">
6793 <Title>Clark-Shenker-Zhang algorithm (CSZ)</Title>
6796 This is so theoretical that not even Alexey (the main CBQ author) claims to
6797 understand it. From his source:
6802 David D. Clark, Scott Shenker and Lixia Zhang
6803 <citetitle>Supporting Real-Time Applications in an Integrated Services Packet
6804 Network: Architecture and Mechanism</citetitle>.
6808 As I understand it, the main idea is to create WFQ flows for each guaranteed
6809 service and to allocate the rest of bandwith to dummy flow-0. Flow-0
6810 comprises the predictive services and the best effort traffic; it is handled
6811 by a priority scheduler with the highest priority band allocated for
6812 predictive services, and the rest --- to the best effort packets.
6816 Note that in CSZ flows are NOT limited to their bandwidth. It is supposed
6817 that the flow passed admission control at the edge of the QoS network and it
6818 doesn't need further shaping. Any attempt to improve the flow or to shape it
6819 to a token bucket at intermediate hops will introduce undesired delays and
6824 At the moment CSZ is the only scheduler that provides true guaranteed
6825 service. Another schemes (including CBQ) do not provide guaranteed delay and
6830 Does not currently seem like a good candidate to use, unless you've read and
6831 understand the article mentioned.
6837 <Sect1 id="lartc.adv-qdisc.dsmark"
6839 <Title>DSMARK</Title>
6843 <author><firstname>Esteve</firstname><surname>Camps</surname></author>
6844 <address><email>marvin@grn.es</email></address>
6845 This text is an extract from my thesis on
6846 <citetitle>QoS Support in Linux</citetitle>, September 2000.
6850 <Para>Source documents:
6856 <ULink URL="ftp://icaftp.epfl.ch/pub/linux/diffserv/misc/dsid-01.txt.gz">
6857 Draft-almesberger-wajhak-diffserv-linux-01.txt</ULink>.
6861 <Para>Examples in iproute2 distribution.
6866 <ULink URL="http://www.qosforum.com/white-papers/qosprot_v3.pdf">
6867 White Paper-QoS protocols and architectures</ULink> and
6868 <ULink URL="http://www.qosforum.com/docs/faq">
6869 IP QoS Frequently Asked Questions</ULink> both by
6870 <citetitle>Quality of Service Forum</citetitle>.
6876 This chapter was written by Esteve Camps <esteve@hades.udg.es>.
6880 <Title>Introduction</Title>
6883 First of all, first of all, it would be a great idea for you to read RFCs
6884 written about this (RFC2474, RFC2475, RFC2597 and RFC2598) at
6885 <ULink URL="http://www.ietf.org/html.charters/diffserv-charter.html">
6886 IETF DiffServ working Group web site</ULink> and
6887 <ULink URL="http://diffserv.sf.net/">
6888 Werner Almesberger web site</ULink>
6889 (he wrote the code to support Differentiated Services on Linux).
6895 <Title>What is Dsmark related to?</Title>
6898 Dsmark is a queueing discipline that offers the capabilities needed in
6899 Differentiated Services (also called DiffServ or, simply, DS). DiffServ is
6900 one of two actual QoS architectures (the other one is called Integrated
6901 Services) that is based on a value carried by packets in the DS field of the
6906 One of the first solutions in IP designed to offer some QoS level was
6907 the Type of Service field (TOS byte) in IP header. By changing that value,
6908 we could choose a high/low level of throughput, delay or reliability.
6909 But this didn't provide sufficient flexibility to the needs of new
6910 services (such as real-time applications, interactive applications and
6911 others). After this, new architectures appeared. One of these was DiffServ
6912 which kept TOS bits and renamed DS field.
6918 <Title>Differentiated Services guidelines</Title>
6921 Differentiated Services is group-oriented. I mean, we don't know anything
6922 about flows (this will be the Integrated Services purpose); we know about
6923 flow aggregations and we will apply different behaviours depending on which
6924 aggregation a packet belongs to.
6928 When a packet arrives to an edge node (entry node to a DiffServ domain)
6929 entering to a DiffServ Domain we'll have to policy, shape and/or mark those
6930 packets (marking refers to assigning a value to the DS field. It's just like the
6931 cows :-) ). This will be the mark/value that the internal/core nodes on our
6932 DiffServ Domain will look at to determine which behaviour or QoS level
6937 As you can deduce, Differentiated Services involves a domain on which
6938 all DS rules will have to be applied. In fact you can think I
6939 will classify all the packets entering my domain. Once they enter my
6940 domain they will be subjected to the rules that my classification dictates
6941 and every traversed node will apply that QoS level.
6945 In fact, you can apply your own policies into your local domains, but some
6946 <Emphasis>Service Level Agreements</Emphasis> should be considered when connecting to
6951 At this point, you maybe have a lot of questions. DiffServ is more than I've
6952 explained. In fact, you can understand that I can not resume more than 3
6953 RFCs in just 50 lines :-).
6959 <Title>Working with Dsmark</Title>
6962 As the DiffServ bibliography specifies, we differentiate boundary nodes and
6963 interior nodes. These are two important points in the traffic path. Both
6964 types perform a classification when the packets arrive. Its result may be
6965 used in different places along the DS process before the packet is released
6966 to the network. It's just because of this that the diffserv code supplies an
6967 structure called sk_buff, including a new field called skb->tc_index
6968 where we'll store the result of initial classification that may be used in
6969 several points in DS treatment.
6973 The skb->tc_index value will be initially set by the DSMARK qdisc,
6974 retrieving it from the DS field in IP header of every received packet.
6975 Besides, cls_tcindex classifier will read all or part of skb->tcindex
6976 value and use it to select classes.
6980 But, first of all, take a look at DSMARK qdisc command and its parameters:
6983 ... dsmark indices INDICES [ default_index DEFAULT_INDEX ] [ set_tc_index ]
6986 What do these parameters mean?
6992 <Emphasis remap="bf">indices</Emphasis>: size of table of (mask,value) pairs. Maximum value is 2ˆn, where n>=0.
6998 <Emphasis remap="bf">Default_index</Emphasis>: the default table entry index if classifier finds no match.
7004 <Emphasis remap="bf">Set_tc_index</Emphasis>: instructs dsmark discipline to retrieve the DS field and store it onto skb->tc_index.
7010 Let's see the DSMARK process.
7016 <Title>How SCH_DSMARK works.</Title>
7019 This qdisc will apply the next steps:
7025 If we have declared set_tc_index option in qdisc command, DS field is retrieved and stored onto
7026 skb->tc_index variable.
7032 Classifier is invoked. The classifier will be executed and it will return a class ID that will be stored in
7033 skb->tc_index variable.If no filter matches are found, we consider the default_index option to be the
7034 classId to store. If neither set_tc_index nor default_index has been declared results may be
7041 After been sent to internal qdiscs where you can reuse the result of the filter, the classid returned by
7042 the internal qdisc is stored into skb->tc_index. We will use this value in the future to index a mask-
7043 value table. The final result to assign to the packet will be that resulting from next operation:
7046 New_Ds_field = ( Old_DS_field & mask ) | value
7055 Thus, new value will result from "anding" ds_field and mask values and next, this result "ORed" with
7056 value parameter. See next diagram to understand all this process:
7065 - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - >
7067 | -- If you declare set_tc_index, we set DS | | <-----May change
7068 | value into skb->tc_index variable | |O DS field
7070 +-|-+ +------+ +---+-+ Internal +-+ +---N|-----|----+
7071 | | | | tc |--->| | |--> . . . -->| | | D| | |
7072 | | |----->|index |--->| | | Qdisc | |---->| v | |
7073 | | | |filter|--->| | | +---------------+ | ---->(mask,value) |
7074 -->| O | +------+ +-|-+--------------^----+ / | (. , .) |
7075 | | | ^ | | | | (. , .) |
7076 | | +----------|---------|----------------|-------|--+ (. , .) |
7077 | | sch_dsmark | | | | |
7078 +-|------------|---------|----------------|-------|------------------+
7079 | | | <- tc_index -> | |
7080 | |(read) | may change | | <--------------Index to the
7081 | | | | | (mask,value)
7082 v | v v | pairs table
7083 - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - ->
7090 How to do marking? Just change the mask and value of the class you want to remark. See next line of code:
7093 tc class change dev eth0 classid 1:1 dsmark mask 0x3 value 0xb8
7096 This changes the (mask,value) pair in hash table, to remark packets belonging to class 1:1.You have to "change" this values
7097 because of default values that (mask,value) gets initially (see table below).
7101 Now, we'll explain how TC_INDEX filter works and how fits into this. Besides, TCINDEX filter can be
7102 used in other configurations rather than those including DS services.
7108 <Title>TC_INDEX Filter</Title>
7111 This is the basic command to declare a TC_INDEX filter:
7114 ... tcindex [ hash SIZE ] [ mask MASK ] [ shift SHIFT ]
7115 [ pass_on | fall_through ]
7116 [ classid CLASSID ] [ police POLICE_SPEC ]
7119 Next, we show the example used to explain TC_INDEX operation mode. Pay attention to bolded words:
7122 tc qdisc add dev eth0 handle 1:0 root dsmark indices 64 <Emphasis remap="bf">set_tc_index</Emphasis>
7124 tc filter add dev eth0 parent 1:0 protocol ip prio 1 tcindex <Emphasis remap="bf">mask 0xfc shift 2</Emphasis>
7126 tc qdisc add dev eth0 parent 1:0 handle 2:0 cbq bandwidth 10Mbit cell 8 avpkt 1000 mpu 64
7128 # EF traffic class
7130 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
7132 # Packet fifo qdisc for EF traffic
7134 tc qdisc add dev eth0 parent 2:1 pfifo limit 5
7136 tc filter add dev eth0 parent 2:0 protocol ip prio 1 <Emphasis remap="bf">handle 0x2e</Emphasis> tcindex <Emphasis remap="bf">classid 2:1 pass_on</Emphasis>
7140 (This code is not complete. It's just an extract from EFCBQ example included in iproute2 distribution).
7144 First of all, suppose we receive a packet marked as EF . If you read RFC2598, you'll see that DSCP
7145 recommended value for EF traffic is 101110. This means that DS field will be 10111000 (remember that
7146 less significant bits in TOS byte are not used in DS) or 0xb8 in hexadecimal codification.
7154 +---+ +-------+ +---+-+ +------+ +-+ +-------+
7155 | | | | | | | |FILTER| +-+ +-+ | | | |
7156 | |----->| MASK | -> | | | -> |HANDLE|->| | | | -> | | -> | |
7157 | | . | =0xfc | | | | |0x2E | | +----+ | | | | |
7158 | | . | | | | | +------+ +--------+ | | | |
7159 | | . | | | | | | | | |
7160 -->| | . | SHIFT | | | | | | | |-->
7161 | | . | =2 | | | +----------------------------+ | | |
7162 | | | | | | CBQ 2:0 | | |
7163 | | +-------+ +---+--------------------------------+ | |
7165 | +-------------------------------------------------------------+ |
7167 +-------------------------------------------------------------------------+
7174 The packet arrives, then, set with 0xb8 value at DS field. As we explained before, dsmark qdisc identified
7175 by 1:0 id in the example, retrieves DS field and store it in skb->tc_index variable.
7176 Next step in the example will correspond to the filter associated to this qdisc (second line in the example).
7177 This will perform next operations:
7180 Value1 = skb->tc_index & MASK
7181 Key = Value1 >> SHIFT
7187 In the example, MASK=0xFC i SHIFT=2.
7190 Value1 = 10111000 & 11111100 = 10111000
7191 Key = 10111000 >> 2 = 00101110 -> 0x2E in hexadecimal
7197 The returned value will correspond to a qdisc internal filter handle (in the example, identifier 2:0). If a
7198 filter with this id exists, policing and metering conditions will be verified (in case that filter includes this)
7199 and the classid will be returned (in our example, classid 2:1) and stored in skb->tc_index variable.
7203 But if any filter with that identifier is found, the result will depend on fall_through flag declaration. If so,
7204 value key is returned as classid. If not, an error is returned and process continues with the rest filters. Be
7205 careful if you use fall_through flag; this can be done if a simple relation exists between values
7207 of skb->tc_index variable and class id's.
7211 The latest parameters to comment on are hash and pass_on. The first one
7212 relates to hash table size. Pass_on will be used to indicate that if no classid
7213 equal to the result of this filter is found, try next filter.
7214 The default action is fall_through (look at next table).
7218 Finally, let's see which possible values can be set to all this TCINDEX parameters:
7221 TC Name Value Default
7222 -----------------------------------------------------------------
7223 Hash 1...0x10000 Implementation dependent
7224 Mask 0...0xffff 0xffff
7226 Fall through / Pass_on Flag Fall_through
7227 Classid Major:minor None
7234 This kind of filter is very powerful. It's necessary to explore all possibilities. Besides, this filter is not only used in DiffServ configurations.
7235 You can use it as any other kind of filter.
7239 I recommend you to look at all DiffServ examples included in iproute2 distribution. I promise I will try to
7240 complement this text as soon as I can. Besides, all I have explained is the result of a lot of tests.
7241 I would thank you tell me if I'm wrong in any point.
7248 <Sect1 id="lartc.adv-qdisc.ingress">
7249 <Title>Ingress qdisc</Title>
7252 All qdiscs discussed so far are egress qdiscs. Each interface however can
7253 also have an ingress qdisc which is not used to send packets
7254 out to the network adaptor. Instead, it allows you to apply tc filters to
7255 packets coming in over the interface, regardless of whether they have a local
7256 destination or are to be forwarded.
7260 As the tc filters contain a full Token Bucket Filter implementation, and are
7261 also able to match on the kernel flow estimator, there is a lot of
7262 functionality available. This effectively allows you to police incoming
7263 traffic, before it even enters the IP stack.
7267 <Title>Parameters & usage</Title>
7270 The ingress qdisc itself does not require any parameters. It differs from
7271 other qdiscs in that it does not occupy the root of a device. Attach it like
7275 # tc qdisc add dev eth0 ingress
7278 This allows you to have other, sending, qdiscs on your device besides the
7283 For a contrived example how the ingress qdisc could be used, see the
7291 <Sect1 id="lartc.adv-qdisc.red">
7292 <Title>Random Early Detection (RED)</Title>
7295 This section is meant as an introduction to backbone routing, which often
7296 involves <100 megabit bandwidths, which requires a different approach than
7297 your ADSL modem at home.
7301 The normal behaviour of router queues on the Internet is called tail-drop.
7302 Tail-drop works by queueing up to a certain amount, then dropping all traffic
7303 that 'spills over'. This is very unfair, and also leads to retransmit
7304 synchronization. When retransmit synchronization occurs, the sudden burst
7305 of drops from a router that has reached its fill will cause a delayed burst
7306 of retransmits, which will over fill the congested router again.
7310 In order to cope with transient congestion on links, backbone routers will
7311 often implement large queues. Unfortunately, while these queues are good for
7312 throughput, they can substantially increase latency and cause TCP
7313 connections to behave very burstily during congestion.
7317 These issues with tail-drop are becoming increasingly troublesome on the
7318 Internet because the use of network unfriendly applications is increasing.
7319 The Linux kernel offers us RED, short for Random Early Detect, also called
7320 Random Early Drop, as that is how it works.
7324 RED isn't a cure-all for this, applications which inappropriately fail to
7325 implement exponential backoff still get an unfair share of the bandwidth,
7326 however, with RED they do not cause as much harm to the throughput and
7327 latency of other connections.
7331 RED statistically drops packets from flows before it reaches its hard
7332 limit. This causes a congested backbone link to slow more gracefully, and
7333 prevents retransmit synchronization. This also helps TCP find its 'fair'
7334 speed faster by allowing some packets to get dropped sooner keeping queue
7335 sizes low and latency under control. The probability of a packet being
7336 dropped from a particular connection is proportional to its bandwidth usage
7337 rather than the number of packets it transmits.
7341 RED is a good queue for backbones, where you can't afford the
7342 complexity of per-session state tracking needed by fairness queueing.
7346 In order to use RED, you must decide on three parameters: Min, Max, and
7347 burst. Min sets the minimum queue size in bytes before dropping will begin,
7348 Max is a soft maximum that the algorithm will attempt to stay under, and
7349 burst sets the maximum number of packets that can 'burst through'.
7353 You should set the min by calculating that highest acceptable base queueing
7354 latency you wish, and multiply it by your bandwidth. For instance, on my
7355 64kbit/s ISDN link, I might want a base queueing latency of 200ms so I set
7356 min to 1600 bytes. Setting min too small will degrade throughput and too
7357 large will degrade latency. Setting a small min is not a replacement for
7358 reducing the MTU on a slow link to improve interactive response.
7362 You should make max at least twice min to prevent synchronization. On slow
7363 links with small Min's it might be wise to make max perhaps four or
7364 more times large then min.
7368 Burst controls how the RED algorithm responds to bursts. Burst must be set
7369 larger then min/avpkt. Experimentally, I've found (min+min+max)/(3*avpkt) to
7374 Additionally, you need to set limit and avpkt. Limit is a safety value, after
7375 there are limit bytes in the queue, RED 'turns into' tail-drop. I typical set
7376 limit to eight times max. Avpkt should be your average packet size. 1000
7377 works OK on high speed Internet links with a 1500byte MTU.
7382 URL="http://www.aciri.org/floyd/papers/red/red.html"
7383 >the paper on RED queueing</ULink
7384 > by Sally Floyd and Van Jacobson for technical
7390 <Sect1 id="lartc.adv-qdisc.gred">
7391 <Title>Generic Random Early Detection</Title>
7394 Not a lot is known about GRED. It looks like GRED with several internal
7395 queues, whereby the internal queue is chosen based on the Diffserv tcindex
7396 field. According to a slide found
7397 <ULink URL="http://www.davin.ottawa.on.ca/ols/img22.htm">here</ULink>,
7398 it contains the capabilities of Cisco's 'Distributed Weighted RED', as well
7399 as Dave Clark's RIO.
7403 Each virtual queue can have its own Drop Parameters specified.
7407 FIXME: get Jamal or Werner to tell us more
7412 <Sect1 id="lartc.adv-qdisc.vc-atm">
7413 <Title>VC/ATM emulation</Title>
7416 This is quite a major effort by Werner Almesberger to allow you to build
7417 Virtual Circuits over TCP/IP sockets. A Virtual Circuit is a concept from
7422 For more information, see the <ULink
7423 URL="http://linux-atm.sourceforge.net/"
7424 >ATM on Linux homepage</ULink
7430 <Sect1 id="lartc.adv-qdisc.wrr">
7431 <Title>Weighted Round Robin (WRR)</Title>
7434 This qdisc is not included in the standard kernels but can be downloaded from
7435 <ULink URL="http://wipl-wrr.dkik.dk/wrr/">here</ULink>.
7436 Currently the qdisc is only tested with Linux 2.2 kernels but it will
7437 probably work with 2.4/2.5 kernels too.
7441 The WRR qdisc distributes bandwidth between its classes using the weighted
7442 round robin scheme. That is, like the CBQ qdisc it contains classes
7443 into which arbitrary qdiscs can be plugged. All classes which have sufficient
7444 demand will get bandwidth proportional to the weights associated with the classes.
7445 The weights can be set manually using the <Literal remap="tt">tc</Literal> program. But they
7446 can also be made automatically decreasing for classes transferring much data.
7450 The qdisc has a built-in classifier which assigns packets coming from or
7451 sent to different machines to different classes. Either the MAC or IP and
7452 either source or destination addresses can be used. The MAC address can only
7453 be used when the Linux box is acting as an ethernet bridge, however. The
7454 classes are automatically assigned to machines based on the packets seen.
7458 The qdisc can be very useful at sites such as dorms where a lot of unrelated
7459 individuals share an Internet connection. A set of scripts setting up a
7460 relevant behavior for such a site is a central part of the WRR distribution.
7467 <chapter id="lartc.cookbook"
7468 xreflabel="Cookbook">
7469 <Title>Cookbook</Title>
7472 This section contains 'cookbook' entries which may help you solve problems.
7473 A cookbook is no replacement for understanding however, so try and comprehend
7477 <Sect1 id="lartc.cookbook.sla">
7478 <Title>Running multiple sites with different SLAs</Title>
7481 You can do this in several ways. Apache has some support for this with a
7482 module, but we'll show how Linux can do this for you, and do so for other
7483 services as well. These commands are stolen from a presentation by Jamal
7484 Hadi that's referenced below.
7488 Let's say we have two customers, with http, ftp and streaming audio, and we
7489 want to sell them a limited amount of bandwidth. We do so on the server itself.
7493 Customer A should have at most 2 megabits, customer B has paid for 5
7494 megabits. We separate our customers by creating virtual IP addresses on our
7501 # ip address add 188.177.166.1 dev eth0
7502 # ip address add 188.177.166.2 dev eth0
7508 It is up to you to attach the different servers to the right IP address. All
7509 popular daemons have support for this.
7513 We first attach a CBQ qdisc to eth0:
7516 # tc qdisc add dev eth0 root handle 1: cbq bandwidth 10Mbit cell 8 avpkt 1000 \
7523 We then create classes for our customers:
7529 # tc class add dev eth0 parent 1:0 classid 1:1 cbq bandwidth 10Mbit rate \
7530 2MBit avpkt 1000 prio 5 bounded isolated allot 1514 weight 1 maxburst 21
7531 # tc class add dev eth0 parent 1:0 classid 1:2 cbq bandwidth 10Mbit rate \
7532 5Mbit avpkt 1000 prio 5 bounded isolated allot 1514 weight 1 maxburst 21
7538 Then we add filters for our two classes:
7541 ##FIXME: Why this line, what does it do?, what is a divisor?:
7542 ##FIXME: A divisor has something to do with a hash table, and the number of
7544 # tc filter add dev eth0 parent 1:0 protocol ip prio 5 handle 1: u32 divisor 1
7545 # tc filter add dev eth0 parent 1:0 prio 5 u32 match ip src 188.177.166.1
7547 # tc filter add dev eth0 parent 1:0 prio 5 u32 match ip src 188.177.166.2
7558 FIXME: why no token bucket filter? is there a default pfifo_fast fallback
7564 <Sect1 id="lartc.cookbook.synflood-protect"
7565 xreflabel="Protecting your host from SYN floods">
7566 <Title>Protecting your host from SYN floods</Title>
7569 From Alexey's iproute documentation, adapted to netfilter and with more
7570 plausible paths. If you use this, take care to adjust the numbers to
7571 reasonable values for your system.
7575 If you want to protect an entire network, skip this script, which is best
7576 suited for a single host.
7580 It appears that you need the very latest version of the iproute2 tools to
7581 get this to work with 2.4.0.
7589 # sample script on using the ingress capabilities
7590 # this script shows how one can rate limit incoming SYNs
7591 # Useful for TCP-SYN attack protection. You can use
7592 # IPchains to have more powerful additions to the SYN (eg
7593 # in addition the subnet)
7595 #path to various utilities;
7596 #change to reflect yours.
7600 IPTABLES=/sbin/iptables
7603 # tag all incoming SYN packets through $INDEV as mark value 1
7604 ############################################################
7605 $iptables -A PREROUTING -i $INDEV -t mangle -p tcp --syn \
7606 -j MARK --set-mark 1
7607 ############################################################
7609 # install the ingress qdisc on the ingress interface
7610 ############################################################
7611 $TC qdisc add dev $INDEV handle ffff: ingress
7612 ############################################################
7616 # SYN packets are 40 bytes (320 bits) so three SYNs equals
7617 # 960 bits (approximately 1kbit); so we rate limit below
7618 # the incoming SYNs to 3/sec (not very useful really; but
7619 #serves to show the point - JHS
7620 ############################################################
7621 $TC filter add dev $INDEV parent ffff: protocol ip prio 50 handle 1 fw \
7622 police rate 1kbit burst 40 mtu 9k drop flowid :1
7623 ############################################################
7627 echo "---- qdisc parameters Ingress ----------"
7628 $TC qdisc ls dev $INDEV
7629 echo "---- Class parameters Ingress ----------"
7630 $TC class ls dev $INDEV
7631 echo "---- filter parameters Ingress ----------"
7632 $TC filter ls dev $INDEV parent ffff:
7634 #deleting the ingress qdisc
7635 #$TC qdisc del $INDEV ingress
7642 <Sect1 id="lartc.cookbook.icmp-ratelimit">
7643 <Title>Rate limit ICMP to prevent dDoS</Title>
7646 Recently, distributed denial of service attacks have become a major nuisance
7647 on the Internet. By properly filtering and rate limiting your network, you can
7648 both prevent becoming a casualty or the cause of these attacks.
7652 You should filter your networks so that you do not allow non-local IP source
7653 addressed packets to leave your network. This stops people from anonymously
7654 sending junk to the Internet.
7658 Rate limiting goes much as shown earlier. To refresh your memory, our
7665 [The Internet] ---<E3, T3, whatever>--- [Linux router] --- [Office+ISP]
7672 We first set up the prerequisite parts:
7678 # tc qdisc add dev eth0 root handle 10: cbq bandwidth 10Mbit avpkt 1000
7679 # tc class add dev eth0 parent 10:0 classid 10:1 cbq bandwidth 10Mbit rate \
7680 10Mbit allot 1514 prio 5 maxburst 20 avpkt 1000
7686 If you have 100Mbit, or more, interfaces, adjust these numbers. Now you need
7687 to determine how much ICMP traffic you want to allow. You can perform
7688 measurements with tcpdump, by having it write to a file for a while, and
7689 seeing how much ICMP passes your network. Do not forget to raise the
7694 If measurement is impractical, you might want to choose 5% of your available
7695 bandwidth. Let's set up our class:
7698 # tc class add dev eth0 parent 10:1 classid 10:100 cbq bandwidth 10Mbit rate \
7699 100Kbit allot 1514 weight 800Kbit prio 5 maxburst 20 avpkt 250 \
7706 This limits at 100Kbit. Now we need a filter to assign ICMP traffic to this
7710 # tc filter add dev eth0 parent 10:0 protocol ip prio 100 u32 match ip
7711 protocol 1 0xFF flowid 10:100
7719 <Sect1 id="lartc.cookbook.interactive-prio">
7720 <Title>Prioritizing interactive traffic</Title>
7723 If lots of data is coming down your link, or going up for that matter, and
7724 you are trying to do some maintenance via telnet or ssh, this may not go too
7725 well. Other packets are blocking your keystrokes. Wouldn't it be great if
7726 there were a way for your interactive packets to sneak past the bulk
7727 traffic? Linux can do this for you!
7731 As before, we need to handle traffic going both ways. Evidently, this works
7732 best if there are Linux boxes on both ends of your link, although other
7733 UNIX's are able to do this. Consult your local Solaris/BSD guru for this.
7737 The standard pfifo_fast scheduler has 3 different 'bands'. Traffic in band 0
7738 is transmitted first, after which traffic in band 1 and 2 gets considered.
7739 It is vital that our interactive traffic be in band 0!
7743 We blatantly adapt from the (soon to be obsolete) ipchains HOWTO:
7747 There are four seldom-used bits in the IP header, called the Type of Service
7748 (TOS) bits. They effect the way packets are treated; the four bits are
7749 "Minimum Delay", "Maximum Throughput", "Maximum Reliability" and "Minimum
7750 Cost". Only one of these bits is allowed to be set. Rob van Nieuwkerk, the
7751 author of the ipchains TOS-mangling code, puts it as follows:
7757 Especially the "Minimum Delay" is important for me. I switch it on for
7758 "interactive" packets in my upstream (Linux) router. I'm
7759 behind a 33k6 modem link. Linux prioritizes packets in 3 queues. This
7760 way I get acceptable interactive performance while doing bulk
7761 downloads at the same time.
7767 The most common use is to set telnet & ftp control connections to "Minimum
7768 Delay" and FTP data to "Maximum Throughput". This would be
7769 done as follows, on your upstream router:
7775 # iptables -A PREROUTING -t mangle -p tcp --sport telnet \
7776 -j TOS --set-tos Minimize-Delay
7777 # iptables -A PREROUTING -t mangle -p tcp --sport ftp \
7778 -j TOS --set-tos Minimize-Delay
7779 # iptables -A PREROUTING -t mangle -p tcp --sport ftp-data \
7780 -j TOS --set-tos Maximize-Throughput
7786 Now, this only works for data going from your telnet foreign host to your
7787 local computer. The other way around appears to be done for you, ie, telnet,
7788 ssh & friends all set the TOS field on outgoing packets automatically.
7792 Should you have an application that does not do this, you can always do it
7793 with netfilter. On your local box:
7799 # iptables -A OUTPUT -t mangle -p tcp --dport telnet \
7800 -j TOS --set-tos Minimize-Delay
7801 # iptables -A OUTPUT -t mangle -p tcp --dport ftp \
7802 -j TOS --set-tos Minimize-Delay
7803 # iptables -A OUTPUT -t mangle -p tcp --dport ftp-data \
7804 -j TOS --set-tos Maximize-Throughput
7811 <Sect1 id="lartc.cookbook.squid">
7812 <Title>Transparent web-caching using <application>netfilter</application>,
7813 <application>iproute2</application>, <application>ipchains</application> and
7814 <application>squid</application></Title>
7817 This section was sent in by reader Ram Narula from Internet for Education
7822 The regular technique in accomplishing this in Linux
7823 is probably with use of ipchains AFTER making sure
7824 that the "outgoing" port 80(web) traffic gets routed through
7825 the server running squid.
7829 There are 3 common methods to make sure "outgoing"
7830 port 80 traffic gets routed to the server running squid
7831 and 4th one is being introduced here.
7837 <Term>Making the gateway router do it.</Term>
7840 If you can tell your gateway router to
7841 match packets that has outgoing destination port
7842 of 80 to be sent to the IP address of squid server.
7850 This would put additional load on the router and
7851 some commercial routers might not even support this.
7856 <Term>Using a Layer 4 switch.</Term>
7859 Layer 4 switches can handle this without any problem.
7867 The cost for this equipment is usually very high. Typical
7868 layer 4 switch would normally cost more than
7869 a typical router+good linux server.
7874 <Term>Using cache server as network's gateway.</Term>
7877 You can force ALL traffic through cache server.
7883 This is quite risky because Squid does utilize lots of CPU power which might
7884 result in slower over-all network performance or the server itself might crash and no one on the
7885 network will be able to access the Internet if that occurs.
7890 <Term>Linux+NetFilter router.</Term>
7893 By using NetFilter another technique can be implemented
7894 which is using NetFilter for "mark"ing the packets
7895 with destination port 80 and using iproute2 to
7896 route the "mark"ed packets to the Squid server.
7907 10.0.0.1 naret (NetFilter server)
7908 10.0.0.2 silom (Squid server)
7909 10.0.0.3 donmuang (Router connected to the Internet)
7910 10.0.0.4 kaosarn (other server on network)
7912 10.0.0.0/24 main network
7913 10.0.0.0/19 total network
7923 ------------hub/switch----------
7925 naret silom kaosarn RAS etc.
7928 First, make all traffic pass through naret by making sure it is the default gateway except for silom.
7929 Silom's default gateway has to be donmuang (10.0.0.3) or this would create web traffic loop.
7932 (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
7933 was changed the IP address of donmuang to 10.0.0.3 and gave naret ip address of 10.0.0.1)
7939 -setup squid and ipchains
7943 Setup Squid server on silom, make sure it does support transparent caching/proxying, the default port is usually
7944 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:
7948 silom# ipchains -N allow1
7949 silom# ipchains -A allow1 -p TCP -s 10.0.0.0/19 -d 0/0 80 -j REDIRECT 3128
7950 silom# ipchains -I input -j allow1
7954 Or, in netfilter lingo:
7956 silom# iptables -t nat -A PREROUTING -i eth0 -p tcp --dport 80 -j REDIRECT --to-port 3128
7960 (note: you might have other entries as well)
7963 For more information on setting Squid server please refer to Squid FAQ page on <ULink
7964 URL="http://squid.nlanr.net">http://squid.nlanr.net</ULink>).
7967 Make sure ip forwarding is enabled on this server and the default gateway for this server is donmuang router (NOT naret).
7973 -setup iptables and iproute2
7974 -disable icmp REDIRECT messages (if needed)
7981 "Mark" packets of destination port 80 with value 2
7983 naret# iptables -A PREROUTING -i eth0 -t mangle -p tcp --dport 80 \
7984 -j MARK --set-mark 2
7990 Setup iproute2 so it will route packets with "mark" 2 to silom
7992 naret# echo 202 www.out >> /etc/iproute2/rt_tables
7993 naret# ip rule add fwmark 2 table www.out
7994 naret# ip route add default via 10.0.0.2 dev eth0 table www.out
7995 naret# ip route flush cache
7999 If donmuang and naret is on the same subnet then naret should not send out icmp REDIRECT messages.
8000 In this case it is, so icmp REDIRECTs has to be disabled by:
8002 naret# echo 0 > /proc/sys/net/ipv4/conf/all/send_redirects
8003 naret# echo 0 > /proc/sys/net/ipv4/conf/default/send_redirects
8004 naret# echo 0 > /proc/sys/net/ipv4/conf/eth0/send_redirects
8011 The setup is complete, check the configuration
8017 naret# iptables -t mangle -L
8018 Chain PREROUTING (policy ACCEPT)
8019 target prot opt source destination
8020 MARK tcp -- anywhere anywhere tcp dpt:www MARK set 0x2
8022 Chain OUTPUT (policy ACCEPT)
8023 target prot opt source destination
8026 0: from all lookup local
8027 32765: from all fwmark 2 lookup www.out
8028 32766: from all lookup main
8029 32767: from all lookup default
8031 naret# ip route list table www.out
8032 default via 203.114.224.8 dev eth0
8035 10.0.0.1 dev eth0 scope link
8036 10.0.0.0/24 dev eth0 proto kernel scope link src 10.0.0.1
8037 127.0.0.0/8 dev lo scope link
8038 default via 10.0.0.3 dev eth0
8040 (make sure silom belongs to one of the above lines, in this case
8041 it's the line with 10.0.0.0/24)
8049 <Title>Traffic flow diagram after implementation</Title>
8053 |-----------------------------------------|
8054 |Traffic flow diagram after implementation|
8055 |-----------------------------------------|
8061 -----------------donmuang router---------------------
8066 *destination port 80 traffic=========>(cache) ||
8069 \\===================================kaosarn, RAS, etc.
8074 Note that the network is asymmetric as there is one extra hop on
8075 general outgoing path.
8081 Here is run down for packet traversing the network from kaosarn
8082 to and from the Internet.
8084 For web/http traffic:
8085 kaosarn http request->naret->silom->donmuang->internet
8086 http replies from Internet->donmuang->silom->kaosarn
8088 For non-web/http requests(eg. telnet):
8089 kaosarn outgoing data->naret->donmuang->internet
8090 incoming data from Internet->donmuang->kaosarn
8099 <Sect1 id="lartc.cookbook.mtu-discovery">
8100 <Title>Circumventing Path MTU Discovery issues with per route MTU settings</Title>
8103 For sending bulk data, the Internet generally works better when using larger
8104 packets. Each packet implies a routing decision, when sending a 1 megabyte
8105 file, this can either mean around 700 packets when using packets that are as
8106 large as possible, or 4000 if using the smallest default.
8110 However, not all parts of the Internet support full 1460 bytes of payload
8111 per packet. It is therefore necessary to try and find the largest packet
8112 that will 'fit', in order to optimize a connection.
8116 This process is called 'Path MTU Discovery', where MTU stands for 'Maximum
8121 When a router encounters a packet that's too big too send in one piece, AND
8122 it has been flagged with the "Don't Fragment" bit, it returns an ICMP
8123 message stating that it was forced to drop a packet because of this. The
8124 sending host acts on this hint by sending smaller packets, and by iterating
8125 it can find the optimum packet size for a connection over a certain path.
8129 This used to work well until the Internet was discovered by hooligans who do
8130 their best to disrupt communications. This in turn lead administrators to
8131 either block or shape ICMP traffic in a misguided attempt to improve
8132 security or robustness of their Internet service.
8136 What has happened now is that Path MTU Discovery is working less and less
8137 well and fails for certain routes, which leads to strange TCP/IP sessions
8138 which die after a while.
8142 Although I have no proof for this, two sites who I used to have this problem
8143 with both run Alteon Acedirectors before the affected systems - perhaps
8144 somebody more knowledgeable can provide clues as to why this happens.
8148 <Title>Solution</Title>
8151 When you encounter sites that suffer from this problem, you can disable Path
8152 MTU discovery by setting it manually. Koos van den Hout, slightly edited,
8158 The following problem: I set the mtu/mru of my leased line running ppp to
8159 296 because it's only 33k6 and I cannot influence the queueing on the
8160 other side. At 296, the response to a key press is within a reasonable
8165 And, on my side I have a masqrouter running (of course) Linux.
8169 Recently I split 'server' and 'router' so most applications are run on a
8170 different machine than the routing happens on.
8174 I then had trouble logging into irc. Big panic! Some digging did find
8175 out that I got connected to irc, even showed up as 'connected' on irc
8176 but I did not receive the motd from irc. I checked what could be wrong
8177 and noted that I already had some previous trouble reaching certain
8178 websites related to the MTU, since I had no trouble reaching them when
8179 the MTU was 1500, the problem just showed when the MTU was set to 296.
8180 Since irc servers block about every kind of traffic not needed for their
8181 immediate operation, they also block icmp.
8185 I managed to convince the operators of a webserver that this was the cause
8186 of a problem, but the irc server operators were not going to fix this.
8190 So, I had to make sure outgoing masqueraded traffic started with the lower
8191 mtu of the outside link. But I want local ethernet traffic to have the
8192 normal mtu (for things like nfs traffic).
8199 ip route add default via 10.0.0.1 mtu 296
8203 (10.0.0.1 being the default gateway, the inside address of the
8204 masquerading router)
8209 In general, it is possible to override PMTU Discovery by setting specific
8210 routes. For example, if only a certain subnet is giving problems, this
8215 ip route add 195.96.96.0/24 via 10.0.0.1 mtu 1000
8222 <Sect1 id="lartc.cookbook.mtu-mss">
8223 <Title>Circumventing Path MTU Discovery issues with MSS Clamping
8224 (for ADSL, cable, PPPoE & PPtP users)</Title>
8227 As explained above, Path MTU Discovery doesn't work as well as it should
8228 anymore. If you know for a fact that a hop somewhere in your network has a
8229 limited (<1500) MTU, you cannot rely on PMTU Discovery finding this out.
8233 Besides MTU, there is yet another way to set the maximum packet size, the so
8234 called Maximum Segment Size. This is a field in the TCP Options part of a
8239 Recent Linux kernels, and a few PPPoE drivers (notably, the excellent
8240 Roaring Penguin one), feature the possibility to 'clamp the MSS'.
8244 The good thing about this is that by setting the MSS value, you are telling
8245 the remote side unequivocally 'do not ever try to send me packets bigger
8246 than this value'. No ICMP traffic is needed to get this to work.
8250 The bad thing is that it's an obvious hack - it breaks 'end to end' by
8251 modifying packets. Having said that, we use this trick in many places and it
8256 In order for this to work you need at least iptables-1.2.1a and Linux 2.4.3
8257 or higher. The basic command line is:
8260 # iptables -A FORWARD -p tcp --tcp-flags SYN,RST SYN -j TCPMSS --clamp-mss-to-pmtu
8266 This calculates the proper MSS for your link. If you are feeling brave, or
8267 think that you know best, you can also do something like this:
8273 # iptables -A FORWARD -p tcp --tcp-flags SYN,RST SYN -j TCPMSS --set-mss 128
8279 This sets the MSS of passing SYN packets to 128. Use this if you have VoIP
8280 with tiny packets, and huge http packets which are causing chopping in your
8286 <Sect1 id="lartc.cookbook.ultimate-tc">
8287 <Title>The Ultimate Traffic Conditioner: Low Latency, Fast Up & Downloads</Title>
8290 Note: This script has recently been upgraded and previously only worked for
8291 Linux clients in your network! So you might want to update if you have
8292 Windows machines or Macs in your network and noticed that they were not able
8293 to download faster while others were uploading.
8297 I attempted to create the holy grail:
8301 <Term>Maintain low latency for interactive traffic at all times</Term>
8304 This means that downloading or uploading files should not disturb SSH or
8305 even telnet. These are the most important things, even 200ms latency is
8306 sluggish to work over.
8310 <Term>Allow 'surfing' at reasonable speeds while up or downloading</Term>
8313 Even though http is 'bulk' traffic, other traffic should not drown it out
8318 <Term>Make sure uploads don't harm downloads, and the other way around</Term>
8321 This is a much observed phenomenon where upstream traffic simply destroys
8326 It turns out that all this is possible, at the cost of a tiny bit of
8327 bandwidth. The reason that uploads, downloads and ssh hurt each other is the
8328 presence of large queues in many domestic access devices like cable or DSL
8333 The next section explains in depth what causes the delays, and how we can
8334 fix them. You can safely skip it and head straight for the script if you
8335 don't care how the magic is performed.
8339 <Title>Why it doesn't work well by default</Title>
8342 ISPs know that they are benchmarked solely on how fast people can download.
8343 Besides available bandwidth, download speed is influenced heavily by packet
8344 loss, which seriously hampers TCP/IP performance. Large queues can help
8345 prevent packet loss, and speed up downloads. So ISPs configure large queues.
8349 These large queues however damage interactivity. A keystroke must first
8350 travel the upstream queue, which may be seconds (!) long and go to your
8351 remote host. It is then displayed, which leads to a packet coming back, which
8352 must then traverse the downstream queue, located at your ISP, before it
8353 appears on your screen.
8357 This HOWTO teaches you how to mangle and process the queue in many ways, but
8358 sadly, not all queues are accessible to us. The queue over at the ISP is
8359 completely off-limits, whereas the upstream queue probably lives inside your
8360 cable modem or DSL device. You may or may not be able to configure it. Most
8365 So, what next? As we can't control either of those queues, they must be
8366 eliminated, and moved to your Linux router. Luckily this is possible.
8373 <Term>Limit upload speed</Term>
8376 By limiting our upload speed to slightly less than the truly available rate,
8377 no queues are built up in our modem. The queue is now moved to Linux.
8381 <Term>Limit download speed</Term>
8384 This is slightly trickier as we can't really influence how fast the internet
8385 ships us data. We can however drop packets that are coming in too fast,
8386 which causes TCP/IP to slow down to just the rate we want. Because we don't
8387 want to drop traffic unnecessarily, we configure a 'burst' size we allow at
8395 Now, once we have done this, we have eliminated the downstream queue totally
8396 (except for short bursts), and gain the ability to manage the upstream queue
8397 with all the power Linux offers.
8401 What remains to be done is to make sure interactive traffic jumps to the
8402 front of the upstream queue. To make sure that uploads don't hurt downloads,
8403 we also move ACK packets to the front of the queue. This is what normally
8404 causes the huge slowdown observed when generating bulk traffic both ways.
8405 The ACKnowledgements for downstream traffic must compete with upstream
8406 traffic, and get delayed in the process.
8410 If we do all this we get the following measurements using an excellent ADSL
8411 connection from xs4all in the Netherlands:
8418 round-trip min/avg/max = 14.4/17.1/21.7 ms
8420 Without traffic conditioner, while downloading:
8421 round-trip min/avg/max = 560.9/573.6/586.4 ms
8423 Without traffic conditioner, while uploading:
8424 round-trip min/avg/max = 2041.4/2332.1/2427.6 ms
8426 With conditioner, during 220kbit/s upload:
8427 round-trip min/avg/max = 15.7/51.8/79.9 ms
8429 With conditioner, during 850kbit/s download:
8430 round-trip min/avg/max = 20.4/46.9/74.0 ms
8432 When uploading, downloads proceed at ~80% of the available speed. Uploads
8433 at around 90%. Latency then jumps to 850 ms, still figuring out why.
8439 What you can expect from this script depends a lot on your actual uplink
8440 speed. When uploading at full speed, there will always be a single packet
8441 ahead of your keystroke. That is the lower limit to the latency you can
8442 achieve - divide your MTU by your upstream speed to calculate. Typical
8443 values will be somewhat higher than that. Lower your MTU for better effects!
8447 Next, two versions of this script, one with Devik's excellent HTB, the other
8448 with CBQ which is in each Linux kernel, unlike HTB. Both are tested and work
8455 <Title>The actual script (CBQ)</Title>
8458 Works on all kernels. Within the CBQ
8459 qdisc we place two Stochastic Fairness Queues that make sure that multiple
8460 bulk streams don't drown each other out.
8464 Downstream traffic is policed using a tc filter containing a Token Bucket
8469 You might improve on this script by adding 'bounded' to the line that starts
8470 with 'tc class add .. classid 1:20'. If you lowered your MTU, also lower the
8471 allot & avpkt numbers!
8479 # The Ultimate Setup For Your Internet Connection At Home
8482 # Set the following values to somewhat less than your actual download
8483 # and uplink speed. In kilobits
8488 # clean existing down- and uplink qdiscs, hide errors
8489 tc qdisc del dev $DEV root 2> /dev/null > /dev/null
8490 tc qdisc del dev $DEV ingress 2> /dev/null > /dev/null
8496 tc qdisc add dev $DEV root handle 1: cbq avpkt 1000 bandwidth 10mbit
8498 # shape everything at $UPLINK speed - this prevents huge queues in your
8499 # DSL modem which destroy latency:
8502 tc class add dev $DEV parent 1: classid 1:1 cbq rate ${UPLINK}kbit \
8503 allot 1500 prio 5 bounded isolated
8505 # high prio class 1:10:
8507 tc class add dev $DEV parent 1:1 classid 1:10 cbq rate ${UPLINK}kbit \
8508 allot 1600 prio 1 avpkt 1000
8510 # bulk and default class 1:20 - gets slightly less traffic,
8511 # and a lower priority:
8513 tc class add dev $DEV parent 1:1 classid 1:20 cbq rate $[9*$UPLINK/10]kbit \
8514 allot 1600 prio 2 avpkt 1000
8516 # both get Stochastic Fairness:
8517 tc qdisc add dev $DEV parent 1:10 handle 10: sfq perturb 10
8518 tc qdisc add dev $DEV parent 1:20 handle 20: sfq perturb 10
8521 # TOS Minimum Delay (ssh, NOT scp) in 1:10:
8522 tc filter add dev $DEV parent 1:0 protocol ip prio 10 u32 \
8523 match ip tos 0x10 0xff flowid 1:10
8525 # ICMP (ip protocol 1) in the interactive class 1:10 so we
8526 # can do measurements & impress our friends:
8527 tc filter add dev $DEV parent 1:0 protocol ip prio 11 u32 \
8528 match ip protocol 1 0xff flowid 1:10
8530 # To speed up downloads while an upload is going on, put ACK packets in
8531 # the interactive class:
8533 tc filter add dev $DEV parent 1: protocol ip prio 12 u32 \
8534 match ip protocol 6 0xff \
8535 match u8 0x05 0x0f at 0 \
8536 match u16 0x0000 0xffc0 at 2 \
8537 match u8 0x10 0xff at 33 \
8540 # rest is 'non-interactive' ie 'bulk' and ends up in 1:20
8542 tc filter add dev $DEV parent 1: protocol ip prio 13 u32 \
8543 match ip dst 0.0.0.0/0 flowid 1:20
8545 ########## downlink #############
8546 # slow downloads down to somewhat less than the real speed to prevent
8547 # queuing at our ISP. Tune to see how high you can set it.
8548 # ISPs tend to have *huge* queues to make sure big downloads are fast
8550 # attach ingress policer:
8552 tc qdisc add dev $DEV handle ffff: ingress
8554 # filter *everything* to it (0.0.0.0/0), drop everything that's
8555 # coming in too fast:
8557 tc filter add dev $DEV parent ffff: protocol ip prio 50 u32 match ip src \
8558 0.0.0.0/0 police rate ${DOWNLINK}kbit burst 10k drop flowid :1
8561 If you want this script to be run by ppp on connect, copy it to
8566 If the last two lines give an error, update your tc tool to a newer version!
8572 <Title>The actual script (HTB)</Title>
8575 The following script achieves all goals using the wonderful HTB queue, see
8576 the relevant chapter. Well worth patching your kernel for!
8581 # The Ultimate Setup For Your Internet Connection At Home
8584 # Set the following values to somewhat less than your actual download
8585 # and uplink speed. In kilobits
8590 # clean existing down- and uplink qdiscs, hide errors
8591 tc qdisc del dev $DEV root 2> /dev/null > /dev/null
8592 tc qdisc del dev $DEV ingress 2> /dev/null > /dev/null
8596 # install root HTB, point default traffic to 1:20:
8598 tc qdisc add dev $DEV root handle 1: htb default 20
8600 # shape everything at $UPLINK speed - this prevents huge queues in your
8601 # DSL modem which destroy latency:
8603 tc class add dev $DEV parent 1: classid 1:1 htb rate ${UPLINK}kbit burst 6k
8605 # high prio class 1:10:
8607 tc class add dev $DEV parent 1:1 classid 1:10 htb rate ${UPLINK}kbit \
8610 # bulk & default class 1:20 - gets slightly less traffic,
8611 # and a lower priority:
8613 tc class add dev $DEV parent 1:1 classid 1:20 htb rate $[9*$UPLINK/10]kbit \
8616 # both get Stochastic Fairness:
8617 tc qdisc add dev $DEV parent 1:10 handle 10: sfq perturb 10
8618 tc qdisc add dev $DEV parent 1:20 handle 20: sfq perturb 10
8620 # TOS Minimum Delay (ssh, NOT scp) in 1:10:
8621 tc filter add dev $DEV parent 1:0 protocol ip prio 10 u32 \
8622 match ip tos 0x10 0xff flowid 1:10
8624 # ICMP (ip protocol 1) in the interactive class 1:10 so we
8625 # can do measurements & impress our friends:
8626 tc filter add dev $DEV parent 1:0 protocol ip prio 10 u32 \
8627 match ip protocol 1 0xff flowid 1:10
8629 # To speed up downloads while an upload is going on, put ACK packets in
8630 # the interactive class:
8632 tc filter add dev $DEV parent 1: protocol ip prio 10 u32 \
8633 match ip protocol 6 0xff \
8634 match u8 0x05 0x0f at 0 \
8635 match u16 0x0000 0xffc0 at 2 \
8636 match u8 0x10 0xff at 33 \
8639 # rest is 'non-interactive' ie 'bulk' and ends up in 1:20
8642 ########## downlink #############
8643 # slow downloads down to somewhat less than the real speed to prevent
8644 # queuing at our ISP. Tune to see how high you can set it.
8645 # ISPs tend to have *huge* queues to make sure big downloads are fast
8647 # attach ingress policer:
8649 tc qdisc add dev $DEV handle ffff: ingress
8651 # filter *everything* to it (0.0.0.0/0), drop everything that's
8652 # coming in too fast:
8654 tc filter add dev $DEV parent ffff: protocol ip prio 50 u32 match ip src \
8655 0.0.0.0/0 police rate ${DOWNLINK}kbit burst 10k drop flowid :1
8661 If you want this script to be run by ppp on connect, copy it to
8666 If the last two lines give an error, update your tc tool to a newer version!
8672 <sect1 id="lartc.ratelimit.single"><title>Rate limiting a single host or netmask</title>
8674 Although this is described in stupendous details elsewhere and in our manpages, this question gets asked a lot and
8675 happily there is a simple answer that does not need full comprehension of traffic control.
8678 This three line script does the trick:
8682 tc qdisc add dev $DEV root handle 1: cbq avpkt 1000 bandwidth 10mbit
8684 tc class add dev $DEV parent 1: classid 1:1 cbq rate 512kbit \
8685 allot 1500 prio 5 bounded isolated
8687 tc filter add dev $DEV parent 1: protocol ip prio 16 u32 \
8688 match ip dst 195.96.96.97 flowid 1:1
8692 The first line installs a class based queue on your interface, and tells the kernel that for calculations,
8693 it can be assumed to be a 10mbit interface. If you get this wrong, no real harm is done. But getting it right will
8694 make everything more precise.
8697 The second line creates a 512kbit class with some reasonable defaults. For details, see the cbq manpages and
8698 <xref linkend="lartc.qdisc">.
8701 The last line tells which traffic should go to the shaped class. Traffic not matched by this rule is NOT shaped. To make more
8702 complicated matches (subnets, source ports, destination ports), see <xref linkend="lartc.filtering.simple">.
8705 If you changed anything and want to reload the script, execute 'tc qdisc del dev $DEV root' to clean up your existing
8709 The script can further be improved by adding a last optional line 'tc qdisc add dev $DEV parent 1:1 sfq perturb 10'. See
8710 <xref linkend="lartc.sfq"> for details on what this does.
8714 <sect1 id="lartc.cookbook.fullnat.intro"><title>Example of a full nat solution with QoS</title>
8716 I'm Pedro Larroy <address><email>piotr@omega.resa.es</email></address>. Here I'm describing a common set up where we have lots of users in a private network connected to the Internet trough a Linux router with a public ip address that is doing network address translation (NAT). I use this QoS setup to give access to the Internet to 198 users in a university dorm, in which I live and I'm netadmin of. The users here do heavy use of peer to peer programs, so proper traffic control is a must. I hope this serves as a practical example for all interested lartc readers.
8720 At first I make a practical approach with step by step configuration, and in the end I explain how to make the process automatic at bootime. The network to which this example applies is a private LAN connected to the Internet through a Linux router which has one public ip address. Extending it to several public ip address should be very easy, a couple of iptables rules should be added.
8721 In order to get things working we need:
8724 <Term>Linux 2.4.18 or higher kernel version installed</Term>
8727 If you use 2.4.18 you will have to apply HTB patch available here.
8732 <Term>iproute</Term>
8735 Also ensure the "tc" binary is HTB ready, a precompiled binary is distributed with HTB.
8740 <Term>iptables</Term>
8751 <Title>Let's begin optimizing that scarce bandwidth</Title>
8753 First we set up some qdiscs in which we will classify the traffic. We create a htb qdisc with 6 classes with ascending priority. Then we have classes that will always get allocated rate, but can use the unused bandwidth that other classes don't need. Recall that classes with higher priority ( i.e with a lower prio number ) will get excess of bandwith allocated first. Our connection is 2Mb down 300kbits/s up Adsl. I use 240kbit/s as ceil rate just because it's the higher I can set it before latency starts to grow, due to buffer filling in whatever place between us and remote hosts. This parameter should be timed experimentally, raising it and lowering while observing latency between some near hosts.
8756 Adjust CEIL to 75% of your upstream bandwith limit by now, and where I use eth0, you should use the interface which has a public Internet address. To begin our example execute the following in a root shell:
8759 tc qdisc add dev eth0 root handle 1: htb default 15
8760 tc class add dev eth0 parent 1: classid 1:1 htb rate ${CEIL}kbit ceil ${CEIL}kbit
8761 tc class add dev eth0 parent 1:1 classid 1:10 htb rate 80kbit ceil 80kbit prio 0
8762 tc class add dev eth0 parent 1:1 classid 1:11 htb rate 80kbit ceil ${CEIL}kbit prio 1
8763 tc class add dev eth0 parent 1:1 classid 1:12 htb rate 20kbit ceil ${CEIL}kbit prio 2
8764 tc class add dev eth0 parent 1:1 classid 1:13 htb rate 20kbit ceil ${CEIL}kbit prio 2
8765 tc class add dev eth0 parent 1:1 classid 1:14 htb rate 10kbit ceil ${CEIL}kbit prio 3
8766 tc class add dev eth0 parent 1:1 classid 1:15 htb rate 30kbit ceil ${CEIL}kbit prio 3
8767 tc qdisc add dev eth0 parent 1:12 handle 120: sfq perturb 10
8768 tc qdisc add dev eth0 parent 1:13 handle 130: sfq perturb 10
8769 tc qdisc add dev eth0 parent 1:14 handle 140: sfq perturb 10
8770 tc qdisc add dev eth0 parent 1:15 handle 150: sfq perturb 10
8772 We have just created a htb tree with one level depth. Something like this:
8779 +---------------------------------------+
8781 +---------------------------------------+
8783 +----+ +----+ +----+ +----+ +----+ +----+
8784 |1:10| |1:11| |1:12| |1:13| |1:14| |1:15|
8785 +----+ +----+ +----+ +----+ +----+ +----+
8790 <Term>classid 1:10 htb rate 80kbit ceil 80kbit prio 0</Term>
8793 This is the higher priority class. The packets in this class will have the lowest delay and would get the excess of bandwith first so it's a good idea to limit the ceil rate to this class. We will send through this class the following packets that benefit from low delay, such as interactive traffic: <emphasis>ssh, telnet, dns, quake3, irc, and packets with the SYN flag</emphasis>.
8799 <Term>classid 1:11 htb rate 80kbit ceil ${CEIL}kbit prio 1</Term>
8802 Here we have the first class in which we can start to put bulk traffic. In my example I have traffic from the local web server and requests for web pages: source port 80, and destination port 80 respectively.
8807 <Term>classid 1:12 htb rate 20kbit ceil ${CEIL}kbit prio 2</Term>
8810 In this class I will put traffic with Maximize-Throughput TOS bit set and the rest of the traffic that goes from <emphasis>local processes</emphasis> on the router to the Internet. So the following classes will only have traffic that is <quote>routed through</quote> the box.
8815 <Term>classid 1:13 htb rate 20kbit ceil ${CEIL}kbit prio 2</Term>
8818 This class is for the traffic of other NATed machines that need higher priority in their bulk traffic.
8824 <Term>classid 1:14 htb rate 10kbit ceil ${CEIL}kbit prio 3</Term>
8827 Here goes mail traffic (SMTP,pop3...) and packets with Minimize-Cost TOS bit set.
8832 <Term>classid 1:15 htb rate 30kbit ceil ${CEIL}kbit prio 3</Term>
8835 And finally here we have bulk traffic from the NATed machines behind the router. All kazaa, edonkey, and others will go here, in order to not interfere with other services.
8848 <Title>Classifying packets</Title>
8850 We have created the qdisc setup but no packet classification has been made, so now all outgoing packets are going out in class 1:15 ( because we used: tc qdisc add dev eth0 root handle 1: htb <emphasis>default 15</emphasis> ). Now we need to tell which packets go where. This is the most important part.
8854 Now we set the filters so we can classify the packets with iptables. I really prefer to do it with iptables, because they are very flexible and you have packet count for each rule. Also with the RETURN target packets don't need to traverse all rules. We execute the following commands:
8856 tc filter add dev eth0 parent 1:0 protocol ip prio 1 handle 1 fw classid 1:10
8857 tc filter add dev eth0 parent 1:0 protocol ip prio 2 handle 2 fw classid 1:11
8858 tc filter add dev eth0 parent 1:0 protocol ip prio 3 handle 3 fw classid 1:12
8859 tc filter add dev eth0 parent 1:0 protocol ip prio 4 handle 4 fw classid 1:13
8860 tc filter add dev eth0 parent 1:0 protocol ip prio 5 handle 5 fw classid 1:14
8861 tc filter add dev eth0 parent 1:0 protocol ip prio 6 handle 6 fw classid 1:15
8863 We have just told the kernel that packets that has a specific FWMARK value ( hanlde x fw ) go in the specified class ( classid x:x). Next you will see how to mark packets with iptables.
8867 First you have to understand how packet traverse the filters with iptables:
8869 +------------+ +---------+ +-------------+
8870 Packet -| PREROUTING |--- routing-----| FORWARD |-------+-------| POSTROUTING |- Packets
8871 input +------------+ decision +-Â-------+ | +-------------+ out
8873 +-------+ +--------+
8874 | INPUT |---- Local process -| OUTPUT |
8875 +-------+ +--------+
8878 I assume you have all your tables creak and with default policy ACCEPT ( -P ACCEPT ) if you haven't poked with iptables yet, It should be ok by default. Ours private network is a class B with address 172.17.0.0/16 and public ip is 212.170.21.172
8882 Next we instruct the kernel to <emphasis>actually do NAT</emphasis>, so clients in the private network can start talking to the outside.
8884 echo 1 > /proc/sys/net/ipv4/ip_forward
8885 iptables -t nat -A POSTROUTING -s 172.17.0.0/255.255.0.0 -o eth0 -j SNAT --to-source 212.170.21.172
8887 Now check that packets are flowing through 1:15:
8889 tc -s class show dev eth0
8894 You can start marking packets adding rules to the PREROUTING chain in the mangle table.
8896 iptables -t mangle -A PREROUTING -p icmp -j MARK --set-mark 0x1
8897 iptables -t mangle -A PREROUTING -p icmp -j RETURN
8899 Now you should be able to see packet count increasing when pinging from machines within the private network to some site on the Internet. Check packet count increasing in 1:10
8901 tc -s class show dev eth0
8903 We have done a -j RETURN so packets don't traverse all rules. Icmp packets won't match other rules below RETURN. Keep that in mind.
8904 Now we can start adding more rules, lets do proper TOS handling:
8906 iptables -t mangle -A PREROUTING -m tos --tos Minimize-Delay -j MARK --set-mark 0x1
8907 iptables -t mangle -A PREROUTING -m tos --tos Minimize-Delay -j RETURN
8908 iptables -t mangle -A PREROUTING -m tos --tos Minimize-Cost -j MARK --set-mark 0x5
8909 iptables -t mangle -A PREROUTING -m tos --tos Minimize-Cost -j RETURN
8910 iptables -t mangle -A PREROUTING -m tos --tos Maximize-Throughput -j MARK --set-mark 0x6
8911 iptables -t mangle -A PREROUTING -m tos --tos Maximize-Throughput -j RETURN
8914 Now prioritize ssh packets:
8916 iptables -t mangle -A PREROUTING -p tcp -m tcp --sport 22 -j MARK --set-mark 0x1
8917 iptables -t mangle -A PREROUTING -p tcp -m tcp --sport 22 -j RETURN
8919 A good idea is to prioritize packets to begin tcp connections, those with SYN flag set:
8921 iptables -t mangle -I PREROUTING -p tcp -m tcp --tcp-flags SYN,RST,ACK SYN -j MARK --set-mark 0x1
8922 iptables -t mangle -I PREROUTING -p tcp -m tcp --tcp-flags SYN,RST,ACK SYN -j RETURN
8926 When we are done adding rules to PREROUTING in mangle, we terminate the PREROUTING table with:
8928 iptables -t mangle -A PREROUTING -j MARK --set-mark 0x6
8930 So previously unmarked traffic goes in 1:15. In fact this last step is unnecessary since default class was 1:15, but I will mark them in order to be consistent with the whole setup, and furthermore it's useful to see the counter in that rule.
8934 It will be a good idea to do the same in the OUTPUT rule, so repeat those commands with -A OUTPUT instead of PREROUTING. ( s/PREROUTING/OUTPUT/ ). Then traffic generated locally (on the Linux router) will also be classified. I finish OUTPUT chain with -j MARK --set-mark 0x3 so local traffic has higher priority.
8940 <Title>Improving our setup</Title>
8942 Now we have all our setup working. Take time looking at the graphs, and watching where your bandwith is spent and how do you want it. Doing that for lots of hours, I finally got the Internet connection working really well. Otherwise continuous timeouts and nearly zero allotment of bandwith to newly created tcp connections will occur.
8945 If you find that some classes are full most of the time it would be a good idea to attach another queueing discipline to them so bandwith sharing is more fair:
8947 tc qdisc add dev eth0 parent 1:13 handle 130: sfq perturb 10
8948 tc qdisc add dev eth0 parent 1:14 handle 140: sfq perturb 10
8949 tc qdisc add dev eth0 parent 1:15 handle 150: sfq perturb 10
8954 <Title>Making all of the above start at boot</Title>
8956 It sure can be done in many ways. In mine, I have a shell script in /etc/init.d/packetfilter that accepts [start | stop | stop-tables | start-tables | reload-tables] it configures qdiscs and loads needed kernel modules, so it behaves much like a daemon. The same script loads iptables rules from /etc/network/iptables-rules. I will beautify it a little and will make it available on my web page <ULink URL="http://omega.resa.es/piotr/files/packetfilter.tar.bz2">here</ULink>
8966 <chapter id="lartc.bridging">
8967 <Title>Building bridges, and pseudo-bridges with Proxy ARP</Title>
8970 Bridges are devices which can be installed in a network without any
8971 reconfiguration. A network switch is basically a many-port bridge. A bridge
8972 is often a 2-port switch. Linux does however support multiple interfaces in
8973 a bridge, making it a true switch.
8977 Bridges are often deployed when confronted with a broken network that needs
8978 to be fixed without any alterations. Because the bridge is a layer-2 device,
8979 one layer below IP, routers and servers are not aware of its existence.
8980 This means that you can transparently block or modify certain packets, or do
8985 Another good thing is that a bridge can often be replaced by a cross cable
8986 or a hub, should it break down.
8990 The bad news is that a bridge can cause great confusion unless it is very
8991 well documented. It does not appear in traceroutes, but somehow packets
8992 disappear or get changed from point A to point B ('this network is
8993 HAUNTED!'). You should also wonder if an organization that 'does not want to
8994 change anything' is doing the right thing.
8998 The Linux 2.4/2.5 bridge is documented on
8999 <ULink URL=" http://bridge.sourceforge.net/">this page</ULink>.
9002 <Sect1 id="lartc.bridging.iptables">
9003 <Title>State of bridging and iptables</Title>
9006 As of Linux 2.4.20, bridging and iptables do not 'see' each other without
9007 help. If you bridge packets from eth0 to eth1, they do not 'pass' by
9008 iptables. This means that you cannot do filtering, or NAT or mangling or
9009 whatever. In Linux 2.5.45 and higher, this is fixed.
9012 You may also see 'ebtables' mentioned which is yet another project - it
9013 allows you to do wild things as MACNAT and 'brouting'. It is truly scary.
9016 <Sect1 id="lartc.bridging.shaping">
9017 <Title>Bridging and shaping</Title>
9020 This does work as advertised. Be sure to figure out which side each
9021 interface is on, otherwise you might be shaping outbound traffic in your
9022 internal interface, which won't work. Use tcpdump if needed.
9027 <Sect1 id="lartc.bridging.proxy-arp">
9028 <Title>Pseudo-bridges with Proxy-ARP</Title>
9031 If you just want to implement a Pseudo-bridge, skip down a few sections
9032 to 'Implementing it', but it is wise to read a bit about how it works in
9037 A Pseudo-bridge works a bit differently. By default, a bridge passes packets
9038 unaltered from one interface to the other. It only looks at the hardware
9039 address of packets to determine what goes where. This in turn means that you
9040 can bridge traffic that Linux does not understand, as long as it has an
9041 hardware address it does.
9045 A 'Pseudo-bridge' works differently and looks more like a hidden router than
9046 a bridge, but like a bridge, it has little impact on network design.
9050 An advantage of the fact that it is not a bridge lies in the fact that
9051 packets really pass through the kernel, and can be filtered, changed,
9052 redirected or rerouted.
9056 A real bridge can also be made to perform these feats, but it needs special
9057 code, like the Ethernet Frame Diverter, or the above mentioned patch.
9061 Another advantage of a pseudo-bridge is that it does not pass packets it
9062 does not understand - thus cleaning your network of a lot of cruft. In cases
9063 where you need this cruft (like SAP packets, or Netbeui), use a real bridge.
9067 <Title>ARP & Proxy-ARP</Title>
9070 When a host wants to talk to another host on the same physical network
9071 segment, it sends out an Address Resolution Protocol packet, which, somewhat
9072 simplified, reads like this 'who has 10.0.0.1, tell 10.0.0.7'. In response
9073 to this, 10.0.0.1 replies with a short 'here' packet.
9077 10.0.0.7 then sends packets to the hardware address mentioned in the 'here'
9078 packet. It caches this hardware address for a relatively long time, and
9079 after the cache expires, it re-asks the question.
9083 When building a Pseudo-bridge, we instruct the bridge to reply to these ARP
9084 packets, which causes the hosts in the network to send its packets to the
9085 bridge. The bridge then processes these packets, and sends them to the
9090 So, in short, whenever a host on one side of the bridge asks for the
9091 hardware address of a host on the other, the bridge replies with a packet
9092 that says 'hand it to me'.
9096 This way, all data traffic gets transmitted to the right place, and always
9097 passes through the bridge.
9103 <Title>Implementing it</Title>
9106 In the bad old days, it used to be possible to instruct the Linux Kernel to
9107 perform 'proxy-ARP' for just any subnet. So, to configure a pseudo-bridge,
9108 you would have to specify both the proper routes to both sides of the bridge
9109 AND create matching proxy-ARP rules. This is bad in that it requires a lot
9110 of typing, but also because it easily allows you to make mistakes which make
9111 your bridge respond to ARP queries for networks it does not know how to
9116 With Linux 2.4/2.5 (and possibly 2.2), this possibility has been withdrawn and
9117 has been replaced by a flag in the /proc directory, called 'proxy_arp'. The
9118 procedure for building a pseudo-bridge is then:
9127 Assign an IP address to both interfaces, the 'left' and the 'right'
9134 Create routes so your machine knows which hosts reside on the left,
9135 and which on the right
9141 Turn on proxy-ARP on both interfaces, echo 1 >
9142 /proc/sys/net/ipv4/conf/ethL/proxy_arp, echo 1 >
9143 /proc/sys/net/ipv4/conf/ethR/proxy_arp, where L and R stand for the numbers
9144 of your interfaces on the left and on the right side
9153 Also, do not forget to turn on the ip_forwarding flag! When converting from
9154 a true bridge, you may find that this flag was turned off as it is not
9155 needed when bridging.
9159 Another thing you might note when converting is that you need to clear the
9160 arp cache of computers in the network - the arp cache might contain old
9161 pre-bridge hardware addresses which are no longer correct.
9165 On a Cisco, this is done using the command 'clear arp-cache', under
9166 Linux, use 'arp -d ip.address'. You can also wait for the cache to expire
9167 manually, which can take rather long.
9170 You can speed this up using the wonderful 'arping' tool, which on many
9171 distributions is part of the 'iputils' package. Using 'arping' you can send
9172 out unsolicited ARP messages so as to update remote arp caches.
9175 This is a very powerful technique that is also used by 'black hats' to
9176 subvert your routing!
9180 On Linux 2.4, you may need to execute
9181 'echo 1 > /proc/sys/net/ipv4/ip_nonlocal_bind' before being able to send
9182 out unsolicited ARP messages!
9186 You may also discover that your network was misconfigured if you are/were of
9187 the habit of specifying routes without netmasks. To explain, some versions
9188 of route may have guessed your netmask right in the past, or guessed wrong
9189 without you noticing. When doing surgical routing like described above, it
9190 is *vital* that you check your netmasks!
9199 <chapter id="lartc.dynamic-routing">
9200 <Title>Dynamic routing - OSPF and BGP</Title>
9203 Once your network starts to get really big, or you start to consider 'the
9204 internet' as your network, you need tools which dynamically route your data.
9205 Sites are often connected to each other with multiple links, and more are
9206 popping up all the time.
9210 The Internet has mostly standardized on OSPF and BGP4 (rfc1771).
9211 Linux supports both, by way of <application>gated</application> and
9212 <application>zebra</application>
9216 While currently not within the scope of this document, we would like to
9217 point you to the definitive works:
9227 URL="http://www.cisco.com/univercd/cc/td/doc/cisintwk/idg4/nd2003.htm"
9228 >Designing large-scale IP Internetworks</ULink
9238 "OSPF. The anatomy of an Internet routing protocol"
9239 Addison Wesley. Reading, MA. 1998.
9243 Halabi has also written a good guide to OSPF routing design, but this
9244 appears to have been dropped from the Cisco web site.
9253 "Internet routing architectures"
9254 Cisco Press (New Riders Publishing). Indianapolis, IN. 1997.
9267 URL="http://www.cisco.com/univercd/cc/td/doc/cisintwk/ics/icsbgp4.htm"
9268 >Using the Border Gateway Protocol for interdomain routing</ULink
9273 Although the examples are Cisco-specific, they are remarkably similar
9274 to the configuration language in Zebra :-)
9277 <Sect1 id="lartc.dynamic-routing.ospf">
9280 <FirstName>Pedro</FirstName><Surname>Larroy Tovar</Surname>
9283 <email>piotr@omega.resa.es</email>
9287 </sect1info><Title>Setting up OSPF with Zebra</Title>
9290 Please, let <ulink url="mailto:piotr@omega.resa.es">me</ulink> know if any of the following information is not accurate or if you have any suggestions.
9291 <ulink url="http://www.zebra.org">Zebra</ulink> is a great dynamic routing software written by Kunihiro Ishiguro, Toshiaki Takada and Yasuhiro Ohara. With Zebra, setting up OSPF is fast an simple, but in practice there's a lot of parameters to tune if you have very specific needs. OSPF stands for Open Shortest Path First, and some of its principal features are:
9294 <Term>Hierachical</Term>
9297 Networks are grouped by <emphasis>areas</emphasis>, which are interconnected by a <emphasis>backbone area</emphasis> which will be designated as <emphasis>area 0</emphasis>. All traffic goes through area 0, and all the routers in area 0 have routing information about all the other areas.
9303 <Term>Short convergence</Term>
9306 Routes are propagated very fast, compared with RIP, for example.
9311 <Term>Bandwith efficient</Term>
9314 Uses multicasting instead of broadcasting, so it doesn't flood other hosts with routing information that may not be of interest for them, thus reducing network overhead. Also, <emphasis>Internal Routers</emphasis> (those which only have interfaces in one area) Don't have routing information about other areas. Routers with interfaces in more than one area are called <emphasis>Area Border Routers</emphasis>, and hold topological information about the areas they are connected to.
9319 <Term>Cpu intensive</Term>
9322 OSPF is based on Dijkstra's <ulink url="http://www.soi.wide.ad.jp/class/99007/slides/13/07.html">Shortest Path First algorithm</ulink>, which is expensive compared to other routing algorithms. But really is not that bad, since the Shortest Path is only calculed for each area, also for small to medium sized networks this won't be an issue, and you won't even notice.
9327 <Term>Link state</Term>
9330 OSPF counts with the special characteristics of networks and interfaces, such as bandwith, link failures, and monetary cost.
9335 <Term>Open protocol and GPLed software</Term>
9338 OSPF is an open protocol, and Zebra is GPL software, which has obvious advantages over propietary software and protocols.
9345 <Sect2 id="lartc.dynamic-routing.ospf.prereq">
9346 <Title>Prerequisites</Title>
9352 <Term>Linux Kernel:</Term>
9355 Compiled with CONFIG_NETLINK_DEV and CONFIG_IP_MULTICAST (I am not sure if anything more is also needed).
9360 <Term>Iproute</Term>
9370 Get it with your favorite package manager or from <ulink url="http://www.zebra.org">http://www.zebra.org</ulink>.
9376 <Sect2 id="lartc.dynamic-routing.ospf.zebracfg">
9377 <Title>Configuring Zebra</Title>
9379 Let's take this network as an example:
9381 ----------------------------------------------------
9384 | Area 0 100BaseTX Switched |
9385 | Backbone Ethernet |
9386 ----------------------------------------------------
9390 |100BaseTX |100BaseTX |100BaseTX |100BaseTX
9392 --------- ------------ ----------- ----------------
9393 |R Omega| |R Atlantis| |R Legolas| |R Frodo |
9394 --------- ------------ ----------- ----------------
9397 |2MbDSL/ATM |100BaseTX |10BaseT |10BaseT |10BaseT
9398 ------------ ------------------------------------ -------------------------------
9399 | Internet | | 172.17.0.0/16 Area 1 | | 192.168.1.0/24 wlan Area 2|
9400 ------------ | Student network (dorm) | | barcelonawireless |
9401 ------------------------------------ -------------------------------
9403 Don't be afraid by this diagram, zebra does most of the work automatically, so it won't take any work to put all the routes up with zebra. It would be painful to mantain all those routes by hand in a day to day basis. The most important thing you must have clear, is the network topology. And take special care with Area 0, since it's the most important.
9404 First configure zebra, editing zebra.conf and adapt it to your needs:
9410 ! Interface's description.
9413 ! description test of desc.
9418 ! Static default route
9420 ip route 0.0.0.0/0 212.170.21.129
9422 log file /var/log/zebra/zebra.log
9424 In Debian, I will also had to edit /etc/zebra/daemons so they start at boot:
9429 Now we have to edit ospfd.conf if you are still runnig IPV4 or ospf6d.conf if you run IPV6. My ospfd.conf looks like:
9436 network 192.168.0.0/24 area 0
9437 network 172.17.0.0/16 area 1
9440 log file /var/log/zebra/ospfd.log
9442 Here we instruct ospf about our network topology.
9446 <Sect2 id="lartc.dynamic-routing.ospf.running">
9447 <Title>Running Zebra</Title>
9449 Now, we have to start Zebra; either by hand by typing "zebra -d" or with some script like "/etc/init.d/zebra start". Then carefully watching the ospdfd logs we should see something like:
9451 2002/12/13 22:46:24 OSPF: interface 192.168.0.1 join AllSPFRouters Multicast group.
9452 2002/12/13 22:46:34 OSPF: SMUX_CLOSE with reason: 5
9453 2002/12/13 22:46:44 OSPF: SMUX_CLOSE with reason: 5
9454 2002/12/13 22:46:54 OSPF: SMUX_CLOSE with reason: 5
9455 2002/12/13 22:47:04 OSPF: SMUX_CLOSE with reason: 5
9456 2002/12/13 22:47:04 OSPF: DR-Election[1st]: Backup 192.168.0.1
9457 2002/12/13 22:47:04 OSPF: DR-Election[1st]: DR 192.168.0.1
9458 2002/12/13 22:47:04 OSPF: DR-Election[2nd]: Backup 0.0.0.0
9459 2002/12/13 22:47:04 OSPF: DR-Election[2nd]: DR 192.168.0.1
9460 2002/12/13 22:47:04 OSPF: interface 192.168.0.1 join AllDRouters Multicast group.
9461 2002/12/13 22:47:06 OSPF: DR-Election[1st]: Backup 192.168.0.2
9462 2002/12/13 22:47:06 OSPF: DR-Election[1st]: DR 192.168.0.1
9463 2002/12/13 22:47:06 OSPF: Packet[DD]: Negotiation done (Slave).
9464 2002/12/13 22:47:06 OSPF: nsm_change_status(): scheduling new router-LSA origination
9465 2002/12/13 22:47:11 OSPF: ospf_intra_add_router: Start
9467 Ignore the SMUX_CLOSE message by now, since it's about SNMP. We can see that 192.168.0.1 is the <emphasis>Designated Router</emphasis> and 192.168.0.2 is the <emphasis>Backup Designated Router</emphasis>
9471 We can also interact with the zebra or the ospfd interface by executing:
9473 <prompt>$ </prompt>telnet localhost zebra
9474 <prompt>$ </prompt>telnet localhost ospfd
9477 Let's see how to view if the routes are propagating, log into zebra and type:
9480 root@atlantis:~# telnet localhost zebra
9482 Connected to atlantis.
9483 Escape character is '^]'.
9485 Hello, this is zebra (version 0.92a).
9486 Copyright 1996-2001 Kunihiro Ishiguro.
9488 User Access Verification
9491 atlantis> show ip route
9492 Codes: K - kernel route, C - connected, S - static, R - RIP, O - OSPF,
9493 B - BGP, > - selected route, * - FIB route
9495 K>* 0.0.0.0/0 via 192.168.0.1, eth1
9496 C>* 127.0.0.0/8 is directly connected, lo
9497 O 172.17.0.0/16 [110/10] is directly connected, eth0, 06:21:53
9498 C>* 172.17.0.0/16 is directly connected, eth0
9499 O 192.168.0.0/24 [110/10] is directly connected, eth1, 06:21:53
9500 C>* 192.168.0.0/24 is directly connected, eth1
9501 atlantis> show ip ospf border-routers
9502 ============ OSPF router routing table =============
9503 R 192.168.0.253 [10] area: (0.0.0.0), ABR
9504 via 192.168.0.253, eth1
9505 [10] area: (0.0.0.1), ABR
9506 via 172.17.0.2, eth0
9508 Or with iproute directly:
9510 root@omega:~# ip route
9511 212.170.21.128/26 dev eth0 proto kernel scope link src 212.170.21.172
9512 192.168.0.0/24 dev eth1 proto kernel scope link src 192.168.0.1
9513 172.17.0.0/16 via 192.168.0.2 dev eth1 proto zebra metric 20
9514 default via 212.170.21.129 dev eth0 proto zebra
9517 We can see the zebra routes, that weren't there before. It's really nice to see routes appearing just a few seconds after you start zebra and ospfd. You can check connectivity to other hosts with ping. Zebra routes are automatic, you can just add another router to the network, configure zebra, and voila!
9523 tcpdump -i eth1 ip[9] == 89
9525 To campture OSPF packets for analisys. OSPF ip protocol number is 89, and the protocol field is the 9nth octet on the ip header.
9529 OSPF has a lot of tunable parameters, specially for large networks. In further ampliations of the howto we will show some methodologies for fine tunning OSPF.
9536 <chapter id="lartc.other"
9537 xreflabel="Other possibilities">
9538 <Title>Other possibilities</Title>
9541 This chapter is a list of projects having to do with advanced Linux routing
9542 & traffic shaping. Some of these links may deserve chapters of their
9543 own, some are documented very well of themselves, and don't need more HOWTO.
9550 <Term>802.1Q VLAN Implementation for Linux <ULink
9551 URL="http://scry.wanfear.com/~greear/vlan.html"
9556 VLANs are a very cool way to segregate your
9557 networks in a more virtual than physical way. Good information on VLANs can
9559 URL="ftp://ftp.netlab.ohio-state.edu/pub/jain/courses/cis788-97/virtual_lans/index.htm"
9561 >. With this implementation, you can have your Linux box talk
9562 VLANs with machines like Cisco Catalyst, 3Com: {Corebuilder, Netbuilder II,
9563 SuperStack II switch 630}, Extreme Ntwks Summit 48, Foundry: {ServerIronXL,
9568 A great HOWTO about VLANs can be found <ULink
9569 URL="http://scry.wanfear.com/~greear/vlan/cisco_howto.html"
9575 Update: has been included in the kernel as of 2.4.14 (perhaps 13).
9579 <Term>Alternate 802.1Q VLAN Implementation for Linux <ULink
9580 URL="http://vlan.sourceforge.net "
9585 Alternative VLAN implementation for linux. This project was started out of
9586 disagreement with the 'established' VLAN project's architecture and coding
9587 style, resulting in a cleaner overall design.
9591 <Term>Linux Virtual Server <ULink
9592 URL="http://www.LinuxVirtualServer.org/"
9597 These people are brilliant. The Linux Virtual Server is a highly scalable and
9598 highly available server built on a cluster of real servers, with the load
9599 balancer running on the Linux operating system. The architecture of the
9600 cluster is transparent to end users. End users only see a single virtual
9605 In short whatever you need to load balance, at whatever level of traffic, LVS
9606 will have a way of doing it. Some of their techniques are positively evil!
9607 For example, they let several machines have the same IP address on a
9608 segment, but turn off ARP on them. Only the LVS machine does ARP - it then
9609 decides which of the backend hosts should handle an incoming packet, and
9610 sends it directly to the right MAC address of the backend server. Outgoing
9611 traffic will flow directly to the router, and not via the LVS machine, which
9612 does therefor not need to see your 5Gbit/s of content flowing to the world,
9613 and cannot be a bottleneck.
9617 The LVS is implemented as a kernel patch in Linux 2.0 and 2.2, but as a
9618 Netfilter module in 2.4/2.5, so it does not need kernel patches! Their 2.4
9619 support is still in early development, so beat on it and give feedback or
9624 <Term>CBQ.init <ULink
9625 URL="ftp://ftp.equinox.gu.net/pub/linux/cbq/"
9630 Configuring CBQ can be a bit daunting, especially if all you want to do is
9631 shape some computers behind a router. CBQ.init can help you configure Linux
9632 with a simplified syntax.
9636 For example, if you want all computers in your 192.168.1.0/24 subnet
9637 (on 10mbit eth1) to be limited to 28kbit/s download speed, put
9638 this in the CBQ.init configuration file:
9644 DEVICE=eth1,10Mbit,1Mbit
9654 By all means use this program if the 'how and why' don't interest you.
9655 We're using CBQ.init in production and it works very well. It can even do
9656 some more advanced things, like time dependent shaping. The documentation is
9657 embedded in the script, which explains why you can't find a README.
9661 <Term>Chronox easy shaping scripts <ULink
9662 URL="http://www.chronox.de"
9667 Stephan Mueller (smueller@chronox.de) wrote two useful scripts, 'limit.conn'
9668 and 'shaper'. The first one allows you to easily throttle a single download
9675 # limit.conn -s SERVERIP -p SERVERPORT -l LIMIT
9681 It works on Linux 2.2 and 2.4/2.5.
9685 The second script is more complicated, and can be used to make lots of
9686 different queues based on iptables rules, which are used to mark packets
9687 which are then shaped.
9691 <Term>Virtual Router
9692 Redundancy Protocol implementation <ULink
9693 URL="http://w3.arobas.net/~jetienne/vrrpd/index.html"
9698 This is purely for redundancy. Two machines with their own IP address and
9699 MAC Address together create a third IP Address and MAC Address, which is
9700 virtual. Originally intended purely for routers, which need constant MAC
9701 addresses, it also works for other servers.
9705 The beauty of this approach is the incredibly easy configuration. No kernel
9706 compiling or patching required, all userspace.
9710 Just run this on all machines participating in a service:
9713 # vrrpd -i eth0 -v 50 10.0.0.22
9719 And you are in business! 10.0.0.22 is now carried by one of your servers,
9720 probably the first one to run the vrrp daemon. Now disconnect that computer
9721 from the network and very rapidly one of the other computers will assume the
9722 10.0.0.22 address, as well as the MAC address.
9726 I tried this over here and had it up and running in 1 minute. For some
9727 strange reason it decided to drop my default gateway, but the -n flag
9732 This is a 'live' fail over:
9738 64 bytes from 10.0.0.22: icmp_seq=3 ttl=255 time=0.2 ms
9739 64 bytes from 10.0.0.22: icmp_seq=4 ttl=255 time=0.2 ms
9740 64 bytes from 10.0.0.22: icmp_seq=5 ttl=255 time=16.8 ms
9741 64 bytes from 10.0.0.22: icmp_seq=6 ttl=255 time=1.8 ms
9742 64 bytes from 10.0.0.22: icmp_seq=7 ttl=255 time=1.7 ms
9748 Not *one* ping packet was lost! Just after packet 4, I disconnected my P200
9749 from the network, and my 486 took over, which you can see from the higher
9758 <chapter id="lartc.further">
9759 <Title>Further reading</Title>
9766 URL="http://snafu.freedom.org/linux2.2/iproute-notes.html"
9767 >http://snafu.freedom.org/linux2.2/iproute-notes.html</ULink
9771 Contains lots of technical information, comments from the kernel
9776 URL="http://www.davin.ottawa.on.ca/ols/"
9777 >http://www.davin.ottawa.on.ca/ols/</ULink
9781 Slides by Jamal Hadi Salim, one of the authors of Linux traffic control
9786 URL="http://defiant.coinet.com/iproute2/ip-cref/"
9787 >http://defiant.coinet.com/iproute2/ip-cref/</ULink
9791 HTML version of Alexeys LaTeX documentation - explains part of iproute2 in
9797 URL="http://www.aciri.org/floyd/cbq.html"
9798 >http://www.aciri.org/floyd/cbq.html</ULink
9802 Sally Floyd has a good page on CBQ, including her original papers. None of
9803 it is Linux specific, but it does a fair job discussing the theory and uses
9805 Very technical stuff, but good reading for those so inclined.
9809 <Term>Differentiated Services on Linux</Term>
9813 URL="ftp://icaftp.epfl.ch/pub/linux/diffserv/misc/dsid-01.txt.gz"
9815 > by Werner Almesberger, Jamal Hadi Salim and Alexey
9816 Kuznetsov describes DiffServ facilities in the Linux kernel, amongst which
9817 are TBF, GRED, the DSMARK qdisc and the tcindex classifier.
9822 URL="http://ceti.pl/~kravietz/cbq/NET4_tc.html"
9823 >http://ceti.pl/~kravietz/cbq/NET4_tc.html</ULink
9827 Yet another HOWTO, this time in Polish! You can copy/paste command lines
9828 however, they work just the same in every language. The author is
9829 cooperating with us and may soon author sections of this HOWTO.
9834 URL="http://www.cisco.com/univercd/cc/td/doc/product/software/ios111/cc111/car.htm"
9835 >IOS Committed Access Rate</ULink
9840 From the helpful folks of Cisco who have the laudable habit of putting
9841 their documentation online. Cisco syntax is different but the concepts are
9842 the same, except that we can do more and do it without routers the price of
9847 <Term>Docum experimental site<ULink
9848 URL="http://www.docum.org"
9853 Stef Coene is busy convincing his boss to sell Linux support, and so he is
9854 experimenting a lot, especially with managing bandwidth. His site has a lot
9855 of practical information, examples, tests and also points out some CBQ/tc bugs.
9860 <Term>TCP/IP Illustrated, volume 1, W. Richard Stevens, ISBN 0-201-63346-9</Term>
9863 Required reading if you truly want to understand TCP/IP. Entertaining as
9873 <chapter id="lartc.ack">
9874 <Title>Acknowledgements </Title>
9878 It is our goal to list everybody who has contributed to this HOWTO, or
9879 helped us demystify how things work. While there are currently no plans
9880 for a Netfilter type scoreboard, we do like to recognize the people who are
9886 <ItemizedList spacing="compact">
9890 <author><firstname>Junk</firstname><surname>Alins</surname></author>
9891 <address><email>juanjo@mat.upc.es</email></address>
9896 <author><firstname>Joe</firstname><surname>Van Andel</surname></author>
9902 <author><firstname>Michael</firstname><othername>T.</othername>
9903 <surname>Babcock</surname></author>
9904 <address><email>mbabcock@fibrespeed.net</email></address>
9911 <author><firstname>Christopher</firstname>
9912 <surname>Barton</surname></author>
9913 <address><email>cpbarton%uiuc.edu</email></address>
9920 <author><firstname>Ard</firstname><surname>van Breemen</surname></author>
9921 <address><email>ard%kwaak.net</email></address>
9926 <author><firstname>Ron</firstname><surname>Brinker</surname></author>
9927 <address><email>service%emcis.com</email></address>
9932 <author><firstname>?ukasz</firstname><surname>Bromirski</surname></author>
9933 <address><email>l.bromirski@mr0vka.eu.org</email></address>
9938 <author><firstname>Lennert</firstname><surname>Buytenhek</surname></author>
9939 <address><email>buytenh@gnu.org</email></address>
9944 <author><firstname>Esteve</firstname><surname>Camps</surname></author>
9945 <address><email>esteve@hades.udg.es</email></address>
9950 <author><firstname>Stef</firstname><surname>Coene</surname></author>
9951 <address><email>stef.coene@docum.org</email></address>
9956 <author><firstname>Don</firstname><surname>Cohen</surname></author>
9957 <address><email>don-lartc%isis.cs3-inc.com</email></address>
9962 <author><firstname>Jonathan</firstname><surname>Corbet</surname></author>
9963 <address><email>lwn%lwn.net</email></address>
9968 <author><firstname>Gerry</firstname><surname>Creager</surname>
9969 <othername>N5JXS</othername></author>
9970 <address><email>gerry%cs.tamu.edu</email></address>
9975 <author><firstname>Marco</firstname><surname>Davids</surname></author>
9976 <address><email>marco@sara.nl</email></address>
9981 <author><firstname>Jonathan</firstname><surname>Day</surname></author>
9982 <address><email>jd9812@my-deja.com</email></address>
9987 <author><firstname>Martin</firstname><surname>Devera</surname>
9988 <othername>aka devik</othername></author>
9989 <address><email>devik@cdi.cz</email></address>
9995 <author><firstname>Hannes</firstname><surname>Ebner</surname>
9997 <address><email>he%fli4l.de</email></address>
10003 <author><firstname>Derek</firstname><surname>Fawcus</surname>
10005 <address><email>dfawcus%cisco.com</email></address>
10012 <author><firstname>Stephan</firstname><othername>"Kobold"</othername>
10013 <surname>Gehring</surname></author>
10014 <address><email>Stephan.Gehring@bechtle.de</email></address>
10019 <author><firstname>Jacek</firstname><surname>Glinkowski</surname></author>
10020 <address><email>jglinkow%hns.com</email></address>
10025 <author><firstname>Andrea</firstname><surname>Glorioso</surname></author>
10026 <address><email>sama%perchetopi.org</email></address>
10031 <author><firstname>Nadeem</firstname><surname>Hasan</surname></author>
10032 <address><email>nhasan@usa.net</email></address>
10037 <author><firstname>Erik</firstname><surname>Hensema</surname></author>
10038 <address><email>erik%hensema.xs4all.nl</email></address>
10043 <author><firstname>Vik</firstname><surname>Heyndrickx</surname></author>
10044 <address><email>vik.heyndrickx@edchq.com</email></address>
10049 <author><firstname>Spauldo</firstname><surname>Da Hippie</surname></author>
10050 <address><email>spauldo%usa.net</email></address>
10055 <author><firstname>Koos</firstname><surname>van den Hout</surname></author>
10056 <address><email>koos@kzdoos.xs4all.nl</email></address>
10062 Stefan Huelbrock <shuelbrock%datasystems.de>
10068 Alexander W. Janssen <yalla%ynfonatic.de>
10074 Gareth John <gdjohn%zepler.org>
10080 <author><firstname>Dave</firstname><surname>Johnson</surname></author>
10081 <address><email>dj@www.uk.linux.org</email></address>
10089 Martin Josefsson <gandalf%wlug.westbo.se>
10095 Andi Kleen <ak%suse.de>
10101 Andreas J. Koenig <andreas.koenig%anima.de>
10107 Pawel Krawczyk <kravietz%alfa.ceti.pl>
10113 Amit Kucheria <amitk@ittc.ku.edu>
10119 Edmund Lau <edlau%ucf.ics.uci.edu>
10125 Philippe Latu <philippe.latu%linux-france.org>
10131 Arthur van Leeuwen <arthurvl%sci.kun.nl>
10136 <author><firstname>Jose Luis Domingo</firstname><surname>Lopez</surname>
10138 <address><email>jdomingo@24x7linux.com</email></address>
10144 Jason Lunz <j@cc.gatech.edu>
10150 Stuart Lynne <sl@fireplug.net>
10156 Alexey Mahotkin <alexm@formulabez.ru>
10162 Predrag Malicevic <pmalic@ieee.org>
10167 Patrick McHardy <kaber@trash.net>
10175 Andreas Mohr <andi%lisas.de>
10180 <para>James Morris <jmorris@intercode.com.au>
10186 Andrew Morton <akpm%zip.com.au>
10198 Stephan Mueller <smueller@chronox.de>
10204 Togan Muftuoglu <toganm%yahoo.com>
10211 Chris Murray <cmurray@stargate.ca>
10217 Patrick Nagelschmidt <dto%gmx.net>
10223 Ram Narula <ram@princess1.net>
10229 Jorge Novo <jnovo@educanet.net>
10235 Patrik <ph@kurd.nu>
10239 <listitem><para>P?l Osgy?ny <oplab%westel900.net></para></listitem>
10244 Lutz Preßler <Lutz.Pressler%SerNet.DE>
10250 Jason Pyeron <jason%pyeron.com>
10256 Rusty Russell <rusty%rustcorp.com.au>
10262 Mihai RUSU <dizzy%roedu.net>
10268 Jamal Hadi Salim <hadi%cyberus.ca>
10274 Ren? Serral <rserral%ac.upc.es>
10281 David Sauer <davids%penguin.cz>
10287 Sheharyar Suleman Shaikh <sss23@drexel.edu>
10293 Stewart Shields <MourningBlade%bigfoot.com>
10299 Nick Silberstein <nhsilber%yahoo.com>
10305 Konrads Smelkov <konrads@interbaltika.com>
10311 <author><firstname>William</firstname><surname>Stearns</surname></author>
10312 <address><email>wstearns@pobox.com</email></address>
10318 Andreas Steinmetz <ast%domdv.de>
10324 Jason Tackaberry <tack@linux.com>
10330 Charles Tassell <ctassell%isn.net>
10336 Glen Turner <glen.turner%aarnet.edu.au>
10342 Tea Sponsor: Eric Veldhuyzen <eric%terra.nu>
10350 Song Wang <wsong@ece.uci.edu>
10356 <author><firstname>Chris</firstname><surname>Wilson</surname></author>
10357 <address><email>chris@netservers.co.uk</email></address>
10364 <author><firstname>Lazar</firstname><surname>Yanackiev</surname></author>
10365 <address><email>Lyanackiev%gmx.net</email></address>
10371 <author><firstname>Pedro</firstname><surname>Larroy</surname></author>
10372 <address><email>piotr%omega.resa.es</email></address>
10376 Chapter 15, section 10: Example of a full nat solution with QoS
10382 Chapter 17,, section 1: Setting up OSPF with Zebra