Complex Traffic Shaping/ControlTomEastepArneBernin2001-2009Thomas M. Eastep2005Arne Bernin & Thomas M. EastepPermission is granted to copy, distribute and/or modify this
document under the terms of the GNU Free Documentation License, Version
1.2 or any later version published by the Free Software Foundation; with
no Invariant Sections, with no Front-Cover, and with no Back-Cover
Texts. A copy of the license is included in the section entitled
GNU Free Documentation
License.Traffic shaping is complex and the Shorewall community is not well
equipped to answer traffic shaping questions. So if you are the type of
person who needs "insert tab A into slot B" instructions for everything
that you do, then please don't try to implement traffic shaping using
Shorewall. You will just frustrate yourself and we won't be able to help
you.Said another way, reading just Shorewall documentation is not going
to give you enough background to use this material.At a minimum, you will need to refer to at least the following
additional information:The LARTC HOWTO: http://www.lartc.orgThe HTB User's Guide: http://luxik.cdi.cz/~devik/qos/htb/manual/userg.htmHFSC Scheduling with Linux: http://linux-ip.net/articles/hfsc.en/Some of the documents listed at http://www.netfilter.org/documentation/index.html#documentation-howto.
The tutorial by Oskar Andreasson is particularly good.The output of man iptablesIntroductionBeginning with Shorewall 4.4.6, Shorewall includes two separate
implementations of traffic shaping. This document describes the original
implementation which is complex and difficult to configure. A much simpler
version is described in Simple Traffic Shaping/Control
and is highly recommended unless you really need to delay certain traffic
passing through your firewall.Shorewall has builtin support for traffic shaping and control. This
support does not cover all options available (and especially all
algorithms that can be used to queue traffic) in the Linux kernel but it
should fit most needs. If you are using your own script for traffic
control and you still want to use it in the future, you will find
information on how to do this, later in this
document. But for this to work, you will also need to enable
traffic shaping in the kernel and Shorewall as covered by the next
sections.Linux traffic shaping and controlThis section gives a brief introduction of how controlling traffic
with the Linux kernel works. Although this might be enough for configuring
it in the Shorewall configuration files, we strongly recommend that you
take a deeper look into the Linux
Advanced Routing and Shaping HOWTO. At the time of writing this,
the current version is 1.0.0.Since kernel 2.2, Linux has extensive support for controlling
traffic. You can define different algorithms that are used to queue the
traffic before it leaves an interface. The standard one is called pfifo
and is (as the name suggests) of the type First In First out. This means,
that it does not shape anything, if you have a connection that eats up all
your bandwidth, this queuing algorithm will not stop it from doing
so.For Shorewall traffic shaping we use three algorithms: HTB
(Hierarchical Token Bucket), HFSC (Hierarchical Fair Service Curves) and
SFQ (Stochastic Fairness Queuing). SFQ is easy to explain: it just tries
to track your connections (tcp or udp streams) and balances the traffic
between them. This normally works well. HTB and HFSC allow you to define a
set of classes, and you can put the traffic you want into these classes.
You can define minimum and maximum bandwidth settings for those classes
and order them hierarchically (the less prioritized classes only get
bandwidth if the more important have what they need). Additionally, HFSC
allows you to specify the maximum queuing delay that a packet may
experience. Shorewall builtin traffic shaping allows you to define these
classes (and their bandwidth limits), and it uses SFQ inside these classes
to make sure, that different data streams are handled equally. If SFQ's
default notion of a 'stream' doesn't work well for you, you can change it
using the flow option described below.You can shape incoming traffic through use of an
Intermediate Functional Block (IFB) device. See below. But beware: using an
IFB can result in queues building up both at your ISPs router and at your
own.You shape and control outgoing traffic by assigning the traffic to
classes. Each class is associated with exactly one
network interface and has a number of attributes:PRIORITY - Used to give preference to one class over another
when selecting a packet to send. The priority is a numeric value with
1 being the highest priority, 2 being the next highest, and so
on.RATE - The minimum bandwidth this class should get, when the
traffic load rises. Classes with a higher priority (lower PRIORITY
value) are served even if there are others that have a guaranteed
bandwidth but have a lower priority (higher PRIORITY value).CEIL - The maximum bandwidth the class is allowed to use when
the link is idle.MARK - Netfilter has a facility for
marking packets. Packet marks have a numeric
value which is limited in Shorewall to the values 1-255 (1-16383 if
you set WIDE_TC_MARKS=Yes in shorewall.conf (5) ). You
assign packet marks to different types of traffic using entries in the
/etc/shorewall/tcrules file.In Shorewall 4.5.0, WIDE_TC_MARKS was superseded by TC_BITS
which specifies the width in bits of the traffic shaping mark field.
The default is based on the setting of WIDE_TC_MARKS so as to
provide upward compatibility.One class for each interface must be designated as the
default class. This is the class to which unmarked
traffic (packets to which you have not assigned a mark value in
/etc/shorewall/tcrules) is assigned.Netfilter also supports a mark value on each connection. You can
assign connection mark values in
/etc/shorewall/tcrules, you can copy the current
packet's mark to the connection mark (SAVE), or you can copy the
connection mark value to the current packet's mark (RESTORE). For more
information, see this
article.Linux Kernel ConfigurationYou will need at least kernel 2.4.18 for this to work, please take a
look at the following screenshot for what settings you need to enable. For
builtin support, you need the HTB scheduler, the Ingress scheduler, the
PRIO pseudoscheduler and SFQ queue. The other scheduler or queue
algorithms are not needed.This screen shot shows how I configured QoS in a 2.6.16
Kernel:And here's my recommendation for a 2.6.21 kernel:Enable TC support in ShorewallYou need this support whether you use the builtin support or whether
you provide your own tcstart script.To enable the builtin traffic shaping and control in Shorewall, you
have to do the following:Set TC_ENABLED to "Internal" in /etc/shorewall/shorewall.conf.
Setting TC_ENABLED=Yes causes
Shorewall to look for an external tcstart file (See a later section for details).Setting CLEAR_TC parameter in
/etc/shorewall/shorewall.conf to Yes
will clear the traffic shaping configuration during Shorewall
[re]start and Shorewall stop. This is normally what you want when
using the builtin support (and also if you use your own tcstart
script)The other steps that follow depend on whether you use your own
script or the builtin solution. They will be explained in the
following sections.Using builtin traffic shaping/controlShorewall's builtin traffic shaping feature provides a thin layer on
top of the ingress qdesc, HTB and SFQ. That translation layer allows you
to:Define HTB and/or HFSC classes using Shorewall-style
column-oriented configuration files.Integrate the reloading of your traffic shaping configuration
with the reloading of your packet-filtering and marking
configuration.Assign traffic to HTB or HFSC classes by TOS value.Assign outgoing TCP ACK packets to an HTB or HFSC class.Assign traffic to HTB and/or HFSC classes based on packet mark
value or based on packet contents.Those few features are really all that builtin traffic
shaping/control provides; consequently, you need to understand HTB and/or
HFSC and Linux traffic shaping as well as Netfilter packet marking in
order to use the facility. Again, please see the links at top of this
article.For defining bandwidths (for either devices or classes) please use
kbit or kbps (for Kilobytes per second) and make sure there is NO space between the number and the unit (it is
100kbit not 100 kbit). Using mbit, mbps
or a raw number (which means bytes) could be used, but note that only
integer numbers are supported (0.5 is not
valid).To properly configure the settings for your
devices you need to find out the real up- and downstream rates you
have. This is especially the case, if you are using a DSL
connection or one of another type that do not have a guaranteed bandwidth.
Don't trust the values your provider tells you for this; especially
measuring the real download speed is important! There are several online
tools that help you find out; search for "dsl speed test" on google (For
Germany you can use arcor speed
check). Be sure to choose a test site located near you./etc/shorewall/tcdevicesThis file allows you to define the incoming and outgoing bandwidth
for the devices you want traffic shaping to be enabled. That means, if
you want to use traffic shaping for a device, you have to define it
here. For additional information, see shorewall-tcdevices
(5).Columns in the file are as follows:INTERFACE - Name of interface. Each interface may be listed
only once in this file. You may NOT specify the name of an alias
(e.g., eth0:0) here; see FAQ #18.
You man NOT specify wildcards here, e.g. if you have multiple ppp
interfaces, you need to put them all in here! Shorewall will
determine if the device exists and will only configure the device if
it does exist. If it doesn't exist or it is DOWN, the following
warning is issued:WARNING: Device <device name> is
not in the UP state -- traffic-shaping configuration
skippedShorewall assigns a sequential interface
number to each interface (the first entry in
/etc/shorewall/tcdevices is interface 1, the
second is interface 2 and so on) You can also explicitly specify the
interface number by prefixing the interface name with the number and
a colon (":"). Example: 1:eth0.Device numbers are expressed in hexidecimal. So the device
following 9 is A, not 10.IN-BANDWIDTH - The incoming Bandwidth of that interface.
Please note that when you use this column, you are not traffic
shaping incoming traffic, as the traffic is already received before
you could do so. This Column allows you to define the maximum
traffic allowed for this interface in total, if the rate is
exceeded, the excess packets are dropped. You want this mainly if
you have a DSL or Cable Connection to avoid queuing at your
providers side. If you don't want any traffic to be dropped set this
to a value faster than your interface maximum rate (or to 0
(zero).To determine the optimum value for this setting, we recommend
that you start by setting it significantly below your measured
download bandwidth (20% or so). While downloading, measure the
ping response time from the firewall to the
upstream router as you gradually increase the setting.The optimal
setting is at the point beyond which the ping
time increases sharply as you increase the setting.For fast lines, the actually download speed may be well
below what you specify here. If you have this problem, then follow
the bandwidth with a ":" and a burst size.
The default burst is 10kb, but on my 50mbit line, I specify 200kb.
(50mbit:200kb).OUT-BANDWIDTH - Specify the outgoing bandwidth of that
interface. This is the maximum speed your connection can handle. It
is also the speed you can refer as "full" if you define the tc
classes. Outgoing traffic above this rate will be dropped.OPTIONS — A comma-separated list of options from the following
list:classifyIf specified, classification of traffic into the various
classes is done by CLASSIFY entries in
/etc/shorewall/tcrules or by entries in
/etc/shorewall/tcfilters. No MARK value
will be associated with classes on this interface.hfscShorewall normally uses the Hierarchical
Token Bucket (HTB) queuing discipline. When
is specified, the
Hierarchical Fair Service Curves (HFSC)
discipline is used instead.REDIRECTED INTERFACES — Entries are appropriate in this column
only if the device in the INTERFACE column names a Intermediate Functional Block (IFB). It lists
the physical interfaces that will have their input shaped using
classes defined on the IFB. Neither the IFB nor any of the
interfaces listed in this column may have an IN-BANDWIDTH specified.
You may specify zero (0) or a dash ("-:) in the IN-BANDWIDTH
column.IFB devices automatically get the classify option.Suppose you are using PPP over Ethernet (DSL) and ppp0 is the
interface for this. The device has an outgoing bandwidth of 500kbit
and an incoming bandwidth of 6000kbit#INTERFACE IN-BANDWITH OUT-BANDWIDTH
ppp0 6000kbit 500kbit/etc/shorewall/tcclassesThis file allows you to define the actual classes that are used to
split the outgoing traffic. For additional information, see shorewall-tcclasses
(5).INTERFACE - Name of interface. Users may also specify the
interface number. Must match the name (or number) of an interface
with an entry in /etc/shorewall/tcdevices. If
the interface has the classify
option in /etc/shorewall/tcdevices, then the
interface name or number must be followed by a colon and a
class number. Examples: eth0:1, 4:9. Class
numbers must be unique for a given interface. Normally, all classes
defined here are sub-classes of a root class that is implicitly
defined from the entry in shorewall-tcdevices(5). You
can establish a class hierarchy by specifying a
parent class (e.g.,
interface:parent-class:class)
-- the number of a class that you have previously defined. The
sub-class may borrow unused bandwidth from its parent.Class numbers are expressed in hexidecimal. So the class
following class 9 is A, not 10.MARK - The mark value which is an integer in the range 1-255
(1-16383 if you set WIDE_TC_MARKS=Yes in shorewall.conf (5) ). You
define these marks in the tcrules file, marking the traffic you want
to go into the queuing classes defined in here. You can use the same
marks for different Interfaces. You must specify "-' in this column
if the device specified in the INTERFACE column has the classify option in
/etc/shorewall/tcdevices.In Shorewall 4.5.0, WIDE_TC_MARKS was superseded by TC_BITS
which specifies the width in bits of the traffic shaping mark
field. The default is based on the setting of WIDE_TC_MARKS so as
to provide upward compatibility.RATE - The minimum bandwidth this class should get, when the
traffic load rises. Please note that first the classes which equal
or a lesser priority value are served even if there are others that
have a guaranteed bandwidth but a lower priority. If the sum of the RATEs for all classes assigned to an
INTERFACE exceed that interfaces's OUT-BANDWIDTH, then the
OUT-BANDWIDTH limit will not be honored.When using HFSC, this column may contain 1, 2 or 3 pieces of
information separated by colons (":"). In addition to the minimum
bandwidth, leaf classes may specify realtime criteria: DMAX (maximum
delay in milliseconds) and optionally UMAX (the largest packet
expected in the class). See below for
details.CEIL - The maximum bandwidth this class is allowed to use when
the link is idle. Useful if you have traffic which can get full
speed when more important services (e.g. interactive like ssh) are
not used. You can use the value "full" in here for setting the
maximum bandwidth to the defined output bandwidth of that
interface.PRIORITY - you have to define a priority for the class.
packets in a class with a higher priority (=lesser value) are
handled before less prioritized ones. You can just define the mark
value here also, if you are increasing the mark values with lesser
priority.OPTIONS - A comma-separated list of options including the
following:default - this is the default class for that interface
where all traffic should go, that is not classified
otherwise.defining default for exactly one class per interface is
mandatory!tos-<tosname> - this lets you define a filter for
the given <tosname> which lets you define a value of the
Type Of Service bits in the ip package which causes the package
to go in this class. Please note, that this filter overrides all
mark settings, so if you define a tos filter for a class all
traffic having that mark will go in it regardless of the mark on
the package. You can use the following for this option:
tos-minimize-delay (16) tos-maximize-throughput (8)
tos-maximize-reliability (4) tos-minimize-cost (2)
tos-normal-service (0)Each of this options is only valid for one class per interface.tcp-ack - if defined causes an tc filter to be created
that puts all tcp ack packets on that interface that have an
size of <=64 Bytes to go in this class. This is useful for
speeding up downloads. Please note that the size of the ack
packets is limited to 64 bytes as some applications (p2p for
example) use to make every package an ack package which would
cause them all into here. We want only packets WITHOUT payload
to match, so the size limit. Bigger packets just take their
normal way into the classes.This option is only valid for class per interface.occurs=number - Typically used with
an IPMARK entry in tcrules. Causes the rule to be replicated for
a total of number rules. Each rule has a
successively class number and mark value.When 'occurs' is used:The associated device may not have the 'classify'
option.The class may not be the default class.The class may not have any 'tos=' options (including
'tcp-ack').The class should not specify a MARK value. If one is
specified, it will be ignored with a warning message.The 'RATE' and 'CEIL' parameters apply to each instance of
the class. So the total RATE represented by an entry with
'occurs' will be the listed RATE multiplied by
number. For additional information, see
tcrules
(5).flow=keys - Shorewall attaches an SFQ
queuing discipline to each leaf HTB and HFSC class. SFQ ensures
that each flow gets equal access to the
interface. The default definition of a flow corresponds roughly
to a Netfilter connection. So if one internal system is running
BitTorrent, for example, it can have lots of 'flows' and can
thus take up a larger share of the bandwidth than a system
having only a single active connection. The
classifier (module cls_flow) works around
this by letting you define what a 'flow' is. The clasifier must
be used carefully or it can block off all traffic on an
interface! The flow option can be specified for an HTB or HFSC
leaf class (one that has no sub-classes). We recommend that you
use the following:Shaping internet-bound traffic: flow=nfct-srcShaping traffic bound for your local net: flow=dstThese will cause a 'flow' to consists of the traffic
to/from each internal system.When more than one key is give, they must be enclosed in
parenthesis and separated by commas.To see a list of the possible flow keys, run this
command:
tc filter add flow help
Those that begin with "nfct-" are Netfilter connection
tracking fields. As shown above, we recommend flow=nfct-src;
that means that we want to use the source IP address
before SNAT as the key.Shorewall cannot determine ahead of time if the flow
classifier is available in your kernel (especially if it was
built into the kernel as opposed to being loaded as a module).
Consequently, you should check ahead of time to ensure that
both your kernel and 'tc' utility support the feature.You can test the 'tc' utility by typing (as
root):
tc filter add flow help
If flow is supported, you will see: Usage: ... flow ...
[mapping mode]: map key KEY [ OPS ] ...
[hashing mode]: hash keys KEY-LIST ...
...If 'flow' is not supported, you will see: Unknown filter "flow", hence option "help" is unparsableIf your kernel supports module autoloading, just type
(as root):
modprobe cls_flow
If 'flow' is supported, no output is produced;
otherwise, you will see: FATAL: Module cls_flow not found.If your kernel is not modularized or does not support
module autoloading, look at your kernel configuration (either
/proc/config.gz or the
.config file in /lib/modules/<kernel-version>/build/If 'flow' is supported, you will see: NET_CLS_FLOW=m or
NET_CLS_FLOW=y.For modularized kernels, Shorewall will attempt to load
/lib/modules/<kernel-version>/net/sched/cls_flow.ko
by default.pfifo - When specified for a leaf class, the pfifo queing
discipline is applied to the class rather than the sfq queuing
discipline.limit=number - Added in Shorewall
4.4.3. When specified for a leaf class, specifies the maximum
number of packets that may be queued within the class. The
number must be > 2 and less than 128. If
not specified, the value 127 is assumed/etc/shorewall/tcrulesUnlike rules in the shorewall-rules(5) file,
evaluation of rules in this file will continue after a match. So the
final mark for each packet will be the one assigned by the LAST tcrule
that matches.Also unlike rules in the shorewall-rules(5) file,
the tcrules file is not stateful. So every packet that goes into, out
of or through your firewall is subject to entries in the tcrules
file.Because tcrules are not stateful, it is necessary to understand
basic IP socket operation. Here is an edited excerpt from a post on
the Shorewall Users list:
For the purposes of this discussion, the world is separated
into clients and servers. Servers provide services to
clients.When a server starts, it creates a socket and
binds the socket to an
address. For AF_INET (IPv4) and AF_INET6
(IPv6) sockets, that address is an ordered triple consisting of an
IPv4 or IPv6 address, a protocol, and possibly a port number. Port
numbers are only used when the protocol is TCP, UDP, SCTP or DCCP.
The protocol and port number used by a server are typically
well-known so that clients will be able to connect to it or send
datagrams to it. So SSH servers bind to TCP port 22, SMTP servers
bind to TCP port 25, etc. We will call this port the SERVER
PORT.When a client want to use the service provided by a server,
it also creates a socket and, like the server's socket, the
client's socket must be bound to an address. But in the case of
the client, the socket is usually given an automatic address
binding. For AF_INET and AF_INET6 sockets. the IP address is the
IP address of the client system (loose generalization) and the
port number is selected from a local port
range. On Linux systems, the local port range can be
seen by cat
/proc/sys/net/ipv4/ip_local_port_range. So it is not
possible in advance to determine what port the client will be
using. Whatever it is, we'll call it the CLIENT PORT.Now:
Packets sent from the client to the server will
have:
SOURCE PORT = CLIENT PORTDEST PORT = SERVER PORT
Packets sent from the server to the client will have:
SOURCE PORT = SERVER PORTDEST PORT = CLIENT PORT
Since the SERVER PORT is generally the only port known ahead
of time, we must categorize traffic from the server to the client
using the SOURCE PORT.
The fwmark classifier provides a convenient way to classify
packets for traffic shaping. The /etc/shorewall/tcrules
file is used for specifying these marks in a tabular fashion. For an
in-depth look at the packet marking facility in Netfilter/Shorewall,
please see this article.For marking forwarded traffic, you must
either set MARK_IN_FORWARD_CHAIN=Yes shorewall.conf or by using the :F
qualifier (see below).Columns in the file are as follows:MARK or CLASSIFY - MARK specifies the mark value is to be
assigned in case of a match. This is an integer in the range 1-255
(1-16383 if you set WIDE_TC_MARKS=Yes in shorewall.conf (5)
).In Shorewall 4.5.0, WIDE_TC_MARKS was superseded by TC_BITS
which specifies the width in bits of the traffic shaping mark
field. The default is based on the setting of WIDE_TC_MARKS so as
to provide upward compatibility.This value may be optionally followed by : and
either F, P or "T" to designate that
the marking will occur in the FORWARD, PREROUTING or POSTROUTING
chains respectively. If this additional specification is omitted,
the chain used to mark packets will be determined as follows:If the SOURCE is
$FW[:<address>], then the rule is
inserted in the OUTPUT chain.Otherwise, the chain is determined by the setting of the
MARK_IN_FORWARD_CHAIN option in shorewall.conf.Use the 'T' qualifier if you want the
rule to apply equally to traffic being routed through the firewall
and to traffic originating on the firewall
itself.Normally, the mark is applied to the packet. If you follow the
mark value with ":" and "C", then the mark is applied to the
connection. "C" can be combined with "F", "P" or "T" to designate
that the connection should be marked in a particular chain (e.g.,
"CF", "CP", "CT").There are additional special values available:RESTORE[/mask] --
restore the packet's mark from the connection's mark using the
supplied mask if any. Your kernel and iptables must include
CONNMARK support.As above, may be followed by :P, :F
or :T.SAVE[/mask] -- save
the packet's mark to the connection's mark using the supplied
mask if any. Your kernel and iptables must include CONNMARK
support.As above, may be followed by :P, :F
or :T.CONTINUE Don't process
any more marking rules in the table.As above, may be followed by :P, :F
or :T.COMMENT -- the rest of
the line will be attached as a comment to the Netfilter rule(s)
generated by the following entries. The comment will appear
delimited by "/* ... */" in the output of shorewall
show mangleTo stop the comment from being attached to further rules,
simply include COMMENT on a line by itself.To use CLASSIFY, your kernel and iptables must include
CLASSIFY target support. In that case, this column contains a
classification (classid) of the form <major>:<minor>
where <major> and <minor> are integers. Corresponds to
the 'class' specification in these traffic shaping modules:atmcbqdsmarkpfifo_fasthtbprioClassification occurs in the POSTROUTING chain except when the SOURCE contains
$FW[:<address>] in which case, the
classify action takes place in the OUTPUT chain. When used with the
builtin traffic shaper, the <major> class is the interface
number and the <minor> class is either:Constructed by Shorewall. The method of construction
depends on the setting of WIDE_TC_MARKS (shorewall.conf
(5)).When WIDE_TC_MARKS=No (the default), the <minor>
class is:the MARK value of the class preceded by the number "1"
or "10" (MARK value 1 is <minor> class 11, MARK value
22 is <minor> class 122, and so on). "10" is used
where there are more than 10 devices defined in /etc/shorewall/tcdevices.When WIDE_TC_MARKS=Yes, the <minor> class is
assigned sequentially beginning with 2.The class number, if specified.SOURCE - Source of the packet.May be:An interface name - matches traffic entering the firewall
on the specified interface. May not be used in classify rules or
in rules using the :T chain qualifier.A comma-separated list of host or network IP addresses or
MAC addresses. This form will not match
traffic that originates on the firewall itself unless either
<major><minor> or the :T chain qualifier is used in
the MARK column.Examples:0.0.0.0/0192.168.1.0/24, 172.20.4.0/24An interface name followed by a colon (":") followed by a
comma-separated list of host or network IP addresses or MAC
addresses. May not be used in classify rules or in rules using
the :T chain qualifier.$FW optionally followed by a colon (":") and a
comma-separated list of host or network IP addresses. matches
packets originating on the firewall. May not be used with a
chain qualifier (:P, :F, etc.) in the MARK column.MAC addresses must be prefixed with "~" and use "-" as a
separator.Example: ~00-A0-C9-15-39-78If your kernel includes iprange match support, then address
ranges may be included in the address lists.DEST - Destination of the packet.May be:An interface name. May not be used in the PREROUTING chain
(:P in the mark column or no chain qualifier and
MARK_IN_FORWARD_CHAIN=No in shorewall.conf (5)). The
interface name may be optionally followed by a colon (":") and
an IP address list.A comma-separated list of host or network IP addresses.
The list may include ip address ranges if your kernel and
iptables include iprange support.PROTO - Protocol - Must be "tcp", "udp", "icmp", "ipp2p",
"ipp2p:udp", "ipp2p:all" a number, or "all". "ipp2p" requires ipp2p
match support in your kernel and iptables.PORT(S) - Destination Ports. A comma-separated list of Port
names (from /etc/services), port numbers or port ranges; if the
protocol is "icmp", this column is interpreted as the destination
icmp-type(s).If the protocol is ipp2p, this column is interpreted as an
ipp2p option without the leading "--" (example "bit" for
bit-torrent). If no PORT is given, "ipp2p" is assumed. Note that the
xtables-addons version of IPP2P does not support the "ipp2p" option;
if the column is empty or contains "ipp2p" when using that version
of IPP2P, Shorewall will substitute "edk,kazaa,gnu,dc".This column is ignored if PROTOCOL = all but must be entered
if any of the following field is supplied. In that case, it is
suggested that this field contain "-"CLIENT PORT(S) - (Optional) Port(s) used by the client. If
omitted, any source port is acceptable. Specified as a
comma-separate list of port names, port numbers or port
ranges.USER/GROUP (Optional) This column may only be non-empty if the
SOURCE is the firewall itself. When this column is non-empty, the
rule applies only if the program generating the output is running
under the effective user and/or group. It may contain :[!][<user name or number>]:[<group name or
number>][+<program name>]The colon is optional when specifying only a user.Examples:joe #program must be run by joe
:kids #program must be run by a member of the 'kids' group
!:kids #program must not be run by a member of the 'kids' group
+upnpd #program named upnpd (This feature was removed from Netfilter in kernel version 2.6.14).TEST (Optional) Defines a test on the existing packet or
connection mark. The rule will match only if the test returns true.
Tests have the format [!]<value>[/<mask>][:C]Where:! Inverts the test (not equal)<value> Value of the packet or connection
mark.<mask> A mask to be applied to the mark before
testing:C Designates a connection mark. If omitted, the packet
mark's value is tested.LENGTH (Optional) This field, if present, allows you to match
the length of a packet against a specific value or range of values.
A range is specified in the form <min>:<max> where
either <min> or <max> (but not both) may be omitted. If
<min> is omitted, then 0 is assumed; if <max> is
omitted, than any packet that is <min> or longer will
match.You must have iptables length support for this to work. If you
let it empty or place an "-" here, no length match will be
done.Examples: 1024, 64:1500, :100TOS (Optional) Type of Service. Either a standard name, or a
numeric value to match.
HELPER (Optional). Names one of the Netfilter protocol helper
modules such as ftp, sip,
amanda, etc.HEADERS (Optioinal, Shorewall6 only, added in Shorewall
4.4.15). List of IPv6 headers that may appear in packets. See shorewall6-tcrules
(5) for details.All packets arriving on eth1 should be marked with 1. All
packets arriving on eth2 and eth3 should be marked with 2. All packets
originating on the firewall itself should be marked with 3.#MARK SOURCE DESTINATION PROTOCOL PORT(S)
1 eth1 0.0.0.0/0 all
2 eth2 0.0.0.0/0 all
2 eth3 0.0.0.0/0 all
3 $FW 0.0.0.0/0 allAll GRE (protocol 47) packets destined for 155.186.235.151
should be marked with 12.#MARK SOURCE DESTINATION PROTOCOL PORT(S)
12:T 0.0.0.0/0 155.182.235.151 47All SSH request packets originating in 192.168.1.0/24 and
destined for 155.186.235.151 should be marked with 22.#MARK SOURCE DESTINATION PROTOCOL PORT(S)
22:T 192.168.1.0/24 155.182.235.151 tcp 22All SSH packets packets going out of the first device in in
/etc/shorewall/tcdevices should be assigned to the class with mark
value 10.#MARK SOURCE DESTINATION PROTOCOL PORT(S) CLIENT
# PORT(S)
1:110 0.0.0.0/0 0.0.0.0/0 tcp 22
1:110 0.0.0.0/0 0.0.0.0/0 tcp - 22Mark all ICMP echo traffic with packet mark 1. Mark all peer to
peer traffic with packet mark 4.This is a little more complex than otherwise expected. Since the
ipp2p module is unable to determine all packets in a connection are
P2P packets, we mark the entire connection as P2P if any of the
packets are determined to match. We assume packet/connection mark 0 to
means unclassified. Traffic originating on the firewall is not covered
by this example.#MARK SOURCE DESTINATION PROTOCOL PORT(S) CLIENT USER/ TEST
# PORT(S) GROUP
1 0.0.0.0/0 0.0.0.0/0 icmp echo-request
1 0.0.0.0/0 0.0.0.0/0 icmp echo-reply
RESTORE 0.0.0.0/0 0.0.0.0/0 all - - - 0
CONTINUE 0.0.0.0/0 0.0.0.0/0 all - - - !0
4 0.0.0.0/0 0.0.0.0/0 ipp2p:all
SAVE 0.0.0.0/0 0.0.0.0/0 all - - - !0The last four rules can be translated as:
"If a packet hasn't been classified (packet mark is 0), copy
the connection mark to the packet mark. If the packet mark is set,
we're done. If the packet is P2P, set the packet mark to 4. If the
packet mark has been set, save it to the connection mark."
Mark all forwarded VOIP connections with connection mark 1 and
ensure that all VOIP packets also receive that mark (assumes that
nf_conntrack_sip is loaded).#MARK SOURCE DESTINATION PROTOCOL PORT(S) CLIENT USER/ TEST CONNBYTES TOS HELPER
# PORT(S) GROUP
RESTORE 0.0.0.0/0 0.0.0.0/0 all - - - 0
CONTINUE 0.0.0.0/0 0.0.0.0/0 all - - - !0
1 0.0.0.0/0 0.0.0.0/0 all - - - - - - sip
SAVE 0.0.0.0/0 0.0.0.0/0 all - - - !0ppp devicesIf you use ppp/pppoe/pppoa) to connect to your Internet provider
and you use traffic shaping you need to restart shorewall traffic
shaping. The reason for this is, that if the ppp connection gets
restarted (and it usually does this at least daily), all
tc filters/qdiscs related to that interface are
deleted.The easiest way to achieve this, is just to restart shorewall once
the link is up. To achieve this add a small executable script
to/etc/ppp/ip-up.d.#! /bin/sh
/sbin/shorewall refreshSharing a TC configuration between Shorewall and
Shorewall6Beginning with Shorewall 4.4.15, the traffic-shaping configuration
in the tcdevices, tcclasses and tcfilters files can be shared between
Shorewall and Shorewall6. Only one of the products can control the
configuration but the other can configure CLASSIFY rules in its own
tcrules file that refer to the shared classes.To defined the configuration in Shorewall and shared it with
Shorewall6:Set TC_ENABLED=Internal in shorewall.conf
(5).Set TC_ENABLED=Shared in shorewall6.conf
(5).Create symbolic links from /etc/shorewall6 to
/etc/shorewall/tcdevices and /etc/shorewall/tcclasses:ln -s ../shorewall/tcdevices /etc/shorewall6/tcdevices
ln -s ../shorewall/tcclasses /etc/shorewall6/tcclassesIf you need to define IPv6 tcfilter entries, do so in
/etc/shorewall/tcfilters. That file now allows entries that apply to
IPv6.Shorewall6 compilations to have access to the tcdevices and
tcclasses files although it will create no output. That access allows
CLASSIFY rules in /etc/shorewall6/tcrules to be validated against the TC
configuration.In this configuration, it is Shorewall that controls TC
configuration (except for IPv6 tcrules). You can reverse the settings in
the files if you want to control the configuration using
Shorewall6.Per-IP Traffic ShapingSome network administrators feel that they have to divy up their
available bandwidth by IP address rather than by prioritizing the
traffic based on the type of traffic. This gets really awkward when
there are a large number of local IP addresses.This section describes the Shorewall facility for making this
configuration less tedious (and a lot more efficient). Note that it
requires that you install
xtables-addons. So before you try this facility, we suggest that
first you add the following OPTION to each external interface described
in /etc/shorewall/tcdevices:flow=nfct-srcIf you shape traffic on your internal interface(s), then add this
to their entries:flow=dstYou may find that this simple change is all that is needed to
control bandwidth hogs like Bit Torrent. If it doesn't, then proceed as
described in this section.The facility has two components:An IPMARK MARKing command in
/etc/shorewall/tcrules.An occurs OPTION in
/etc/shorewall/tcclasses.The facility is currently only available with IPv4.In a sense, the IPMARK target is more like an IPCLASSIFY target in
that the mark value is later interpreted as a class ID. A packet mark is
32 bits wide; so is a class ID. The major class
occupies the high-order 16 bits and the minor class
occupies the low-order 16 bits. So the class ID 1:4ff (remember that
class IDs are always in hex) is equivalent to a mark value of 0x104ff.
Remember that Shorewall uses the interface number as the
major number where the first interface in tcdevices
has major number 1, the second has
major number 2, and so on.The IPMARK target assigns a mark to each matching packet based on
the either the source or destination IP address. By default, it assigns
a mark value equal to the low-order 8 bits of the source address.The syntax is as follows:
Default values are:srcmask1 = 0xFFmask2 = 0x00shift = 0src and dst specify whether the mark is to be based on
the source or destination address respectively. The selected address is
first shifted right by shift, then LANDed with
mask1 and then LORed with
mask2. The shift argument is
intended to be used primarily with IPv6 addresses.Example:IPMARK(src,0xff,0x10100)
Source IP address is 192.168.4.3 = 0xc0a80403
0xc0a80403 >> 0 = 0xc0a80403
0xc0a80403 LAND 0xFF = 0x03
0x03 LOR 0x10100 = 0x10103
So the mark value is 0x10103 which corresponds to class id 1:103.It is important to realize that, while class IDs are composed of a
major and a minor value, the
set of minor values must be unique. You must keep
this in mind when deciding how to map IP addresses to class IDs. For
example, suppose that your internal network is 192.168.1.0/29 (host IP
addresses 192.168.1.1 - 192.168.1.6). Your first notion might be to use
IPMARK(src,0xFF,0x10000) so as to produce class IDs 1:1 through 1:6. But
1:1 is the class ID of the base HTB class on interface 1. So you might
chose instead to use IPMARK(src,0xFF,0x10100) as shown in the example
above so as to avoid minor class 1.The occurs option in
/etc/shorewall/tcclasses causes the class
definition to be replicated many times.The synax is:
occurs=number
When occurs is used:The associated device may not have the classify option.The class may not be the default class.The class may not have any tos= options (including tcp-ack).The class should not specify a MARK value. Any MARK value given is
ignored with a warning. The RATE and CEIL parameters apply to each
instance of the class. So the total RATE represented by an entry with
occurs will be the listed RATE
multiplied by number.Example:/etc/shorewall/tcdevices:#INTERFACE IN-BANDWIDTH OUT-BANDWIDTH
eth0 100mbit 100mbit/etc/shorewall/tcclasses:#DEVICE MARK RATE CEIL PRIORITY OPTIONS
eth0:101 - 1kbit 230kbit 4 occurs=6The above defines 6 classes with class IDs 0x101-0x106. Each class
has a guaranteed rate of 1kbit/second and a ceiling of 230kbit./etc/shoreall/tcrules:#MARK SOURCE DEST
IPMARK(src,0xff,0x10100):F 192.168.1.0/29 eth0This facility also alters the way in which Shorewall generates a
class number when none is given. Prior to the implementation of this
facility, the class number was constructed by concatinating the MARK
value with the either '1' or '10'. '10' was used when there were more
than 10 devices defined in
/etc/shorewall/tcdevices.With this facility, a new method is added; class numbers are
assigned sequentially beginning with 2. The WIDE_TC_MARKS option in
shorewall.conf selects which construction to use.
WIDE_TC_MARKS=No (the default) produces pre-Shorewall 4.4 behavior.
WIDE_TC_MARKS=Yes produces the new behavior.Real life examplesA Shorewall User's ExperienceChuck Kollars has provided an excellent
writeup about his traffic shaping experiences.Configuration to replace WondershaperYou are able to fully replace the wondershaper script by using
the buitin traffic control.. In this example it is assumed that your
interface for your Internet connection is ppp0 (for DSL), if you use
another connection type, you have to change it. You also need to
change the settings in the tcdevices.wondershaper file to reflect your
line speed. The relevant lines of the config files follow here. Please
note that this is just a 1:1 replacement doing exactly what
wondershaper should do. You are free to change it...tcdevices file#INTERFACE IN-BANDWITH OUT-BANDWIDTH
ppp0 5000kbit 500kbittcclasses file#INTERFACE MARK RATE CEIL PRIORITY OPTIONS
ppp0 1 5*full/10 full 1 tcp-ack,tos-minimize-delay
ppp0 2 3*full/10 9*full/10 2 default
ppp0 3 2*full/10 8*full/10 2tcrules file#MARK SOURCE DEST PROTO PORT(S) CLIENT USER
# PORT(S)
1:F 0.0.0.0/0 0.0.0.0/0 icmp echo-request
1:F 0.0.0.0/0 0.0.0.0/0 icmp echo-reply
# mark traffic which should have a lower priority with a 3:
# mldonkey
3 0.0.0.0/0 0.0.0.0/0 udp - 4666Wondershaper allows you to define a set of hosts and/or ports
you want to classify as low priority. To achieve this , you have to
add these hosts to tcrules and set the mark to 3 (true if you use
the example configuration files).Setting hosts to low prioritylets assume the following settings from your old wondershaper
script (don't assume these example values are really useful, they
are only used for demonstrating ;-):
# low priority OUTGOING traffic - you can leave this blank if you want
# low priority source netmasks
NOPRIOHOSTSRC="192.168.1.128/25 192.168.3.28"
# low priority destination netmasks
NOPRIOHOSTDST=60.0.0.0/24
# low priority source ports
NOPRIOPORTSRC="6662 6663"
# low priority destination ports
NOPRIOPORTDST="6662 6663" This would result in the following additional settings to the
tcrules file:3 192.168.1.128/25 0.0.0.0/0 all
3 192.168.3.28 0.0.0.0/0 all
3 0.0.0.0/0 60.0.0.0/24 all
3 0.0.0.0/0 0.0.0.0/0 udp 6662,6663
3 0.0.0.0/0 0.0.0.0/0 udp - 6662,6663
3 0.0.0.0/0 0.0.0.0/0 tcp 6662,6663
3 0.0.0.0/0 0.0.0.0/0 tcp - 6662,6663A simple setupThis is a simple setup for people sharing an Internet connection
and using different computers for this. It just basically shapes
between 2 hosts which have the ip addresses 192.168.2.23 and
192.168.2.42tcdevices file#INTERFACE IN-BANDWITH OUT-BANDWIDTH
ppp0 6000kbit 700kbitWe have 6mbit down and 700kbit upstream.tcclasses file#INTERFACE MARK RATE CEIL PRIORITY OPTIONS
ppp0 1 10kbit 50kbit 1 tcp-ack,tos-minimize-delay
ppp0 2 300kbit full 2
ppp0 3 300kbit full 2
ppp0 4 90kbit 200kbit 3 defaultWe add a class for tcp ack packets with highest priority, so
that downloads are fast. The following 2 classes share most of the
bandwidth between the 2 hosts, if the connection is idle, they may
use full speed. As the hosts should be treated equally they have the
same priority. The last class is for the remaining traffic.tcrules file#MARK SOURCE DEST PROTO PORT(S) CLIENT USER
# PORT(S)
1:F 0.0.0.0/0 0.0.0.0/0 icmp echo-request
1:F 0.0.0.0/0 0.0.0.0/0 icmp echo-reply
2:F 192.168.2.23 0.0.0.0/0 all
3:F 192.168.2.42 0.0.0.0/0 allWe mark icmp ping and replies so they will go into the fast
interactive class and set a mark for each host.A Warning to Xen UsersIf you are running traffic shaping in your dom0 and traffic shaping
doesn't seem to be limiting outgoing traffic properly, it may be due to
"checksum offloading" in your domU(s). Check the output of "shorewall show
tc". Here's an excerpt from the output of that command:class htb 1:130 parent 1:1 leaf 130: prio 3 quantum 1500 rate 76000bit ceil 230000bit burst 1537b/8 mpu 0b overhead 0b cburst 1614b/8 mpu 0b overhead 0b level 0
Sent 559018700 bytes 75324 pkt (dropped 0, overlimits 0 requeues 0)
rate 299288bit 3pps backlog 0b 0p requeues 0
lended: 53963 borrowed: 21361 giants: 90174
tokens: -26688 ctokens: -14783There are two obvious problems in the above output:The rate (299288) is considerably larger than the ceiling
(230000).There are a large number (90174) of giants reported.This problem will be corrected by disabling "checksum offloading" in
your domU(s) using the ethtool utility. See the one of the Xen articles for
instructions.An HFSC ExampleAs mentioned at the top of this article, there is an excellent
introduction to HFSC at http://linux-ip.net/articles/hfsc.en/.
At the end of that article are 'tc' commands that implement the
configuration in the article. Those tc commands correspond to the
following Shorewall traffic shaping configuration./etc/shorewall/tcdevices:#INTERFACE IN-BANDWITH OUT-BANDWIDTH OPTIONS
eth0 - 1000kbit hfsc/etc/shorewall/tcclasses:#INTERFACE:CLASS MARK RATE: CEIL PRIORITY OPTIONS
# DMAX:UMAX
1:10 1 500kbit full 1
1:20 2 500kbit full 1
1:10:11 3 400kbit:53ms:1500b full 2
1:10:12 4 100kbit:30ms:1500b full 2The following sub-section offers some notes about the
article.Where Did all of those Magic Numbers come from?As you read the article, numbers seem to be introduced out of thin
air. I'll try to shed some light on those.There is very clear development of these numbers:12ms to transfer a 1500b packet at 1000kbits/second.100kbits per second with 1500b packets, requires 8 packets per
second.A packet from class 1:12 must be sent every 120ms.Total transmit delay can be no more than 132ms (120 +
12).We then learn that the queuing latency can be reduced to 30ms if
we use a two-part service curve whose first part is 400kbits/second.
Where did those come from?The latency is calculated from the rate. If it takes 12ms to
transmit a 1500 byte packet at 1000kbits/second, it takes 30ms to
transmit a 1500b at 400kbits/second.For the slope of the first part of the service curve, in
theory we can pick any number between 100 (the rate of class 1:12)
and 500 (the rate of the parent class) with higher numbers providing
lower latency.The final curious number is the latency for class 1:11 - 52.5ms.
It is a consequence of everything that has gone before.To acheive 400kbits/second with 1500-byte packets, 33.33 packets
per second are required. So a packet from class 1:11 must be sent every
30 ms. As the article says, "...the maximum transmission delay of this
class increases from 30ms to a total of 52.5 ms.". So we are looking for
an additional 22.5 ms.Assume that both class 1:11 and 1:12 transmit for 30 ms at
400kbits/second. That is a total of 800kbits/second for 30ms. So Class
1:11 is punished for the excess. How long is the punishment? The two
classes sent 24,000 bits in 30ms; they are only allowed 0.030 * 500,000
= 15,000. So they are 9,000 bits over their quota. The amount of time
required to transmit 9,000 bits at 400,000 bits/second is
22.5ms!.Intermediate Functional Block (IFB) DevicesThe principles behind an IFB is fairly simple:It looks like a network interface although it is never given an
IPv4 configuration.Because it is a network interface, queuing disciplines can be
associated with an IFB.The magic of an IFB comes in the fact that a filter may be defined
on a real network interface such that each packet that arrives on that
interface is queued for the IFB! In that way, the IFB provides a means for
shaping input traffic.To use an IFB, you must have IFB support in your kernel
(configuration option CONFIG_IFB). Assuming that you have a modular
kernel, the name of the IFB module is 'ifb' and may be loaded using the
command modprobe ifb (if you have modprobe installed)
or insmod /path/to/module/ifb.By default, two IFB devices (ifb0 and ifb1) are created. You can
control that using the numifbs option (e.g., modprobe ifb
numifbs=1).To create a single IFB when Shorewall starts, place the following
two commands in /etc/shorewall/init:modprobe ifb numifbs=1
ip link set ifb0 upEntries in /etc/shorewall/tcrules have no
effect on shaping traffic through an IFB. To allow classification of such
traffic, the /etc/shorewall/tcfilters file has been added. Entries in that
file create u32
classification rules./etc/shorewall/tcfiltersWhile this file was created to allow shaping of traffic through an
IFB, the file may be used for general traffic classification as well.
The file is similar to shorewall-tcrules(5) with the
following key exceptions:The first match determines the classification, whereas in the
tcrules file, the last match determines the classification.ipsets are not supportedDNS Names are not supportedAddress ranges and lists are not supportedExclusion is not supported.filters are applied to packets as they appear on the
wire. So incoming packets will not have DNAT applied yet
(the destination IP address will be the external address) and
outgoing packets will have had SNAT applied.The last point warrants elaboration. When looking at traffic being
shaped by an IFB, there are two cases to consider:Requests — packets being sent from remote clients to local
servers. These packets may undergo subsequent DNAT, either as a
result of entries in /etc/shorewall/nat or as a
result of DNAT or REDIRECT rules.Example: /etc/shorewall/rules:#ACTION SOURCE DEST PROTO DEST SOURCE ORIGINAL
# PORT(S) PORT(S) DEST
DNAT net dmz:192.168.4.5 tcp 80 - 206.124.146.177Requests redirected by this rule will have destination IP
address 206.124.146.177 and destination port 80.Responses — packets being sent from remote servers to local
clients. These packets may undergo subsequent DNAT as a result of
entries in /etc/shorewall/nat or in
/etc/shorewall/masq. The packet's destination
IP address will be the external address specified in the
entry.Example:
/etc/shorewall/masq:#INTERFACE SOURCE ADDRESS
eth0 192.168.1.0/24 206.124.146.179HTTP response packets corresponding to requests that fall
under that rule will have destination IP address 206.124.146.179 and
source port 80.Beginning with Shorewall 4.4.15, both IPv4 and IPv6 rules can be
defined in this file. See shorewall-tcfilters (5)
for details.Columns in the file are as follow. As in all Shorewall
configuration files, a hyphen ("-") may be used to indicate that no
value is supplied in the column.CLASSThe interface name or number followed by a colon (":") and
the class number.SOURCESOURCE IP address (host or network). DNS names are not
allowed.DESTDESTINATION IP address (host or network). DNS names are not
allowed.PROTOProtocol name or number.DEST PORT(S)Comma-separated list of destination port names or numbers.
May only be specified if the protocol is TCP, UDP, SCTP or ICMP.
Port ranges are supported except for ICMP.SOURCE PORTComma-separated list of source port names or numbers. May
only be specified if the protocol is TCP, UDP or SCTP. Port ranges
are supported.TOSSpecifies the value of the TOS field. The value can be any
of the following:hex-numberhex-number/hex-numberThe hex-numbers must be exactly
two digits (e.g., 0x04).LENGTHMust be a power of 2 between 32 and 8192 inclusive. Packets
with a total length that is strictly less than the specified value
will match the rule.Example:I've used this configuration on my own firewall. The IFB portion
is more for test purposes rather than to serve any well-reasoned QOS
strategy./etc/shorewall/init:qt modprobe ifb numifbs=1
qt ip link set dev ifb0 up/etc/shorewall/interfaces:#ZONE INTERFACE BROADCAST
- ifb0/etc/shorewall/tcdevices:
#INTERFACE IN-BANDWITH OUT-BANDWIDTH OPTIONS REDIRECTED
# INTERFACES
1:eth0 - 384kbit classify
2:ifb0 - 1300kbit - eth0/etc/shorewall/tcclasses:#INTERFACE MARK RATE CEIL PRIORITY OPTIONS
1:110 - 5*full/10 full 1 tcp-ack,tos-minimize-delay
1:120 - 2*full/10 6*full/10 2 default
1:130 - 2*full/10 6*full/10 3
2:110 - 5*full/10 full 1 tcp-ack,tos-minimize-delay
2:120 - 2*full/10 6*full/10 2 default
2:130 - 2*full/10 6*full/10 3/etc/shorewall/tcfilters:#INTERFACE: SOURCE DEST PROTO DEST SOURCE
#CLASS PORT(S) PORT(S)
#
# OUTGOING TRAFFIC
#
1:130 206.124.146.178 - tcp - 49441,49442 #BITTORRENT on wookie
1:110 206.124.146.178 #wookie
1:110 206.124.146.179 #SNAT of internal systems
1:110 206.124.146.180 #Work Laptop
1:110 - - icmp echo-request,echo-reply
1:110 - - icmp echo-reply
1:130 206.124.146.177 - tcp - 873,25 #Bulk Traffic
#
# INCOMING TRAFFIC
#
2:110 - 206.124.146.178 #Wookie
2:110 - 206.124.146.179 #SNAT Responses
2:110 - 206.124.146.180 #Work Laptop
2:130 - 206.124.146.177 tcp 25 #Incoming Email.You can examine the installed filters with the shorewall
show filters command. What follows shows the output for
eth0 with the filters shown
above. Bold font are comments
explaining the rules.gateway:~ # shorewall-lite show filters
Shorewall Lite 4.1.6 Classifiers at gateway - Fri Mar 21 08:06:47 PDT 2008
Device eth1:
Device eth2:
Device eth0:
filter parent 1: protocol ip pref 10 u32
filter parent 1: protocol ip pref 10 u32 fh 3: ht divisor 1 <========= Start of table 3. parses TCP header
filter parent 1: protocol ip pref 10 u32 fh 3::800 order 2048 key ht 3 bkt 0 flowid 1:130 (rule hit 102 success 0)
match 03690000/ffff0000 at nexthdr+0 (success 0 ) <========= SOURCE PORT 873 goes to class 1:130
filter parent 1: protocol ip pref 10 u32 fh 2: ht divisor 1 <========= Start of table 2. parses ICMP header
filter parent 1: protocol ip pref 10 u32 fh 2::800 order 2048 key ht 2 bkt 0 flowid 1:110 (rule hit 0 success 0)
match 08000000/ff000000 at nexthdr+0 (success 0 ) <========= ICMP Type 8 goes to class 1:110
filter parent 1: protocol ip pref 10 u32 fh 2::801 order 2049 key ht 2 bkt 0 flowid 1:110 (rule hit 0 success 0)
match 00000000/ff000000 at nexthdr+0 (success 0 ) <========= ICMP Type 0 goes to class 1:110
filter parent 1: protocol ip pref 10 u32 fh 1: ht divisor 1 <========= Start of table 1. parses TCP header
filter parent 1: protocol ip pref 10 u32 fh 1::800 order 2048 key ht 1 bkt 0 flowid 1:130 (rule hit 0 success 0)
match c1210000/ffff0000 at nexthdr+0 (success 0 ) <========= SOURCE PORT 49441 goes to class 1:130
filter parent 1: protocol ip pref 10 u32 fh 1::801 order 2049 key ht 1 bkt 0 flowid 1:130 (rule hit 0 success 0)
match c1220000/ffff0000 at nexthdr+0 (success 0 ) <========= SOURCE PORT 49442 goes to class 1:130
filter parent 1: protocol ip pref 10 u32 fh 800: ht divisor 1 <========= Start of Table 800. Packets start here!=============== The following 2 rules are generated by the class definition in /etc/shorewall/classes ==================
filter parent 1: protocol ip pref 10 u32 fh 800::800 order 2048 key ht 800 bkt 0 flowid 1:110 (rule hit 2204 success 138)
match 00060000/00ff0000 at 8 (success 396 ) <========= TCP
match 05000000/0f00ffc0 at 0 (success 250 ) <========= Header length 20 and Packet Length < 64
match 00100000/00ff0000 at 32 (success 138 ) <========= ACK
filter parent 1: protocol ip pref 10 u32 fh 800::801 order 2049 key ht 800 bkt 0 flowid 1:110 (rule hit 2066 success 0)
match 00100000/00100000 at 0 (success 0 ) <========= Minimize-delay goes to class 1:110 =============== Jump to Table 1 if the matches are met ==================
filter parent 1: protocol ip pref 10 u32 fh 800::802 order 2050 key ht 800 bkt 0 link 1: (rule hit 2066 success 0)
match ce7c92b2/ffffffff at 12 (success 1039 ) <========= SOURCE 206.124.146.178
match 00060000/00ff0000 at 8 (success 0 ) <========= PROTO TCP
offset 0f00>>6 at 0 eat
filter parent 1: protocol ip pref 10 u32 fh 800::803 order 2051 key ht 800 bkt 0 flowid 1:110 (rule hit 2066 success 1039)
match ce7c92b2/ffffffff at 12 (success 1039 ) <========= SOURCE 206.124.146.178 goes to class 1:110
filter parent 1: protocol ip pref 10 u32 fh 800::804 order 2052 key ht 800 bkt 0 flowid 1:110 (rule hit 1027 success 132)
match ce7c92b3/ffffffff at 12 (success 132 ) <========= SOURCE 206.124.146.179 goes to class 1:110
filter parent 1: protocol ip pref 10 u32 fh 800::805 order 2053 key ht 800 bkt 0 flowid 1:110 (rule hit 895 success 603)
match ce7c92b4/ffffffff at 12 (success 603 ) <========= SOURCE 206.124.146.180 goes to class 1:110 =============== Jump to Table 2 if the matches are met ==================
filter parent 1: protocol ip pref 10 u32 fh 800::806 order 2054 key ht 800 bkt 0 link 2: (rule hit 292 success 0)
match 00010000/00ff0000 at 8 (success 0 ) <========= PROTO ICMP
offset 0f00>>6 at 0 eat
=============== Jump to Table 3 if the matches are met ==================
filter parent 1: protocol ip pref 10 u32 fh 800::807 order 2055 key ht 800 bkt 0 link 3: (rule hit 292 success 0)
match ce7c92b1/ffffffff at 12 (success 265 ) <========= SOURCE 206.124.146.177
match 00060000/00ff0000 at 8 (success 102 ) <========= PROTO TCP
offset 0f00>>6 at 0 eat Understanding the output of 'shorewall show tc'The shorewall show tc (shorewall-lite
show tc) command displays information about the current state of
traffic shaping. For each device, it executes the following
commands: echo Device $device:
tc -s -d qdisc show dev $device
echo
tc -s -d class show dev $device
echo So, the traffic-shaping output is generated entirely by the
tc utility.Here's the output of for eth0. The configuration is the one shown in
the preceding section (the output was obtained almost 24 hours later than
the shorewall show filters output shown above).Device eth0:
============== The primary queuing discipline is HTB (Hierarchical Token Bucket) ====================
qdisc htb 1: r2q 10 default 120 direct_packets_stat 9 ver 3.17
Sent 2133336743 bytes 4484781 pkt (dropped 198, overlimits 4911403 requeues 21) <=========== Note the overlimits and dropped counts
rate 0bit 0pps backlog 0b 8p requeues 21
============== The ingress filter. If you specify IN-BANDWIDTH, you can see the 'dropped' count here. =========In this case, the packets are being sent to the IFB for shaping
qdisc ingress ffff: ----------------
Sent 4069015112 bytes 4997252 pkt (dropped 0, overlimits 0 requeues 0)
rate 0bit 0pps backlog 0b 0p requeues 0
============ Each of the leaf HTB classes has an SFQ qdisc to ensure that each flow gets its turn ============
qdisc sfq 110: parent 1:110 limit 128p quantum 1514b flows 128/1024 perturb 10sec
Sent 613372519 bytes 2870225 pkt (dropped 0, overlimits 0 requeues 6)
rate 0bit 0pps backlog 0b 0p requeues 6
qdisc sfq 120: parent 1:120 limit 128p quantum 1514b flows 128/1024 perturb 10sec
Sent 18434920 bytes 60961 pkt (dropped 0, overlimits 0 requeues 0)
rate 0bit 0pps backlog 0b 0p requeues 0
qdisc sfq 130: parent 1:130 limit 128p quantum 1514b flows 128/1024 perturb 10sec
Sent 1501528722 bytes 1553586 pkt (dropped 198, overlimits 0 requeues 15)
rate 0bit 0pps backlog 11706b 8p requeues 15
============= Class 1:110 -- the high-priority class ===========
Note the rate and ceiling calculated from 'full'
class htb 1:110 parent 1:1 leaf 110: prio 1 quantum 4800 rate 192000bit ceil 384000bit burst 1695b/8 mpu 0b overhead 0b cburst 1791b/8 mpu 0b overhead 0b level 0
Sent 613372519 bytes 2870225 pkt (dropped 0, overlimits 0 requeues 0)
rate 195672bit 28pps backlog 0b 0p requeues 0 <=========== Note the current rate of traffic. There is no queuing (no packet backlog)
lended: 2758458 borrowed: 111773 giants:
tokens: 46263 ctokens: 35157
============== The root class ============
class htb 1:1 root rate 384000bit ceil 384000bit burst 1791b/8 mpu 0b overhead 0b cburst 1791b/8 mpu 0b overhead 0b level 7
Sent 2133276316 bytes 4484785 pkt (dropped 0, overlimits 0 requeues 0) <==== Total output traffic since last 'restart'
rate 363240bit 45pps backlog 0b 0p requeues 0
lended: 1081936 borrowed: 0 giants: 0
tokens: -52226 ctokens: -52226
============= Bulk Class (outgoing rsync, email and bittorrent) ============
class htb 1:130 parent 1:1 leaf 130: prio 3 quantum 1900 rate 76000bit ceil 230000bit burst 1637b/8 mpu 0b overhead 0b cburst 1714b/8 mpu 0b overhead 0b level 0
Sent 1501528722 bytes 1553586 pkt (dropped 198, overlimits 0 requeues 0)
rate 162528bit 14pps backlog 0b 8p requeues 0 <======== Queuing is occurring (8 packet backlog). The rate is still below the ceiling.
lended: 587134 borrowed: 966459 giants: 0 During peak activity, the rate tops out at around 231000 (just above ceiling).
tokens: -30919 ctokens: -97657
================= Default class (mostly serving web pages) ===============
class htb 1:120 parent 1:1 leaf 120: prio 2 quantum 1900 rate 76000bit ceil 230000bit burst 1637b/8 mpu 0b overhead 0b cburst 1714b/8 mpu 0b overhead 0b level 0
Sent 18434920 bytes 60961 pkt (dropped 0, overlimits 0 requeues 0)
rate 2240bit 2pps backlog 0b 0p requeues 0
lended: 57257 borrowed: 3704 giants: 0
tokens: 156045 ctokens: 54178
Using your own tc scriptReplacing builtin tcstart fileIf you prefer your own tcstart file, just install it in
/etc/shorewall/tcstart.In your tcstart script, when you want to run the tc
utility, use the run_tc function supplied by Shorewall if you want tc
errors to stop the firewall.Set TC_ENABLED=Yes and CLEAR_TC=YesSupply an /etc/shorewall/tcstart script to configure your
traffic shaping rules.Optionally supply an /etc/shorewall/tcclear script to stop
traffic shaping. That is usually unnecessary.If your tcstart script uses the fwmark
classifier, you can mark packets using entries in
/etc/shorewall/tcrules.Traffic control outside ShorewallTo start traffic shaping when you bring up your network
interfaces, you will have to arrange for your traffic shaping
configuration script to be run at that time. How you do that is
distribution dependent and will not be covered here. You then
should:Set TC_ENABLED=No and CLEAR_TC=NoIf your script uses the fwmark classifier, you
can mark packets using entries in /etc/shorewall/tcrules.Testing ToolsAt least one Shorewall user has found this tool helpful: http://e2epi.internet2.edu/network-performance-toolkit.html