.TH "Universal 32bit classifier in tc" 8 "25 Sep 2015" "iproute2" "Linux"

.SH NAME
u32 \- universal 32bit traffic control filter
.SH SYNOPSIS
.in +8
.ti -8
.BR tc " " filter " ... [ " handle
.IR HANDLE " ] "
.B u32
.IR OPTION_LIST " [ "
.B offset
.IR OFFSET " ] [ "
.B hashkey
.IR HASHKEY " ] [ "
.B classid
.IR CLASSID " ] [ "
.B divisor
.IR uint_value " ] [ "
.B order
.IR u32_value " ] [ "
.B ht
.IR HANDLE " ] [ "
.B sample
.IR SELECTOR " [ "
.B divisor
.IR uint_value " ] ] [ "
.B link
.IR HANDLE " ] [ "
.B indev
.IR ifname " ] [ "
.B skip_hw
.R  "|"
.B skip_sw
.R " ] [ "
.BR help " ]"

.ti -8
.IR HANDLE " := { "
\fIu12_hex_htid\fB:\fR[\fIu8_hex_hash\fB:\fR[\fIu12_hex_nodeid\fR] | \fB0x\fIu32_hex_value\fR }

.ti -8
.IR OPTION_LIST " := [ " OPTION_LIST " ] " OPTION

.ti -8
.IR HASHKEY " := [ "
.B mask
.IR u32_hex_value " ] [ "
.B at
.IR 4*int_value " ]"

.ti -8
.IR CLASSID " := { "
.BR root " | "
.BR none " | "
[\fIu16_major\fR]\fB:\fIu16_minor\fR | \fIu32_hex_value\fR }

.ti -8
.IR OFFSET " := [ "
.B plus
.IR int_value " ] [ "
.B at
.IR 2*int_value " ] [ "
.B mask
.IR u16_hex_value " ] [ "
.B shift
.IR int_value " ] [ "
.BR eat " ]"

.ti -8
.IR OPTION " := { "
.B match
.IR SELECTOR " | "
.B action
.IR ACTION " } "

.ti -8
.IR SELECTOR " := { "
.B u32
.IR VAL_MASK_32 " | "
.B u16
.IR VAL_MASK_16 " | "
.B u8
.IR VAL_MASK_8 " | "
.B ip
.IR IP " | "
.B ip6
.IR IP6 " | { "
.BR tcp " | " udp " } "
.IR TCPUDP " | "
.B icmp
.IR ICMP " | "
.B mark
.IR VAL_MASK_32 " | "
.B ether
.IR ETHER " }"

.ti -8
.IR IP " := { { "
.BR src " | " dst " } { " default " | " any " | " all " | "
.IR ip_address " [ "
.BR / " { "
.IR prefixlen " | " netmask " } ] } " AT " | { "
.BR dsfield " | " ihl " | " protocol " | " precedence " | "
.BR icmp_type " | " icmp_code " } "
.IR VAL_MASK_8 " | { "
.BR sport " | " dport " } "
.IR VAL_MASK_16 " | "
.BR nofrag " | " firstfrag " | " df " | " mf " }"

.ti -8
.IR IP6 " := { { "
.BR src " | " dst " } { " default " | " any " | " all " | "
.IR ip6_address " [/" prefixlen " ] } " AT " | "
.B priority
.IR VAL_MASK_8 " | { "
.BR protocol " | " icmp_type " | " icmp_code " } "
.IR VAL_MASK_8 " | "
.B flowlabel
.IR VAL_MASK_32 " | { "
.BR sport " | " dport " } "
.IR VAL_MASK_16 " }"

.ti -8
.IR TCPUDP " := { "
.BR src " | " dst " } "
.I VAL_MASK_16

.ti -8
.IR ICMP " := { "
.B type
.IR VAL_MASK_8 " | "
.B code
.IR VAL_MASK_8 " }"

.ti -8
.IR ETHER " := { "
.BR src " | " dst " } "
.IR ether_address " " AT

.ti -8
.IR VAL_MASK_32 " := " u32_value " " u32_hex_mask " [ " AT " ]"

.ti -8
.IR VAL_MASK_16 " := " u16_value " " u16_hex_mask " [ " AT " ]"

.ti -8
.IR VAL_MASK_8 " := " u8_value " " u8_hex_mask " [ " AT " ]"

.ti -8
.IR AT " := [ "
.BR at " [ " nexthdr+ " ] "
.IR int_value " ]"
.SH DESCRIPTION
The Universal/Ugly 32bit filter allows to match arbitrary bitfields in the
packet. Due to breaking everything down to values, masks and offsets, It is
equally powerful and hard to use. Luckily many abstracting directives are
present which allow defining rules on a higher level and therefore free the
user from having to fiddle with bits and masks in many cases.

There are two general modes of invocation: The first mode creates a new filter
to delegate packets to different destinations. Apart from the obvious ones,
namely classifying the packet by specifying a
.I CLASSID
or calling an
.BR action ,
one may
.B link
one filter to another one (or even a list of them), effectively organizing
filters into a tree-like hierarchy.

Typically filter delegation is done by means of a hash table, which leads to the
second mode of invocation: it merely serves to set up these hash tables. Filters
can select a hash table and provide a key selector from which a hash is to be
computed and used as key to lookup the table's bucket which contains filters for
further processing. This is useful if a high number of filters is in use, as the
overhead of performing the hash operation and table lookup becomes negligible in
that case. Using hashtables with
.B u32
basically involves the following pattern:
.IP (1) 4
Creating a new hash table, specifying it's size using the
.B divisor
parameter and ideally a handle by which the table can be identified. If the
latter is not given, the kernel chooses one on it's own, which has to be
guessed later.
.IP (2) 4
Creating filters which link to the created table in
.I (1)
using the
.B link
parameter and defining the packet data which the kernel will use to calculate
the
.BR hashkey .
.IP (3) 4
Adding filters to buckets in the hash table from
.IR (1) .
In order to avoid having to know how exactly the kernel creates the hash key,
there is the
.B sample
parameter, which gives sample data to hash and thereby define the table bucket
the filter should be added to.

.RE
In fact, even if not explicitly requested
.B u32
creates a hash table for every
.B priority
a filter is being added with. The table's size is 1 though, so it is in fact
merely a linked list.
.SH VALUES
Options and selectors require values to be specified in a specific format, which
is often non-intuitive. Therefore the terminals in
.I SYNOPSIS
have been given descriptive names to indicate the required format and/or maximum
allowed numeric value: Prefixes
.IR u32 ", " u16 " and " u8
indicate four, two and single byte unsigned values. E.g.
.I u16
indicates a two byte-sized value in range between 0 and 65535 (0xFFFF)
inclusive. A prefix of
.I int
indicates a four byte signed value. A middle part of
.I _hex_
indicates that the value is parsed in hexadecimal format. Otherwise, the
value's base is automatically detected, i.e. values prefixed with
.I 0x
are considered hexadecimal, a leading
.I 0
indicates octal format and decimal format otherwise. There are some values with
special formatting as well:
.IR ip_address " and " netmask
are in dotted-quad formatting as usual for IPv4 addresses. An
.I ip6_address
is specified in common, colon-separated hexadecimal format. Finally,
.I prefixlen
is an unsigned, decimal integer value in range from 0 to the address width in
bits (32 for IPv4 and 128 for IPv6).

Sometimes values need to be dividable by a certain number. In that case a name
of the form
.I N*val
was chosen, indicating that
.I val
must be dividable by
.IR N .
Or the other way around: the resulting value must be a multiple of
.IR N .
.SH OPTIONS
.B U32
recognizes the following options:
.TP
.BI handle " HANDLE"
The handle is used to reference a filter and therefore must be unique. It
consists of a hash table identifier
.B htid
and optional
.B hash
(which identifies the hash table's bucket) and
.BR nodeid .
All these values are parsed as unsigned, hexadecimal numbers with length 12bits
(
.BR htid " and " nodeid )
or 8bits (
.BR hash ).
Alternatively one may specify a single, 32bit long hex number which contains
the three fields bits in concatenated form. Other than the fields themselves, it
has to be prefixed by
.BR 0x .
.TP
.BI offset " OFFSET"
Set an offset which defines where matches of subsequent filters are applied to.
Therefore this option is useful only when combined with
.BR link " or a combination of " ht " and " sample .
The offset may be given explicitly by using the
.B plus
keyword, or extracted from the packet data with
.BR at .
It is possible to mangle the latter using
.BR mask " and/or " shift
keywords. By default, this offset is recorded but not implicitly applied. It is
used only to substitute the
.B nexthdr+
statement. Using the keyword
.B eat
though inverses this behaviour: the offset is applied always, and
.B nexthdr+
will fall back to zero.
.TP
.BI hashkey " HASHKEY"
Spefify what packet data to use to calculate a hash key for bucket lookup. The
kernel adjusts the value according to the hash table's size. For this to work,
the option
.B link
must be given.
.TP
.BI classid " CLASSID"
Classify matching packets into the given
.IR CLASSID ,
which consists of either 16bit
.BR major " and " minor
numbers or a single 32bit value combining both.
.TP
.BI divisor " u32_value"
Specify a modulo value. Used when creating hash tables to define their size or
for declaring a
.B sample
to calculate hash table keys from. Must be a power of two with exponent not
exceeding eight.
.TP
.BI order " u32_value"
A value to order filters by, ascending. Conflicts with
.B handle
which serves the same purpose.
.TP
.BI sample " SELECTOR"
Used together with
.B ht
to specify which bucket to add this filter to. This allows one to avoid having
to know how exactly the kernel calculates hashes. The additional
.B divisor
defaults to 256, so must be given for hash tables of different size.
.TP
.BI link " HANDLE"
Delegate matching packets to filters in a hash table.
.I HANDLE
is used to only specify the hash table, so only
.BR htid " may be given, " hash " and " nodeid
have to be omitted. By default, bucket number 0 will be used and can be
overridden by the
.B hashkey
option.
.TP
.BI indev " ifname"
Filter on the incoming interface of the packet. Obviously works only for
forwarded traffic.
.TP
.BI skip_sw
Do not process filter by software. If hardware has no offload support for this
filter, or TC offload is not enabled for the interface, operation will fail.
.TP
.BI skip_hw
Do not process filter by hardware.
.TP
.BI help
Print a brief help text about possible options.
.SH SELECTORS
Basically the only real selector is
.B u32 .
All others merely provide a higher level syntax and are internally translated
into
.B u32 .
.TP
.BI u32 " VAL_MASK_32"
.TQ
.BI u16 " VAL_MASK_16"
.TQ
.BI u8 " VAL_MASK_8"
Match packet data to a given value. The selector name defines the sample length
to extract (32bits for
.BR u32 ,
16bits for
.B u16
and 8bits for
.BR u8 ).
Before comparing, the sample is binary AND'ed with the given mask. This way
uninteresting bits can be cleared before comparison. The position of the sample
is defined by the offset specified in
.IR AT .
.TP
.BI ip " IP"
.TQ
.BI ip6 " IP6"
Assume packet starts with an IPv4 (
.BR ip )
or IPv6 (
.BR ip6 )
header.
.IR IP / IP6
then allows to match various header fields:
.RS
.TP
.BI src " ADDR"
.TQ
.BI dst " ADDR"
Compare Source or Destination Address fields against the value of
.IR ADDR .
The reserved words
.BR default ", " any " and " all
effectively match any address. Otherwise an IP address of the particular
protocol is expected, optionally suffixed by a prefix length to match whole
subnets. In case of IPv4 a netmask may also be given.
.TP
.BI dsfield " VAL_MASK_8"
IPv4 only. Match the packet header's DSCP/ECN field. Synonyms to this are
.BR tos " and " precedence .
.TP
.BI ihl " VAL_MASK_8"
IPv4 only. Match the Internet Header Length field. Note that the value's unit is
32bits, so to match a packet with 24byte header length
.I u8_value
has to be 6.
.TP
.BI protocol " VAL_MASK_8"
Match the Protocol (IPv4) or Next Header (IPv6) field value, e.g. 6 for TCP.
.TP
.BI icmp_type " VAL_MASK_8"
.TQ
.BI icmp_code " VAL_MASK_8"
Assume a next-header protocol of icmp or ipv6-icmp and match Type or Code
field values. This is dangerous, as the code assumes minimal header size for
IPv4 and lack of extension headers for IPv6.
.TP
.BI sport " VAL_MASK_16"
.TQ
.BI dport " VAL_MASK_16"
Match layer four source or destination ports. This is dangerous as well, as it
assumes a suitable layer four protocol is present (which has Source and
Destination Port fields right at the start of the header and 16bit in size).
Also minimal header size for IPv4 and lack of IPv6 extension headers is assumed.
.TP
.B nofrag
.TQ
.B firstfrag
.TQ
.B df
.TQ
.B mf
IPv4 only, check certain flags and fragment offset values. Match if the packet
is not a fragment
.RB ( nofrag ),
the first fragment
.RB ( firstfrag ),
if Don't Fragment
.RB ( df )
or More Fragments
.RB ( mf )
bits are set.
.TP
.BI priority " VAL_MASK_8"
IPv6 only. Match the header's Traffic Class field, which has the same purpose
and semantics of IPv4's ToS field since RFC 3168: upper six bits are DSCP, the
lower two ECN.
.TP
.BI flowlabel " VAL_MASK_32"
IPv6 only. Match the Flow Label field's value. Note that Flow Label itself is
only 20bytes long, which are the least significant ones here. The remaining
upper 12bytes match Version and Traffic Class fields.
.RE
.TP
.BI tcp " TCPUDP"
.TQ
.BI udp " TCPUDP"
Match fields of next header of protocol TCP or UDP. The possible values for
.I TCPDUP
are:
.RS
.TP
.BI src " VAL_MASK_16"
Match on Source Port field value.
.TP
.BI dst " VALMASK_16"
Match on Destination Port field value.
.RE
.TP
.BI icmp " ICMP"
Match fields of next header of protocol ICMP. The possible values for
.I ICMP
are:
.RS
.TP
.BI type " VAL_MASK_8"
Match on ICMP Type field.
.TP
.BI code " VAL_MASK_8"
Match on ICMP Code field.
.RE
.TP
.BI mark " VAL_MASK_32"
Match on netfilter fwmark value.
.TP
.BI ether " ETHER"
Match on ethernet header fields. Possible values for
.I ETHER
are:
.RS
.TP
.BI src " ether_address" " " AT
.TQ
.BI dst " ether_address" " " AT
Match on source or destination ethernet address. This is dangerous: It assumes
an ethernet header is present at the start of the packet. This will probably
lead to unexpected things if used with layer three interfaces like e.g. tun or
ppp.
.SH EXAMPLES
.RS
.EX
tc filter add dev eth0 parent 999:0 prio 99 protocol ip u32 \\
        match ip src 192.168.8.0/24 classid 1:1
.EE
.RE

This attaches a filter to the qdisc identified by
.BR 999:0.
It's priority is
.BR 99 ,
which affects in which order multiple filters attached to the same
.B parent
are consulted (the lower the earlier). The filter handles packets of
.B protocol
type
.BR ip ,
and
.BR match es
if the IP header's source address is within the
.B 192.168.8.0/24
subnet. Matching packets are classified into class
.BR 1.1 .
The effect of this command might be surprising at first glance:

.RS
.EX
filter parent 1: protocol ip pref 99 u32
filter parent 1: protocol ip pref 99 u32 \\
        fh 800: ht divisor 1
filter parent 1: protocol ip pref 99 u32 \\
        fh 800::800 order 2048 key ht 800 bkt 0 flowid 1:1 \\
        match c0a80800/ffffff00 at 12
.EE
.RE

So parent
.B 1:
is assigned a new
.B u32
filter, which contains a hash table of size 1 (as the
.B divisor
indicates). The table ID is
.BR 800 .
The third line then shows the actual filter which was added above: it sits in
table
.B 800
and bucket
.BR 0 ,
classifies packets into class ID
.B 1:1
and matches the upper three bytes of the four byte value at offset
.B 12
to be
.BR 0xc0a808 ,
which is 192, 168 and 8.

Now for something more complicated, namely creating a custom hash table:

.RS
.EX
tc filter add dev eth0 prio 99 handle 1: u32 divisor 256
.EE
.RE

This creates a table of size 256 with handle
.B 1:
in priority
.BR 99 .
The effect is as follows:

.RS
.EX
filter parent 1: protocol all pref 99 u32
filter parent 1: protocol all pref 99 u32 fh 1: ht divisor 256
filter parent 1: protocol all pref 99 u32 fh 800: ht divisor 1
.EE
.RE

So along with the requested hash table (handle
.BR 1: ),
the kernel has created his own table of size 1 to hold other filters of the same
priority.

The next step is to create a filter which links to the created hash table:

.RS
.EX
tc filter add dev eth0 parent 1: prio 1 u32 \\
        link 1: hashkey mask 0x0000ff00 at 12 \\
        match ip src 192.168.0.0/16
.EE
.RE

The filter is given a lower priority than the hash table itself so
.B u32
consults it before manually traversing the hash table. The options
.BR link " and " hashkey
determine which table and bucket to redirect to. In this case the hash key
should be constructed out of the second byte at offset 12, which corresponds to
an IP packet's third byte of the source address field. Along with the
.B match
statement, this effectively maps all class C networks below 192.168.0.0/16 to
different buckets of the hash table.

Filters for certain subnets can be created like so:

.RS
.EX
tc filter add dev eth0 parent 1: prio 99 u32 \\
        ht 1: sample u32 0x00000800 0x0000ff00 at 12 \\
        match ip src 192.168.8.0/24 classid 1:1
.EE
.RE

The bucket is defined using the
.B sample
option: In this case, the second byte at offset 12 must be 0x08, exactly. In
this case, the resulting bucket ID is obviously 8, but as soon as
.B sample
selects an amount of data which could exceed the
.BR divisor ,
one would have to know the kernel-internal algorithm to deduce the destination
bucket. This filter's
.B match
statement is redundant in this case, as the entropy for the hash key does not
exceed the table size and therefore no collisions can occur. Otherwise it's
necessary to prevent matching unwanted packets.

Matching upper layer fields is problematic since IPv4 header length is variable
and IPv6 supports extension headers which affect upper layer header offset. To
overcome this, there is the possibility to specify
.B nexthdr+
when giving an offset, and to make things easier there are the
.BR tcp " and " udp
matches which use
.B nexthdr+
implicitly. This offset has to be calculated in beforehand though, and the only
way to achieve that is by doing it in a separate filter which then links to the
filter which wants to use it. Here is an example of doing so:

.RS
.EX
tc filter add dev eth0 parent 1:0 protocol ip handle 1: \\
        u32 divisor 1
tc filter add dev eth0 parent 1:0 protocol ip \\
        u32 ht 1: \\
        match tcp src 22 FFFF \\
        classid 1:2
tc filter add dev eth0 parent 1:0 protocol ip \\
        u32 ht 800: \\
        match ip protocol 6 FF \\
        match ip firstfrag \\
        offset at 0 mask 0f00 shift 6 \\
        link 1:
.EE
.RE

This is what is being done: In the first call, a single element sized hash table
is created so there is a place to hold the linked to filter and a known handle
.RB ( 1: )
to reference to it. The second call then adds the actual filter, which pushes
packets with TCP source port 22 into class
.BR 1:2 .
Using
.BR ht ,
it is moved into the hash table created by the first call. The third call then
does the actual magic: It matches IPv4 packets with next layer protocol 6 (TCP),
only if it's the first fragment (usually TCP sets DF bit, but if it doesn't and
the packet is fragmented, only the first one contains the TCP header), and then
sets the offset based on the IP header's IHL field (right-shifting by 6
eliminates the offset of the field and at the same time converts the value into
byte unit). Finally, using
.BR link ,
the hash table from first call is referenced which holds the filter from second
call.
.SH SEE ALSO
.BR tc (8),
.br
.BR cls_u32.txt " at " http://linux-tc-notes.sourceforge.net/