QoS in Linux with TC and Filters

TC, the Traffic Control utility, has been there for a very long time - forever
in my humble perception. It is still (and has ever been if I'm not mistaken) the
only tool to configure QoS in Linux.
Standard practice when transmitting packets over a medium which may block (due
to congestion, e.g.) is to use a queue which temporarily holds these packets. In
Linux, this queueing approach is where QoS happens: A Queueing Discipline
(qdisc) holds multiple packet queues with different priorities for dequeueing to
the network driver. The classification (i.e. deciding which queue a packet
should go into) is typically done based on Type Of Service (IPv4) or Traffic
Class (IPv6) header fields but depending on qdisc implementation, might be
controlled by the user as well.
Qdiscs come in two flavors, classful or classless. While classless qdiscs are
not as flexible as classful ones, they also require much less customizing. Often
it is enough to just attach them to an interface, without exact knowledge of
what is done internally. Classful qdiscs are the exact opposite: flexible in
application, they are often not even usable without insightful configuration.
As the name implies, classful qdiscs provide configurable classes to sort
traffic into. In it's basic form, this is not much different than, say, the
classless pfifo_fast which holds three queues and classifies per packet upon
priority field. Though typically classes go beyond that by supporting nesting
and additional characteristics like e.g. maximum traffic rate or quantum.
When it comes to controlling the classification process, filters come into play.
They attach to the parent of a set of classes (i.e. either the qdisc itself or
a parent class) and specify how a packet (or it's associated flow) has to look
like in order to suit a given class. To overcome this simplification, it is
possible to attach multiple filters to the same parent, which then consults each
of them in row until the first one accepts the packet.
Before getting into detail about what filters there are and how to use them, a
simple setup of a qdisc with classes is necessary:
  |                                                       |
  |  HTB                                                  |
  |                                                       |
  | .----------------------------------------------------.|
  | |                                                    ||
  | |  Class 1:1                                         ||
  | |                                                    ||
  | | .---------------..---------------..---------------.||
  | | |               ||               ||               |||
  | | |  Class 1:10   ||  Class 1:20   ||  Class 1:30   |||
  | | |               ||               ||               |||
  | | | .------------.|| .------------.|| .------------.|||
  | | | |            ||| |            ||| |            ||||
  | | | |  fq_codel  ||| |  fq_codel  ||| |  fq_codel  ||||
  | | | |            ||| |            ||| |            ||||
  | | | '------------'|| '------------'|| '------------'|||
  | | '---------------''---------------''---------------'||
  | '----------------------------------------------------'|
The following commands establish the basic setup shown:
(1) # tc qdisc replace dev eth0 root handle 1: htb default 30
(2) # tc class add dev eth0 parent 1: classid 1:1 htb rate 95mbit
(3) # alias tclass='tc class add dev eth0 parent 1:1'
(4) # tclass classid 1:10 htb rate 1mbit ceil 20mbit prio 1
(4) # tclass classid 1:20 htb rate 90mbit ceil 95mbit prio 2
(4) # tclass classid 1:30 htb rate 1mbit ceil 95mbit prio 3
(5) # tc qdisc add dev eth0 parent 1:10 fq_codel
(5) # tc qdisc add dev eth0 parent 1:20 fq_codel
(5) # tc qdisc add dev eth0 parent 1:30 fq_codel
A little explanation for the unfamiliar reader:
1) Replace the root qdisc of eth0 by an instance of HTB. Specifying the handle
   is necessary so it can be referenced in consecutive calls to `tc`. The
   default class for unclassified traffic is set to 30.
2) Create a single top-level class with handle 1:1 which limits the total
   bandwidth allowed to 95mbit/s. It is assumed that eth0 is a 100mbit/s link,
   staying a little below that helps to keep the main point of enqueueing in
   the qdisc layer instead of the interface hardware queue or at another
   bottleneck in the network.
3) Define an alias for the common part of the remaining three calls in order
   to improve readability. This means all remaining classes are attached to the
   common parent class from (2).
4) Create three child classes for different uses: Class 1:10 has highest
   priority but is tightly limited in bandwidth - fine for interactive
   connections.  Class 1:20 has mid priority and high guaranteed bandwidth, for
   high priority bulk traffic. Finally, there's the default class 1:30 with
   lowest priority, low guaranteed bandwidth and the ability to use the full
   link in case it's unused otherwise. This should be fine for uninteresting
   traffic not explicitly taken care of.
5) Attach a leaf qdisc to each of the child classes created in (4). Since
   HTB by default attaches `pfifo` as leaf qdisc, this step is optional. Still,
   the fairness between different flows provided by the classless `fq_codel` is
   worth the effort.
More information about the qdiscs and fine-tuning parameters can be found in
tc-htb(8) and tc-fq_codel(8).
Without any additional setup done, now all traffic leaving eth0 is shaped to
95mbit/s and directed through class 1:30. This can be verified by looking at the
`Sent` field of the class statistics printed via `tc -s class show dev eth0`:
Only the root class 1:1 and it's child 1:30 should show any traffic.
Finally time to start filtering!
Let's begin with a simple one, i.e. reestablishing what pfifo_fast did
automatically based on TOS/Priority field. Linux internally translates the
header field into the priority field of struct skbuff, which pfifo_fast uses for
classification. tc-prio(8) contains a table listing the priority (and
ultimately, pfifo_fast queue index) each TOS value is being translated into.
Here is a shorter version:
 TOS Values      Linux Priority (Number)         Queue Index
 0x0  - 0x6      Best Effort (0)                 1
 0x8  - 0xe      Bulk (2)                        2
 0x10 - 0x16     Interactive (6)                 0
 0x18 - 0x1e     Interactive Bulk (4)            1
Using the `basic` filter, it is possible to match packets based on that skbuff
field, which has the added benefit of being IP version agnostic. Since the HTB
setup above defaults to class ID 1:30, the Bulk priority can be ignored. The
basic filter allows to combine matches, therefore we get along with only two
# tc filter add dev eth0 parent 1: basic 
        match 'meta(priority eq 6)' classid 1:10
# tc filter add dev eth0 parent 1: basic 
        match 'meta(priority eq 0)' 
        or 'meta(priority eq 4)' classid 1:20
A detailed description of the basic filter and the ematch syntax it uses can be
found in tc-basic(8) and tc-ematch(8).
Obviously, this first example cries for optimization. A simple one would be to
just change the default class from 1:30 to 1:20, so filters are only needed for
Bulk and Interactive priorities:
# tc filter add dev eth0 parent 1: basic 
        match 'meta(priority eq 6)' classid 1:10
# tc filter add dev eth0 parent 1: basic 
        match 'meta(priority eq 2)' classid 1:20
Given that class IDs are random, choosing them wisely allows for a direct
mapping. So first, recreate the qdisc and classes configuration:
# tc qdisc replace dev eth0 root handle 1: htb default 10
# tc class add dev eth0 parent 1: classid 1:1 htb rate 95mbit
# alias tclass='tc class add dev eth0 parent 1:1'
# tclass classid 1:16 htb rate 1mbit ceil 20mbit prio 1
# tclass classid 1:10 htb rate 90mbit ceil 95mbit prio 2
# tclass classid 1:12 htb rate 1mbit ceil 95mbit prio 3
# tc qdisc add dev eth0 parent 1:16 fq_codel
# tc qdisc add dev eth0 parent 1:10 fq_codel
# tc qdisc add dev eth0 parent 1:12 fq_codel
This is basically identical to above, but with changed leaf class IDs and the
second priority class being the default. Using the flow filter with it's `map`
functionality, a single filter command is enough:
# tc filter add dev eth0 parent 1: handle 0x1337 flow 
        map key priority baseclass 1:10
The flow filter now uses the priority value to construct a destination class ID
by adding it to the value of `baseclass`. While this works for priority values of
0, 2 and 6, it will result in non-existent class ID 1:14 for Interactive Bulk
traffic. In that case, the HTB default applies so that traffic goes into class
ID 1:10 just as intended. Please note that specifying a handle is a mandatory
requirement by the flow filter, although I didn't see where one would use that
later. For more information about `flow`, see tc-flow(8).
While flow and basic filters are relatively easy to apply and understand, they
are as well quite limited to their intended purpose. A more flexible option is
the u32 filter, which allows to match on arbitrary parts of the packet data -
yet only on that, not any meta data associated to it by the kernel (with the
exception of firewall mark value). So in order to continue this little
exercise with u32, we have to base classification directly upon the actual TOS
value. An intuitive attempt might look like this:
# alias tcfilter='tc filter add dev eth0 parent 1:'
# tcfilter u32 match ip dsfield 0x10 0x1e classid 1:16
# tcfilter u32 match ip dsfield 0x12 0x1e classid 1:16
# tcfilter u32 match ip dsfield 0x14 0x1e classid 1:16
# tcfilter u32 match ip dsfield 0x16 0x1e classid 1:16
# tcfilter u32 match ip dsfield 0x8 0x1e classid 1:12
# tcfilter u32 match ip dsfield 0xa 0x1e classid 1:12
# tcfilter u32 match ip dsfield 0xc 0x1e classid 1:12
# tcfilter u32 match ip dsfield 0xe 0x1e classid 1:12
The obvious drawback here is the amount of filters needed. And without the
default class, eight more filters would be necessary. This also has performance
implications: A packet with TOS value 0xe will be checked eight times in total
in order to determine it's destination class. While there's not much to be done
about the number of filters, at least the performance problem can be eliminated
by using u32's hash table support:
# tc filter add dev eth0 parent 1: prio 99 handle 1: u32 divisor 16
This creates a hash table with 16 buckets. The table size is arbitrary, but not
random: Since the first bit of the TOS field is not interesting, it can be
ignored and therefore the range of values to consider is just [0;15], i.e. a
number of 16 different values. The next step is to populate the hash table:
# alias tcfilter='tc filter add dev eth0 parent 1: prio 99 u32 match u8 0 0'
# tcfilter ht 1:0: classid 1:16
# tcfilter ht 1:1: classid 1:16
# tcfilter ht 1:2: classid 1:16
# tcfilter ht 1:3: classid 1:16
# tcfilter ht 1:4: classid 1:12
# tcfilter ht 1:5: classid 1:12
# tcfilter ht 1:6: classid 1:12
# tcfilter ht 1:7: classid 1:12
# tcfilter ht 1:8: classid 1:16
# tcfilter ht 1:9: classid 1:16
# tcfilter ht 1:a: classid 1:16
# tcfilter ht 1:b: classid 1:16
# tcfilter ht 1:c: classid 1:10
# tcfilter ht 1:d: classid 1:10
# tcfilter ht 1:e: classid 1:10
# tcfilter ht 1:f: classid 1:10
The parameter `ht` denotes the hash table and bucket the filter should be added
to. Since the first TOS bit is ignored, it's value has to be divided by two in
order to get to the bucket it maps to. E.g. a TOS value of 0x10 will therefore
map to bucket 0x8.  For the sake of completeness, all possible values are mapped
and therefore a configurable default class is not required. Note that the used
match expression is not necessary, but mandatory. Therefore anything that
matches any packet will suffice. Finally, a filter which links to the defined
hash table is needed:
# tc filter add dev eth0 parent 1: prio 1 protocol ip u32 
        link 1: hashkey mask 0x001e0000 match u8 0 0
Here again, the actual match statement is not necessary, but syntactically
required. All the magic lies within the `hashkey` parameter, which defines which
part of the packet should be used directly as hash key. Here's a drawing of the
first four bytes of the IPv4 header, with the area selected by `hashkey mask`
 0                1                2                3
 |        |       | ########  |    |                               |
 | Version|  IHL  | #DSCP###  | ECN|  Total Length                 |
 |        |       | ########  |    |                               |
This may look confusing at first, but keep in mind that bit- as well as
byte-ordering here is LSB while the mask value is written in MSB we humans use.
Therefore reading the mask is done like so, starting from left:
1) Skip the first byte (which contains Version and IHL fields).
2) Skip the lowest bit of the second byte (0x1e is even).
3) Mark the four following bits (0x1e is 11110 in binary).
4) Skip the remaining three bits of the second byte as well as the remaining two
Before doing the lookup, the kernel right-shifts the masked value by the amount
of zero-bits in `mask`, which implicitly also does the division by two which the
hash table depends on. With this setup, every packet has to pass exactly two
filters to be classified. Note that this filter is limited to IPv4 packets: Due
to the related Traffic Class field being at a different offset in the packet, it
would not work for IPv6. To use the same setup for IPv6 as well, a second
entry-level filter is necessary:
# tc filter add dev eth0 parent 1: prio 2 protocol ipv6 u32 
        link 1: hashkey mask 0x01e00000 match u8 0 0
For illustration purposes, here again is a drawing of the first four bytes of
the IPv6 header, again with masked area highlighted:
 0                1                2                3
 |        | ########      |                                        |
 | Version| #Traffic Class|   Flow Label                           |
 |        | ########      |                                        |
Reading the mask value is analogous to IPv4 with the added complexity that
Traffic Class spans over two bytes. Yet, for comparison there's a simple trick:
IPv6 has the interesting field shifted by four bits to the left, and the new
mask's value is shifted by the same amount. For further information about `u32`
and what can be done with it, consult it's man page tc-u32(8).
Of course, the kernel provides many more filters than just `basic`, `flow` and
`u32` which have been presented above. As of now, the remaining ones are:
        Filtering using Berkeley Packet Filter programs. The program's return
        code determines the packet's destination class ID.
        Filter packets based on control groups. This is only useful for packets
        originating from the local host, as control groups only exist in that
        An extended variant of the flow filter.
        Matches on firewall mark values previously assigned to the packet by
        netfilter (or a filter action, see below for details). This allows to
        export the classification algorithm into netfilter, which is very
        convenient if appropriate rules exist on the same system in there
        Filter packets based on matching routing table entry. Basically
        equivalent to the `fw` filter above, to make use of an already existing
        extensive routing table setup.
rsvp, rsvp6
        Implementation of the Resource Reservation Protocol in Linux, to react
        upon requests sent by an RSVP daemon.
        Match packets based on tcindex value, which is usually set by the dsmark
        qdisc. This is part of an approach to support Differentiated Services in
        Linux, which is another topic on it's own.
Filter Actions
The tc filter framework provides the infrastructure to another extensible set of
tools as well, namely tc actions. As the name suggests, they allow to do things
with packets (or associated data). (The list of) Actions are part of a given
filter. If it matches, each action it contains is executed in order before
returning the classification result. Since the action has direct access to the
latter, it is in theory possible for an action to react upon or even change the
filtering result - as long as the packet matched, of course. Yet none of the
currently in-tree actions make use of this.
The Generic Actions framework originally evolved out of the filters' ability to
police traffic to a given maximum bandwidth. One common use case for that is to
limit ingress traffic, dropping packets which exceed the threshold. A classic
setup example is like so:
# tc qdisc add dev eth0 handle ffff: ingress
# tc filter add dev eth0 parent ffff: u32 
        match u32 0 0
        police rate 1mbit burst 100k
The ingress qdisc is not a real one, but merely a point of reference for filters
to attach to which should get applied to incoming traffic. The u32 filter added
above matches on any packet and therefore limits the total incoming bandwidth to
1mbit/s, allowing bursts of up to 100kbytes. Using the new syntax, the filter
command changes slightly:
# tc filter add dev eth0 parent ffff: u32 
        match u32 0 0 
        action police rate 1mbit burst 100k
The important detail is that this syntax allows to define multiple actions.
E.g. for testing purposes, it is possible to redirect exceeding traffic to the
loopback interface instead of dropping it:
# tc filter add dev eth0 parent ffff: u32 
        match u32 0 0 
        action police rate 1mbit burst 100k conform-exceed pipe 
        action mirred egress redirect dev lo
The added parameter `conform-exceed pipe` tells the police action to allow for
further actions to handle the exceeding packet.
Apart from `police` and `mirred` actions, there are a few more. Here's a full
list of the currently implemented ones:
        Apply a Berkeley Packet Filter program to the packet.
        Set the packet's firewall mark to that of it's connection. This works by
        searching the conntrack table for a matching entry. If found, the mark
        is restored.
        Trigger recalculation of packet checksums. The supported protocols are:
        IPv4, ICMP, IGMP, TCP, UDP and UDPLite.
        Pass the packet to an iptables target. This allows to use iptables
        extensions directly instead of having to go the extra mile via setting
        an arbitrary firewall mark and matching on that from within netfilter.
        Mirror or redirect packets. This is often combined with the ifb pseudo
        device to share a common QoS setup between multiple interfaces or even
        ingress traffic.
        Perform stateless Native Address Translation. This is certainly not
        complete and therefore inferior to NAT using iptables: Although the
        kernel module decides between TCP, UDP and ICMP traffic, it does not
        handle typical problematic protocols such as active FTP or SIP.
        Generic packet editing. This allows to alter arbitrary bytes of the
        packet, either by specifying an offset into the packet or by naming a
        packet header and field name to change. Currently, the latter is
        implemented only for IPv4 yet.
        Apply a bandwidth rate limiting policy. Packets exceeding it are dropped
        by default, but may optionally be handled differently.
        This is rather an example than real action. All it does is print a
        user-defined string together with a packet counter. Useful maybe for
        debugging when filter statistics are not available or too complicated.
        Edit associated packet data, supports changing queue mapping, priority
        field and firewall mark value.
        Add/remove a VLAN header to/from the packet. This might serve as
        alternative to using 802.1Q pseudo-interfaces in combination with
        routing rules when e.g. packets for a given destination need to be
Intermediate Functional Block
The Intermediate Functional Block (ifb) pseudo network interface acts as a QoS
concentrator for multiple different sources of traffic. Packets from or to other
interfaces have to be redirected to it using the `mirred` action in order to be
handled, regularly routed traffic will be dropped. This way, a single stack of
qdiscs, classes and filters can be shared between multiple interfaces.
Here's a simple example to feed incoming traffic from multiple interfaces
through a Stochastic Fairness Queue (sfq):
(1) # modprobe ifb
(2) # ip link set ifb0 up
(3) # tc qdisc add dev ifb0 root sfq
The first step is to load the ifb kernel module (1). By default, this will
create two ifb devices: ifb0 and ifb1. After setting ifb0 up in (2), the root
qdisc is replaced by sfq in (3). Finally, one can start redirecting ingress
traffic to ifb0, e.g. from eth0:
# tc qdisc add dev eth0 handle ffff: ingress
# tc filter add dev eth0 parent ffff: u32 
        match u32 0 0 
        action mirred egress redirect dev ifb0
The same can be done for other interfaces, just replacing `eth0` in the two
commands above. One thing to keep in mind here is the asymmetrical routing this
creates within the host doing the QoS: Incoming packets enter the system via
ifb0, while corresponding replies leave directly via eth0. This can be observed
using tcpdump on ifb0, which shows the input part of the traffic only. What's
more confusing is that tcpdump on eth0 shows both incoming and outgoing traffic,
but the redirection is still effective - a simple prove is setting ifb0 down,
which will interrupt the communication. Obviously tcpdump catches the packets to
dump before they enter the ingress qdisc, which is why it sees them while the
kernel itself doesn't.
My personal impression is that although the `tc` utility is an absolute
necessity for anyone aiming at doing QoS in Linux professionally, there are way
too many loose ends and trip wires present in it's environment. Contributing to
this is the fact, that much of the non-essential functionality is redundantly
available in netfilter. Another problem which adds weight to the first one is a
general lack of documentation. Of course, there are many HOWTOs and guides in
the internet, but since it's often not clear how up to date these are, I prefer
the usual resources such as man or info pages. Surely nothing one couldn't fix
in hindsight, but quality certainly suffers if the original author of the code
does not or can not contribute to that.
All that being said, once the steep learning curve has been mastered, the
conglomerate of (classful) qdiscs, filters and actions provides a highly
sophisticated and flexible infrastructure to perform QoS, which plays nicely
along with routing and firewalling setups.
Further Reading
A good starting point for novice users and experienced ones diving into unknown
areas is the extensive HOWTO at http://lartc.org. The iproute2 package ships
some examples (usually in /usr/share/doc/, depending on distribution) as well as
man pages for `tc` in general, qdiscs and filters. The latter have been added
just recently though, so if your distribution does not ship iproute2 version
4.3.0 yet, these are not in there. Apart from that, the internet is a spring of
HOWTOs and scripts people wrote - though these should be taken with a grain of
salt: The complexity of the matter often leads to copying others' solutions
without much validation, which allows for less optimal or even obsolete
implementations to survive much longer than desired.