One document matched: draft-ietf-behave-turn-05.xml
<?xml version="1.0" encoding="US-ASCII"?>
<!DOCTYPE rfc SYSTEM "rfc2629.dtd">
<?rfc toc="yes"?>
<?rfc compact='yes'?>
<?rfc subcompact='no'?>
<?rfc symrefs="yes"?>
<rfc category="std" docName="draft-ietf-behave-turn-05" ipr="full3978">
<front>
<title abbrev="TURN">Traversal Using Relays around NAT (TURN): Relay
Extensions to Session Traversal Utilities for NAT (STUN)</title>
<author fullname="Jonathan Rosenberg" initials="J.R." surname="Rosenberg">
<organization abbrev="Cisco">Cisco Systems, Inc.</organization>
<address>
<postal>
<street></street>
<city>Edison</city>
<region>NJ</region>
<country>US</country>
</postal>
<email>jdrosen@cisco.com</email>
<uri>http://www.jdrosen.net</uri>
</address>
</author>
<author fullname="Rohan Mahy" initials="R.M." surname="Mahy">
<organization abbrev="Plantronics">Plantronics, Inc.</organization>
<address>
<email>rohan@ekabal.com</email>
</address>
</author>
<author fullname="Philip Matthews" initials="P.M." surname="Matthews">
<organization abbrev="Avaya">Avaya, Inc.</organization>
<address>
<postal>
<street>1135 Innovation Drive</street>
<city>Ottawa</city>
<region>Ontario</region>
<code>K2K 3G7</code>
<country>Canada</country>
</postal>
<phone>+1 613 592-4343 x223</phone>
<facsimile></facsimile>
<email>philip_matthews@magma.ca</email>
<uri></uri>
</address>
</author>
<author fullname="Dan Wing" initials="D.W." surname="Wing">
<organization abbrev="Cisco">Cisco Systems, Inc.</organization>
<address>
<postal>
<street>170 West Tasman Drive</street>
<city>San Jose</city>
<region>CA</region>
<code>95134</code>
<country>USA</country>
</postal>
<phone></phone>
<facsimile></facsimile>
<email>dwing@cisco.com</email>
<uri></uri>
</address>
</author>
<date day="15" month="November" year="2007" />
<area>Transport</area>
<workgroup>Behave</workgroup>
<keyword>NAT</keyword>
<keyword>TURN</keyword>
<keyword>STUN</keyword>
<abstract>
<t>This specification defines an extension of the Session Traversal
Utilities for NAT (STUN) Protocol for asking the STUN server to relay
packets towards a client. This extension, called Traversal Using Relays
around NAT (TURN), is useful for hosts behind address dependent NATs.
The extension purposefully restricts the ways in which the relayed
address can be used. In particular, it prevents users from running
general purpose servers on ports obtained from the TURN server.</t>
</abstract>
</front>
<middle>
<section title="Introduction">
<t>Session Traversal Utilities for NAT (STUN) <xref
target="I-D.ietf-behave-rfc3489bis"></xref> provides a suite of tools
for facilitating the traversal of NAT. Specifically, it defines the
Binding method, which is used by a client to determine its reflexive
transport address towards the STUN server. The reflexive transport
address can be used by the client for receiving packets from peers, but
only when the client is behind "good" NATs. In particular, if a client
is behind a NAT whose mapping behavior <xref target="RFC4787"></xref> is
address or address and port dependent (sometimes called "bad" NATs), the
reflexive transport address will not be usable for communicating with a
peer.</t>
<t>The only way to obtain a UDP transport address that can be used for
corresponding with a peer through such a NAT is to make use of a relay.
The relay sits on the public side of the NAT, and allocates transport
addresses to clients reaching it from behind the private side of the
NAT. These allocated transport addresses are from IP addresses belonging
to the relay. When the relay receives a packet on one of these allocated
addresses, the relay forwards it toward the client.</t>
<t>This specification defines an extension to STUN, called TURN, that
allows a client to request an address on the TURN server, so that the
TURN server acts as a relay. This extension defines a handful of new
STUN methods. The Allocate method is the most fundamental component of
this set of extensions. It is used to provide the client with a
transport address that is relayed through the TURN server. A transport
address which relays through an intermediary is called a relayed
transport address.</t>
<t>Though a relayed transport address is highly likely to work when
corresponding with a peer, it comes at high cost to the provider of the
relay service. As a consequence, relayed transport addresses should only
be used as a last resort. Protocols using relayed transport addresses
should make use of mechanisms to dynamically determine whether such an
address is actually needed. One such mechanism, defined for multimedia
session establishment protocols based on the offer/answer protocol in
<xref target="RFC3264">RFC 3264</xref>, is Interactive Connectivity
Establishment (ICE) <xref target="I-D.ietf-mmusic-ice"></xref>.</t>
<t>Though originally invented for Voice over IP applications, TURN is
designed to be a general-purpose relay mechanism for NAT traversal.</t>
</section>
<section title="Terminology">
<t>The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
"SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this
document are to be interpreted as described in <xref
target="RFC2119">RFC 2119</xref>.</t>
<t><list style="hanging">
<t hangText="Relayed Transport Address:">A transport address that
terminates on a server, and is forwarded towards the client. The
TURN Allocate request can be used to obtain a relayed transport
address, for example.</t>
<t hangText="TURN client:">A STUN client that implements this
specification. It obtains a relayed transport address that it
provides to a small number of peers (usually one).</t>
<t hangText="TURN server:">A STUN server that implements this
specification. It relays data between a TURN client and its
peer(s).</t>
<t hangText="Peer:">A node with which the TURN client wishes to
communicate. The TURN server relays traffic between the TURN client
and its peer(s).</t>
<t hangText="Allocation:">The IP address and port granted to a
client through an Allocate request, along with related state, such
as permissions and expiration timers.</t>
<t hangText="5-tuple:">A combination of the source IP address and
port, destination IP address and port, and transport protocol (UDP,
or TCP). It uniquely identifies a TCP connection or bi-directional
flow of UDP datagrams.</t>
<t hangText="Permission:">A record of an IP address and transport of
a peer that is permitted to send traffic to the TURN client. The
TURN server will only forward traffic to its client from remote
peers that match an existing permission.</t>
</list></t>
</section>
<section title="Overview of Operation">
<t>In a typical configuration, a TURN client is connected to a private
network and through one or more NATs to the public Internet. On the
public Internet is a TURN server. Elsewhere in the Internet are one or
more peers that the TURN client wishes to communicate with. This
specification defines a framing mechanism and several new STUN methods.
Together, these add the ability for a STUN server to act as a packet
relay.</t>
<t>The framing mechanism serves two purposes. First, it contains a
length field that allow TURN nodes to find the boundaries between chunks
of application data when the communication with the TURN server is over
a stream-based transport such as TCP. Second, it carries a channel
number. Channel zero is used for TURN control messages, while the other
channel numbers are used for application data traveling to or from
various peers. The channel number allows the client to know which peer
sent data to it, and to specify which peer is to be the recipient of
data. Application data flowing on any non-zero channel is
unencapsulated, meaning that the application data starts immediately
after the framing header. The framing header is just four bytes. This
allows TURN to operate with minimal overhead, which is important for the
real-time protocols it is designed to support. Application data can also
flow in encapsulated format, meaning that it is carried in certain TURN
messages on channel 0. Channel numbers are independent in each
direction: for example, channel 5 might indicate one peer in the client
to server direction, but a different peer in the server to client
direction.</t>
<t>When the client wants to obtain a relayed transport address, the
client first sends an Allocate request to the server, which the server
authenticates. The server generates an Allocate response with the
allocated address, port, and target transport. All other STUN messages
defined by this specification happen in the context of an
allocation.</t>
<t>A successful Allocate transaction just reserves a transport address
on the TURN server. Data does not flow through an allocated transport
address until the TURN client asks the TURN server to open a permission,
which is done with a Send Indication. While the client can request more
than one permission per allocation, it needs to request each permission
explicitly and one at a time. This insures that a client can't use a
TURN server to run a traditional server, and partially protects the
client from DoS attacks.</t>
<t>Once a permission is open, the client can then receive data flowing
back from its peer. Initially this data is encapsulated in a Data
Indication. Since multiple permissions can be open simultaneously, the
Data Indication contains the PEER-ADDRESS attribute so the TURN client
knows which peer sent the data, and a CHANNEL-NUMBER attribute so the
client knows how the server will refer to traffic from this peer when
sent unencapsulated. Likewise when the client initially sends to a new
peer, it uses a Send Indication with the peer address in the
PEER-ADDRESS attribute, along with a channel number so the server knows
how the client will refer to unencapsulated data to this peer.</t>
<figure anchor="fig-channels" title="Example Usage of Channels">
<artwork><![CDATA[
TURN TURN peer
client server
|--- Allocate Req -->| |
|<-- Allocate Resp ---| |
| | |
|--- Send (chan 2) -->| data |
| |============>|
|<-- ChannelConfirm --| |
| | data |
| |<============|
|<-- Data (chan 5) ---| |
|--- ChannelConfirm ->| |
| | |
|--- [2] + data ----->| data |
| |============>|
| | data |
| |<============|
|<-- [5] + data ------| |
]]></artwork>
</figure>
<t>When the client and server communicate over UDP, data and control
messages can arrive out of order. For this reason, the client needs to
verify the server knows the client channel mapping before the client
sends unencapsulated, and the server needs to verify the client knows
the server channel mapping before the server sends unencapsulated. When
the client and server communicate over UDP, a Channel Confirmation
indication is sent after the Send (or Data) indication so the client (or
server) knows that it can send unencapsulated.</t>
<t><xref target="fig-channels"></xref> demonstrates how this works. The
client performs an Allocate Request, and gets a response. It decides to
send data to a specific peer. Initially, it sends data to that peer
using a TURN Send indication on channel 0. That Send Indication tells
the TURN server that, once confirmed, the client will send data
unencapsulated to that peer on channel 2. Whenever the TURN server
receives a Send indication, it stores the mapping from channel number to
peer, and sends a ChannelConfirm indication (on channel 0). Once the
confirmation has been received by the client, the client can send data
to the peer on channel 2. Prior to receipt of the ChannelConfirm, any
other data the client wishes to send to the peer is sent using Send
indications, all of which indicate that channel 2 is to be used for
unencapsulated data. The same procedure happens from server to client;
the TURN server initially sends data using a Data indication on channel
0, and once confirmed with a ChannelConfirm, it can send it
unencapsulated on its selected channel (channel 5 in the example).</t>
<t>Over a reliable transport, such as TCP, the confirmation step is not
needed so the Channel Confirmation indication is not used. Clients can
immediately send the next piece of data to the peer on the requested
channel.</t>
<t>Allocations can also request specific attributes such as the desired
Lifetime of the allocation and the maximum Bandwidth. Clients can also
request specific port assignment behavior, for example, a specific port
number, odd or even port numbers, or pairs of sequential port
numbers.</t>
<section title="Transports">
<t>TURN clients can communicate with a TURN server using UDP, TCP, or
TLS over TCP. A TURN server can then relay traffic between a reliable
transport used between the client and server (TCP or TLS over TCP),
and UDP used from server to peer. When relaying data sent from a
stream-based protocol to a UDP peer, the TURN server emits datagrams
which are the same length as the length field in the TURN framing or
the length of the DATA attribute in a Send Indication. Likewise, when
a UDP datagram is received by the TURN server and relayed to the
client over a stream-based transport, the length of the datagram is
the length of the TURN framing or Data Indication's DATA
attribute.</t>
<t>The following table shows the possible combinations of transport
protocols from client to server and from server to peer:</t>
<texttable>
<ttcol>client to TURN server</ttcol>
<ttcol>TURN server to peer</ttcol>
<c>UDP</c>
<c>UDP</c>
<c>TCP</c>
<c>UDP</c>
<c>TLS</c>
<c>UDP</c>
</texttable>
<t>For TURN clients, using TLS over TCP provides two benefits. When
using TLS, the client can be assured that the address of the client's
peers are not visible to an attacker except by traffic analysis
downstream of the TURN server. Second, the client may be able to
communicate with TURN servers using TLS when it would not be able to
communicate with the same server using TCP or UDP, due to the
configuration of a firewall between the TURN client and its server.
TLS between the client and TURN server in this case just facilitates
traversal.</t>
<t>In addition, an extension to TURN is planned to add support for TCP
allocations <xref target="I-D.ietf-behave-turn-tcp"></xref>.</t>
</section>
<section title="About Tuples">
<t>To relay data to and from the correct location, the TURN server
maintains an association between the 5-tuple used to communicate with
the client and the 5-tuple used to communicate with each of the
client's peers. The 5-tuple on the client side will consist of the
client's reflexive address -- the apparent source address and port of
the client (typically as rewritten by the last NAT)--and the
destination address and port used by the TURN server. The <xref
target="fig:turn-model">figure below</xref> shows a typical topology.
In this diagram, the client 5-tuple is for a UDP flow between
192.0.2.1:7000 and 192.0.2.15:3490. The 5-tuple between the TURN
server and Peer B is for a UDP flow between 192.0.2.15:9000 (the TURN
allocated address) and 192.0.2.210:18200.</t>
<t><list>
<t>While the terminology used in this document refers to 5-tuples,
the TURN server can store whatever identifier it likes that yields
identical results. Specifically, many implementations may use a
file-descriptor in place of a 5-tuple to represent a TCP
connection.</t>
</list></t>
<figure anchor="fig:turn-model">
<artwork><![CDATA[
+---------+
| |
| |
/ | Peer A |
Client's TURN // | |
Host Server / | |
Address Address // +---------+
10.1.1.2:17240 192.0.2.15:3490 / 192.0.2.180:16400
| | //
| +-+ | /
| | | | /
v | | | // 192.0.2.210:18200
+---------+ | | |+---------+ / +---------+
| | |N| || | // | |
| TURN | | | v| TURN |/ | |
| Client |----|A|----------| Server |------------------| Peer B |
| | | |^ | |^ ^| |
| | |T|| | || || |
+---------+ | || +---------+| |+---------+
| || | |
| || | |
+-+| | |
| | |
|
Client's TURN Peer B
Reflexive Allocated Transport
Address Address Address
192.0.2.1:7000 192.0.2.15:9000 192.0.2.210:18200
]]></artwork>
</figure>
</section>
<section title="Keepalives">
<t>Since the main purpose of STUN and TURN is to traverse NATs, it is
natural to consider which elements are responsible for generating
sufficient periodic traffic to insure that NAT bindings stay alive.
TURN clients need to send data frequently enough to keep both NAT
bindings and the TURN server permissions fresh. Like NAT bindings, the
TURN server permissions are refreshed by ordinary data traffic relayed
from the client to the peer. Unlike permissions, allocations on the
TURN server have an explicit expiration time and need to be refreshed
explicitly by the client with a TURN Refresh request. When an
allocation expires, all permissions associated with that allocation
are automatically deleted.</t>
</section>
</section>
<!-- Overview -->
<section anchor="framing" title="TURN Framing Mechanism">
<t>All TURN control messages and all application data sent between the
client and the server MUST start with the TURN framing header. This
header is used for two purposes: indicating the channel number, and for
framing.</t>
<t>TURN uses a channel number to distinguish control traffic from data,
and to distinguish among multiple peers using the same allocation.
Channel number zero is reserved for TURN control messages. All TURN
requests, responses and indications between the client and server MUST
be sent on channel 0, and MUST NOT be sent on any other channel. Channel
0xFFFF is reserved for future use and MUST NOT be used by clients or
servers compliant to this specification. Other channel numbers are
assigned and communicated as described in <xref
target="sendrecv"></xref>. Because the framing is always used, TURN
needs to run on a separate port number from unframed STUN requests.</t>
<t>Over stream-based transports, the TURN client and server also need to
include an explicit length so that the TURN server can perform
conversion from streams to datagrams and vice versa. TURN framing has a
2 octet channel number and a 2 octet length field. Over stream-based
transports, the length field counts the number of octets immediately
after the length field itself. Over UDP the length is always set to
zero.</t>
<figure>
<artwork><![CDATA[
0 1 2 3
0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
| Channel Number | Length |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
]]></artwork>
</figure>
<t></t>
<t>Channel numbers are always defined within a particular allocation. If
a client has multiple allocations on a TURN server, there is no
relationship whatsoever between the channel numbers in each allocation.
Once created, a channel number persists for the lifetime of the
allocation. There is no way to explicitly remove a channel.
Consequently, a client which obtains an allocation with the intent of
holding it for extremely long periods, possibly for communication with
many different peers over time, may eventually exhaust the set of
channels. In that case, the client will need to obtain a new
allocation.</t>
</section>
<!-- Framing -->
<section title="General Behavior">
<t>After the initial Allocate transaction, all subsequent TURN
transactions need to be sent in the context of a valid allocation. The
source and destination IP address and ports for these TURN messages MUST
match the internal 5-tuple of an existing allocation. These are
processed using the general server procedures in <xref
target="I-D.ietf-behave-rfc3489bis"></xref> with a few important
additions. For requests (in this specification, the only subsequent
request possible is a Refresh request), if there is no matching
allocation, the server MUST generate a 437 (Allocation Mismatch) error
response. For indications, if there is no matching allocation, the
indication is silently discarded. An Allocate request MUST NOT be sent
by a client within the context of an existing allocation. Such a request
MUST be rejected by the server with a 437 (Allocation Mismatch) error
response.</t>
<t>A subsequent request MUST be authenticated using the same username
and realm as the one used in the Allocate request that created the
allocation. If the request was authenticated but not with the matching
credential, the server MUST reject the request with a 401 (Unauthorized)
error response.</t>
<t>When a server returns an error response, it MAY include an
ALTERNATE-SERVER attribute if it has positive knowledge that the problem
reported in the error response will not be a problem on the alternate
server. For example, a 443 response (Invalid IP Address) with an
ALTERNATE-SERVER means that the other server is responsible for that IP
address. A 442 (Unsupported Transport Protocol) with this attribute
means that the other server is known to support that transport protocol.
A 507 (Insufficient Capacity) means that the other server is known to
have sufficient capacity. Using the ALTERNATE-SERVER mechanism in the
507 (Insufficient Capacity) response can only be done if the rejecting
server has definitive knowledge of available capacity on the target.
This will require some kind of state sharing mechanism between TURN
servers, which is beyond the scope of this specification. If a TURN
server attempts to redirect to another server without knowledge of
available capacity, it is possible that all servers are in a congested
state, resulting in series of rejections that only serve to further
increase the load on the system. This can cause congestion collapse.</t>
<t>If a client sends a request to a server and gets a 500 class error
response without an ALTERNATE-SERVER, or the transaction times out
without a response, and the client was utilizing the SRV procedures of
<xref target="I-D.ietf-behave-rfc3489bis"></xref> to contact the server,
the client SHOULD try another server based on those procedures. However,
the client SHOULD cache the fact that the request to this server failed,
and not retry that server again for a configurable period of time. Five
minutes is RECOMMENDED.</t>
<t>TURN clients and servers MUST NOT include the FINGERPRINT attribute
in any of the methods defined in this document.</t>
</section>
<section anchor="allocations" title="Managing Allocations">
<t>Communications between a TURN client and a TURN server on a new flow
begin with an Allocate transaction. All subsequent transactions happen
in the context of that allocation. The client refreshes allocations and
deallocates them using a Refresh transaction.</t>
<section title="Client Behavior">
<section anchor="sec:alloc" title="Initial Allocate Requests">
<t>When a client wishes to obtain a transport address, it sends an
Allocate request to the server. This request is constructed and sent
using the general procedures defined in <xref
target="I-D.ietf-behave-rfc3489bis"></xref>. Clients MUST implement
the long term credential mechanism defined in <xref
target="I-D.ietf-behave-rfc3489bis"></xref>, and be prepared for the
server to use it.</t>
<t>The client SHOULD include a BANDWIDTH attribute, which indicates
the maximum bandwidth that will be used with this binding. If the
maximum is unknown, the attribute is not included in the
request.</t>
<t><list style="empty">
<t>OPEN ISSUE: Bandwidth is very much underspecified. Is anyone
actually using it for capacity planning? If not we should
remove.</t>
</list></t>
<t>The client MAY request a particular lifetime for the allocation
by including it in the LIFETIME attribute in the request.</t>
<t>The client MUST include a REQUESTED-TRANSPORT attribute. In this
specification, the REQUESTED-TRANSPORT will always be UDP. This
attribute is included to allow for future extensions to TURN.</t>
<t>The client MAY include a REQUESTED-PORT-PROPS or REQUESTED-IP
attribute in the request to obtain specific types of transport
addresses. Whether these are needed depends on the application using
the TURN server. As an example, the Real Time Transport Protocol
(RTP) <xref target="RFC3550"></xref> requires that RTP and RTCP
ports be an adjacent pair, even and odd respectively, for
compatibility with a previous version of that specification. The
REQUESTED-PORT-PROPS attribute allows the client to ask the relay
for those properties.</t>
<t>Processing of the response follows the general procedures of
<xref target="I-D.ietf-behave-rfc3489bis"></xref>. A successful
response will include both a RELAY-ADDRESS and an XOR-MAPPED-ADDRESS
attribute, providing both a relayed transport address and a
reflexive transport address, respectively, to the client. The value
of the LIFETIME attribute in the response indicates the amount of
time after which the server will expire the allocation, if not
refreshed with a Refresh request. The server will allow the user to
send and receive at least the amount of data indicated in the
BANDWIDTH attribute per allocation. (At its discretion the server
can optionally discard UDP data above this threshold.)</t>
<t>If the response is an error response and contains a 442, 443 or
444 error code, the client knows that its requested properties could
not be met. The client MAY retry with different properties, with the
same properties (in a hope that something has changed on the
server), or give up, depending on the needs of the application.
However, if the client retries, it SHOULD wait 500ms, and if the
request fails again, wait 1 second, then 2 seconds, and so on,
exponentially backing off.</t>
</section>
<section anchor="sec:refresh" title="Refresh Requests">
<t>Before 3/4 of the lifetime of the allocation has passed (the
lifetime of the allocation is conveyed in the LIFETIME attribute of
the Allocate Response), the client SHOULD refresh the allocation
with a Refresh transaction if it wishes to keep the allocation.</t>
<t>To perform a refresh, the client generates a Refresh Request. The
client MUST use the same username, realm and password for the
Refresh request as it used in its initial Allocate Request. The
Refresh request MAY contain a proposed LIFETIME attribute. The
client MAY include a BANDWIDTH attribute if it wishes to request
more or less bandwidth than in the original request. If absent, it
indicates no change in the requested bandwidth from the Allocate
request. The client MUST NOT include a REQUESTED-IP,
REQUESTED-TRANSPORT, or REQUESTED-PORT-PROPS attribute in the
Refresh request.</t>
<t>In a successful response, the LIFETIME attribute indicates the
amount of additional time (the number of seconds after the response
is received) that the allocation will live without being refreshed.
A successful response will also contain a BANDWIDTH attribute,
indicating the bandwidth the server is allowing for this allocation.
Note that an error response does not imply that the allocation has
expired, just that the refresh has failed.</t>
<t>If a client no longer needs an allocation, it SHOULD perform an
explicit deallocation. If the client wishes to explicitly remove the
allocation because it no longer needs it, it sends a Refresh
request, but sets the LIFETIME attribute to zero. This will cause
the server to remove the allocation, and all associated permissions
and channel numbers. For connection-oriented transports such as TCP,
the client can also remove the allocation (and all associated
bindings) by closing the relevant connection with the TURN
server.</t>
</section>
<!-- close client -->
</section>
<section title="Server Behavior">
<t>The server first processes the request according to the base
protocol procedures in <xref
target="I-D.ietf-behave-rfc3489bis"></xref>, extended with the
procedures for the long-term credential mechanism.</t>
<section anchor="sec:initial"
title="Initial Allocate Requests">
<t>When the server receives an Allocate request, the server attempts
to allocate a relayed transport address. It first looks for the
BANDWIDTH attribute in the request. If present, the server
determines whether or not it has sufficient capacity to handle a
binding that will generate the requested bandwidth.</t>
<t>If it does, the server attempts to allocate a transport address
for the client. The Allocate Request can contain several additional
attributes that allow the client to request specific characteristics
of the transport address.</t>
<section title="REQUESTED-TRANSPORT">
<t>First, the server checks for the REQUESTED-TRANSPORT attribute.
This indicates the transport protocol requested by the client.
This specification defines a value for UDP only, but support for
TCP allocations is planned in <xref
target="I-D.ietf-behave-turn-tcp"></xref>.</t>
<t><list>
<t>As a consequence of the REQUESTED-TRANSPORT attribute, it
is possible for a client to connect to the server over TCP or
TLS over TCP and request a UDP transport address. In this
case, the server will relay data between the transports.</t>
</list></t>
<t>If the requested transport is supported, the server allocates a
port using the requested transport protocol. If the
REQUESTED-TRANSPORT attribute contains a value of the transport
protocol unknown to the server, or known to the server but not
supported by the server in the context of this request, the server
MUST reject the request and include a 442 (Unsupported Transport
Protocol) in the response. If the request did not contain a
REQUESTED-TRANSPORT attribute, the server MUST use the same
transport protocol as the request arrived on.</t>
</section>
<section title="REQUESTED-IP">
<t>Next, the server checks for the REQUESTED-IP attribute. If
present, it indicates a specific IP address from which the client
would like its transport address allocated. (The client could do
this if it requesting the second address in a specific port pair).
If this IP address is not a valid one for allocations on the
server, the server MUST reject the request and include a 443
(Invalid IP Address) error code in the response, or else redirect
the request to a server that is known to support this IP address.
If the IP address is one that is valid for allocations
(presumably, the server is configured to know the set of IP
addresses from which it performs allocations), the server MUST
provide an allocation from that IP address. If the attribute is
not present, the selection of an IP address is at the discretion
of the server.</t>
</section>
<section title="REQUESTED-PORT-PROPS">
<t>Finally, the server checks for the REQUESTED-PORT-PROPS
attribute. If present, it indicates specific port properties
desired by the client. This attribute is split into two portions:
one portion for port behavior and the other for requested port
alignment (whether the allocated port is odd, even, reserved as a
pair, or at the discretion of the server).</t>
<t>If the port behavior requested is for a Specific Port, the
server MUST attempt to allocate that specific port for the client.
If the specific port is not available (in use or reserved), the
server MUST reject the request with a 444 (Invalid Port) response.
For example, the STUN server could reject a request for a Specific
Port because the port is temporarily reserved as part of an
adjacent pair of ports, or because the requested port is a
well-known port (1-1023).</t>
<t>If the client requests "even" port alignment, the server MUST
attempt to allocate an even port for the client. If an even port
cannot be obtained, the server MUST reject the request with a 444
(Invalid Port) response or redirect to an alternate server. If the
client requests odd port alignment, the server MUST attempt to
allocate an odd port for the client. If an odd port cannot be
obtained, the server MUST reject the request with a 444 (Invalid
Port) response or redirect to an alternate server. Finally, the
"Even port with hold of the next higher port" alignment is similar
to requesting an even port. It is a request for an even port, and
MUST be rejected by the server if an even port cannot be provided,
or redirected to an alternate server. However, it is also a hint
from the client that the client will request the next higher port
with a separate Allocate request. As such, it is a request for the
server to allocate an even port whose next higher port is also
available, and furthermore, a request for the server to not
allocate that one higher port to any other request except for one
that asks for that port explicitly. The server can honor this
request for adjacency at its discretion. The only constraint is
that the allocated port has to be even.</t>
<t><list>
<t>Port alignment requests exist for compatibility with
implementations of RTP which predate RFC 3550. These
implementations use the port numbering conventions in (now
obsolete) RFC 1889.</t>
</list></t>
</section>
<section title="Creating the Allocation">
<t>If any of the requested or desired constraints cannot be met,
whether it be bandwidth, transport protocol, IP address or port,
instead of rejecting the request, the server can alternately
redirect the client to a different server that may be able to
fulfill the request. This is accomplished using the 300 error
response and ALTERNATE-SERVER attribute. If the server does not
redirect and cannot service the request because the server has
reached capacity, it sends a 507 (Insufficient Capacity) response.
The server can also reject the request with a 486 (Allocation
Quota Reached) if the user or client is not authorized to request
additional allocations.</t>
<t>The server SHOULD only allocate ports in the range 1024-65535.
This is one of several ways to prohibit relayed transport
addresses from being used to attempt to run standard services.</t>
<t>Once a port is allocated, the server associates the allocation
with the 5-tuple used to communicate between the client and the
server. For TCP, this amounts to associating the TCP connection
from the TURN client with the allocated transport address.</t>
<t>The new allocation MUST also be associated with the username,
password and realm used to authenticate the request. These
credentials are used in all subsequent requests to ensure that
only the same client can use or modify the allocation it was
given.</t>
<t>In addition, the allocation created by the server is associated
with a set of permissions. Each permission is a specific IP
address identifying an external client. Initially, this list is
null.</t>
<t>If the LIFETIME attribute was present in the request, and the
value is larger than the maximum duration the server is willing to
use for the lifetime of the allocation, the server MAY lower it to
that maximum. However, the server MUST NOT increase the duration
requested in the LIFETIME attribute. If there was no LIFETIME
attribute, the server may choose a duration at its discretion. Ten
minutes is RECOMMENDED. In either case, the resulting duration is
added to the current time, and a timer, called the allocation
expiration timer, is set to fire at or after that time. <xref
target="sec:lte"></xref> discusses behavior when the timer fires.
Note that the LIFETIME attribute an Allocate request can be zero,
though this is effectively a no-op, since it will create and
destroy the allocation in one transaction.</t>
</section>
<section title="Sending the Allocate Response">
<t>Once the port has been obtained and the allocation expiration
timer has been started, the server generates an Allocate Response
using the general procedures defined in <xref
target="I-D.ietf-behave-rfc3489bis"></xref>, including the ones
for long term authentication. The transport address allocated to
the client MUST be included in the RELAY-ADDRESS attribute in the
response. In addition, this response MUST contain the
XOR-MAPPED-ADDRESS attribute. This allows the client to determine
its reflexive transport address in addition to a relayed transport
address, from the same Allocate request.</t>
<t>The server MUST add a LIFETIME attribute to the Allocate
Response. This attribute contains the duration, in seconds, of the
allocation expiration timer associated with this allocation.</t>
<t>The server MUST add a BANDWIDTH attribute to the Allocate
Response. This MUST be equal to the attribute from the request, if
one was present. Otherwise, it indicates a per-allocation limit
that the server is placing on the bandwidth usage on each binding.
Such limits are needed to prevent against denial-of-service
attacks (See <xref target="sec:sec"></xref>).</t>
</section>
</section>
<section anchor="sec:subsequent" title="Refresh Requests">
<t>A Refresh request is processed using the general server and long
term authentication procedures in <xref
target="I-D.ietf-behave-rfc3489bis"></xref>. It is used to refresh
and extend an allocation, or to cause an immediate deallocation. It
is processed as follows.</t>
<t>First, the request MUST be authenticated using the same shared
secret as the one associated with the allocation. If the request was
authenticated but not with such a matching credential, the server
MUST generate a Refresh Error Response with a 401 response.</t>
<t>If the Refresh request contains a BANDWIDTH attribute, the server
checks that it can relay the requested volume of traffic.</t>
<t>Finally, a Refresh Request will set a new allocation expiration
timer for the allocation, effectively canceling the previous
allocation expiration timer. As with an Allocate request, the server
can offer a shorter allocation lifetime, but never a longer one.</t>
<t>A success Refresh response MUST contain a LIFETIME attribute and
a BANDWIDTH attribute.</t>
</section>
<!-- close server -->
</section>
<!-- close allocate behavior -->
</section>
<section anchor="sendrecv" title="Sending and Receiving Data">
<t>As described in <xref target="framing"></xref>, TURN allows a client
to send and receive data without utilizing TURN Send and Data
indications, by sending and receiving them on channels. Before sending
client-to-peer or peer-to-client data for a new peer, a TURN client or
server needs to assign a channel number that corresponds to that remote
peer. Once a channel number is assigned, it remains assigned through the
duration of the allocation. It cannot be unassigned or reassigned to a
different peer.</t>
<section title="Client Behavior">
<section title="Sending">
<t>When the client wants to forward data to a peer, it checks if it
has assigned a channel number for communications with this peer (as
identified by its IP address and port) over this allocation: <list
style="symbols">
<t>If one has not been assigned, the client assigns one of its
own choosing. This channel number MUST be one that is currently
unassigned by the client for this allocation. It MUST be between
1 and 65534. It is RECOMMENDED that the client choose one of the
unassigned numbers randomly, rather than sequentially. The state
of the channel is set to unconfirmed.</t>
<t>If one has been assigned, that channel MUST be selected.</t>
</list></t>
<t>Next, the client checks if the channel number has been confirmed
by the server. If the channel number has been confirmed, the client
simply sends the data to the TURN server with the appropriate
channel number in the TURN framing.</t>
<t>If the channel number has not been confirmed, the client creates
a Send indication. It places the selected channel number in a
CHANNEL-NUMBER attribute, the peer IP address and port in a
PEER-ADDRESS attribute, and puts the data to be sent in a DATA
attribute. (If the client just wishes to create a permission, it can
omit the DATA attribute.) If the Send indication is sent over a
reliable transport (ex: TCP), the client marks that the channel
number as confirmed. When the client receives a ChannelConfirmation
Indication, and the channel number, IP address and port match the
channel number assigned to that peer, the client marks that the
channel number is confirmed.</t>
<t>Since Send is an Indication, it generates no response. The client
must rely on application layer mechanisms to determine if the data
was received by the peer. A ChannelConfirmation Indication just
means that some Send indication was received by the TURN server. It
does not mean that a specific Send indication was received by the
peer.</t>
<t><list>
<t>Note that Send Indications are not authenticated and do not
contain a MESSAGE-INTEGRITY attribute. Just like non-relayed
data sent over UDP or TCP, the authenticity and integrity of
this data can only be assured using security mechanisms at
higher layers.</t>
</list></t>
</section>
<section title="Receiving">
<t>When the client receives a Data indication, it: <list
style="symbols">
<t>records the channel number used by the server (from the
CHANNEL-NUMBER attribute) and associates it with the IP address
and port in the PEER-ADDRESS attribute, which identify the peer
that sent the data. The resulting mapping from channel number to
transport address MUST be stored by the client for the duration
of the allocation.</t>
<t>delivers the contents of the DATA attribute to the client
application as if it was received from the peer's IP address and
port.</t>
<t>If the Data indication was received over UDP, the client MUST
confirm the channel used by the server, by sending a
ChannelConfirmation Indication to the server. This indication
MUST contain the same PEER-ADDRESS and CHANNEL-NUMBER attributes
included in the Data indication. This indication is sent to the
server on channel 0 using the 5-tuple associated with this
allocation. Note that, due to round trip delays, a client may
receive several Data indications with the same channel number
for the same remote peer. It MUST process each as defined here,
resulting in several ChannelConfirmation indications.</t>
</list></t>
<t>When the client receives unencapsulated data, it checks the
received channel number. If the client has a mapping associated with
the server channel number it delivers the data to the client
application as if it was received directly from that peer.
Otherwise, it silently discards the data.</t>
</section>
</section>
<section title="Server Behavior">
<section title="Receiving Data from the Client">
<t>When the server receives a Data indication from the client, it:
<list style="symbols">
<t>records the channel number used by the client (from the
CHANNEL-NUMBER attribute) and associates it with the IP address
and port in the PEER-ADDRESS attribute, which identify the peer
to which the data is to be sent. The resulting mapping from
channel number to peer transport address MUST be stored by the
server for the duration of the allocation.</t>
<t>sends the contents of the DATA attribute in a UDP datagram,
sending it to the PEER-ADDRESS and sending from the allocated
transport address.</t>
<t>if one doesn't exist, creates a permission for the IP address
from the PEER-ADDRESS (the port is ignored), and attaches the
permission to the allocation</t>
<t>checks if a timer has been set for this permission. If none
has been started, the server starts one. It is RECOMMENDED that
it have a value of sixty seconds. If the timer is already
running, it MUST be reset.</t>
<t>If the Send indication was received over UDP, the server MUST
confirm the channel used by the client, by sending a
ChannelConfirmation Indication to the client. This indication
MUST contain the same PEER-ADDRESS and CHANNEL-NUMBER attributes
included in the Send indication. This indication is sent to the
client on channel 0 using the 5-tuple associated with this
allocation. Note that, due to round trip delays, a server may
receive several Send indications with the same channel number
for the same remote peer. It MUST process each as defined here,
resulting in several ChannelConfirmation indications.</t>
</list></t>
<t>When the server receives unencapsulated data, it checks the
received channel number: <list style="symbols">
<t>If the server has a mapping associated with the client
channel number it: <list style="symbols">
<t>sends a UDP datagram to the peer using the transport
address from the mapping, and sends from the allocated
transport address.</t>
<t>checks if a permission activity timer is running for the
destination IP address of the peer. If one is not running,
the server starts one. It is RECOMMENDED that it have a
value of sixty seconds. If the timer is already running, it
MUST be reset.</t>
</list></t>
<t>If the server has no mapping, it silently discards the
data.</t>
</list></t>
</section>
<section anchor="sec:recv-alloc" title="Receiving Data from Peers">
<t>If a server receives a UDP packet on an allocated UDP transport
address, it checks the permissions associated with that allocation.
If the source IP address of the UDP packet matches one of the
permissions (the source port is not used), the UDP packet is
accepted. Otherwise, it is discarded. If the packet is accepted, it
is forwarded to the client as described below.</t>
<t>The server checks if it has assigned a channel number for
communications from this peer (as identified by its IP address and
port) over this allocation: <list style="symbols">
<t>If one has not been assigned, the client assigns one of its
own choosing. This channel number MUST be one that is currently
unassigned by the server for this allocation. It MUST be between
1 and 65534. It is RECOMMENDED that the server choose one of the
unassigned numbers randomly, rather than sequentially. The state
of the channel is set to unconfirmed.</t>
<t>If one has been assigned, that channel MUST be selected.</t>
</list></t>
<t>Note that data from peers does not reset the permission activity
timer.</t>
<t>Next, the server checks if the channel number has been confirmed
by the client. If the channel number has been confirmed, the server
simply sends the data to the client with the appropriate channel
number in the TURN framing.</t>
<t>If the channel number has not been confirmed, the server creates
a Data indication. It places the selected channel number in a
CHANNEL-NUMBER attribute, the peer IP address and port in a
PEER-ADDRESS attribute, and puts the data to be sent in a DATA
attribute. If the Data indication is sent over a reliable transport
(ex: TCP), the server marks that the channel number as confirmed.
When the server receives a ChannelConfirmation Indication, and the
channel number, IP address and port match the channel number
assigned to that peer, the server marks that the channel number is
confirmed.</t>
<t>Since Data is an Indication, it generates no response. The server
does not provide reliability for the data. When sending over a
reliable transport to the client, if the server is unable to send
the data received from the peer (for example, because the TCP
connection cannot accept any more messages right now), it can
silently discards UDP data received from the peer.</t>
<t><list>
<t>Note that Send Indications are not authenticated and do not
contain a MESSAGE-INTEGRITY attribute. Just like non-relayed
data sent over UDP or TCP, the authenticity and integrity of
this data can only be assured using security mechanisms at
higher layers.</t>
</list></t>
</section>
<section anchor="sec:lte"
title="Allocation Activity Timer and Permission Timeout">
<t>When the allocation activity timer expires, the server MUST
destroy the allocation. This involves freeing the allocated
transport address, deleting permissions and channel numbers, and
removing other state associated with the allocation.</t>
<t>When a permission times out, the TURN server MUST NOT forward a
packet from that TURN peer to the TURN client.</t>
</section>
</section>
</section>
<!-- Sending and Receiving Data -->
<section title="New Attributes">
<t>This STUN extension defines the following new attributes:</t>
<figure>
<preamble></preamble>
<artwork><![CDATA[
0x000C: CHANNEL-NUMBER
0x000D: LIFETIME
0x0010: BANDWIDTH
0x0012: PEER-ADDRESS
0x0013: DATA
0x0016: RELAY-ADDRESS
0x0018: REQUESTED-PORT-PROPS
0x0019: REQUESTED-TRANSPORT
0x0022: REQUESTED-IP
]]></artwork>
</figure>
<section anchor="channelnums" title="CHANNEL-NUMBER">
<t>The channel number attribute represents the channel number assigned
by the sender, that corresponds with the peer specified in the
PEER-ADDRESS attribute. It is a 16-bit unsigned integer, plus two
octets of padding which MUST be set to zero.</t>
<figure>
<artwork><![CDATA[
0 1 2 3
0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
| Channel Number | Reserved = 0 |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
]]></artwork>
</figure>
</section>
<section title="LIFETIME">
<t>The lifetime attribute represents the duration for which the server
will maintain an allocation in the absence of a refresh. It is a 32
bit unsigned integral value representing the number of seconds
remaining until expiration.</t>
</section>
<section title="BANDWIDTH">
<t>The bandwidth attribute represents the peak bandwidth, measured in
kilobits per second, that the client expects to use on the allocation
in each direction.</t>
</section>
<section title="PEER-ADDRESS">
<t>The PEER-ADDRESS specifies the address and port of the peer as seen
from the TURN server. It is encoded in the same way as
XOR-MAPPED-ADDRESS.</t>
</section>
<section title="DATA">
<t>The DATA attribute is present in most Send Indications and Data
Indications. It contains raw payload data that is to be sent (in the
case of a Send Request) or was received (in the case of a Data
Indication).</t>
</section>
<section title="RELAY-ADDRESS">
<t>The RELAY-ADDRESS is present in Allocate responses. It specifies
the address and port that the server allocated to the client. It is
encoded in the same way as XOR-MAPPED-ADDRESS.</t>
</section>
<section title="REQUESTED-PORT-PROPS">
<t>This attribute allows the client to request certain properties for
the port that is allocated by the server. The attribute can be used
with any transport protocol that has the notion of a 16 bit port space
(including TCP and UDP). The attribute is 32 bits long. Its format
is:</t>
<figure>
<preamble></preamble>
<artwork><![CDATA[
0 1 2 3
0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
| Reserved = 0 | A | Specific Port Number |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
]]></artwork>
</figure>
<t>The two bits labeled A in the diagram above are for requested port
alignment and have the following meaning:</t>
<figure>
<artwork><![CDATA[
00 no specific port alignment
01 odd port number
10 even port number
11 even port number; reserve next higher port
]]></artwork>
</figure>
<t>If the value of the A field is 00 (no specific port alignment),
then the Specific Port Number field can either be 0 or some non-zero
port number. If the Specific Port Number field is 0, then the client
is not putting any restrictions on the port number it would like
allocated. If the Specific Port Number is some non-zero port number,
then the client is requesting that the server allocate the specified
port.</t>
<t>If the value of the A field is 01 (odd port number), then the
Specific Port Number field must be zero, and the client is requesting
the server allocate an odd-numbered port.</t>
<t>If the value of the A field is 10 (even port number), then the
Specific Port number field must be zero, and the client is requesting
the server allocate an even-numbered port.</t>
<t>If the value of the A field is 11 (even port number; reserve next
higher port), then the Specific Port Number field must be zero, and
the client is requesting the server allocate an even-numbered port. In
addition, the client is requesting the server reserve the next higher
port (i.e., N+1 if the server allocates port N), and should only
allocate the N+1 port number if it is explicit requested (with a
subsequent request specifying that exact port number)</t>
<t>In all cases, if a port with the requested properties cannot be
allocated, the server responds with a error response with an error
code of 444 (Invalid Port).</t>
</section>
<section title="REQUESTED-TRANSPORT">
<t>This attribute is used by the client to request a specific
transport protocol for the allocated transport address. It is a 32 bit
unsigned integer. Its values are:</t>
<figure>
<preamble></preamble>
<artwork><![CDATA[
0x0000 0000: UDP
0x0000 0001: Reserved for TCP
]]></artwork>
</figure>
<t>If an Allocate request is sent over TCP and requests a UDP
allocation, or an Allocate request is sent over TLS over TCP and
requests a UDP allocation, the server will relay data between the two
transports.</t>
<t>Extensions to TURN can define additional transport protocols in an
IETF-consensus RFC.</t>
</section>
<section title="REQUESTED-IP">
<t>The REQUESTED-IP attribute is used by the client to request that a
specific IP address be allocated to it. This attribute is needed since
it is anticipated that TURN servers will be multi-homed so as to be
able to allocate more than 64k transport addresses. As a consequence,
a client needing a second transport address on the same interface as a
previous one can make that request.</t>
<t>The format of this attribute is identical to XOR-MAPPED-ADDRESS.
However, the port component of the attribute is ignored by the server.
If a client wishes to request a specific IP address and port, it uses
both the REQUESTED-IP and REQUESTED-PORT-PROPS attributes.</t>
</section>
</section>
<!-- new attributes -->
<section title="New Error Response Codes">
<t>This document defines the following new Error response codes:</t>
<t><list style="hanging">
<t hangText="437">(Allocation Mismatch): A request was received by
the server that requires an allocation to be in place, but there is
none, or a request was received which requires no allocation, but
there is one.</t>
<t hangText="442">(Unsupported Transport Protocol): The Allocate
request asked for a transport protocol to be allocated that is not
supported by the server. If the server is aware of another server
that supports the requested protocol, it SHOULD include the other
server's address in an ALTERNATE-SERVER attribute in the error
response.</t>
<t hangText="443">(Invalid IP Address): The Allocate request asked
for a transport address to be allocated from a specific IP address
that is not valid on the server.</t>
<t hangText="444">(Invalid Port): The Allocate request asked for a
port to be allocated that is not available on the server.</t>
<t hangText="486">(Allocation Quota Reached): The user or client is
not authorized to request additional allocations.</t>
<t hangText="507">(Insufficient Capacity): The server cannot
allocate a new port for this client as it has exhausted its relay
capacity.</t>
</list></t>
</section>
<!-- error codes -->
<section title="Client Discovery of TURN Servers">
<t>The STUN extensions introduced by TURN differ from the binding
requests defined in <xref target="I-D.ietf-behave-rfc3489bis"></xref> in
that they are sent with additional framing and demand substantial
resources from the TURN server. In addition, it seems likely that
administrators might want to block connections from clients to the TURN
server for relaying separately from connections for the purposes of
binding discovery. As a consequence, TURN runs on a separate port from
STUN. The client discovers the address and port of the TURN server using
the same DNS procedures defined in <xref
target="I-D.ietf-behave-rfc3489bis"></xref>, but using an SRV service
name of "turn" (or "turns" for TURN over TLS) instead of just
"stun".</t>
<t>For example, to find TURN servers in the example.com domain, the TURN
client performs a lookup for '_turn._udp.example.com',
'_turn._tcp.example.com', and '_turns._tcp.example.com' if the STUN
client wants to communicate with the TURN server using UDP, TCP, or TLS
over TCP, respectively.</t>
</section>
<!-- Client Discovery -->
<section anchor="sec:sec" title="Security Considerations">
<t>TURN servers allocate bandwidth and port resources to clients, in
contrast to the Binding method defined in <xref
target="I-D.ietf-behave-rfc3489bis"></xref>. Therefore, a TURN server
requires authentication and authorization of STUN requests. This
authentication is provided by mechanisms defined in the STUN
specification itself, in particular digest authentication.</t>
<t>Because TURN servers allocate resources, they can be susceptible to
denial-of-service attacks. All Allocate transactions are authenticated,
so that an unknown attacker cannot launch an attack. An authenticated
attacker can generate multiple Allocate Requests, however. To prevent a
single malicious user from allocating all of the resources on the
server, it is RECOMMENDED that a server implement a modest per user
limit on the amount of bandwidth that can be allocated. Such a mechanism
does not prevent a large number of malicious users from each requesting
a small number of allocations. Attacks such as these are possible using
botnets, and are difficult to detect and prevent. Implementors of TURN
should keep up with best practices around detection of anomalous botnet
attacks.</t>
<t>A client will use the transport address learned from the
RELAY-ADDRESS attribute of the Allocate Response to tell other users how
to reach them. Therefore, a client needs to be certain that this address
is valid, and will actually route to them. Such validation occurs
through the message integrity checks provided in the Allocate response.
They can guarantee the authenticity and integrity of the allocated
addresses. Note that TURN is not susceptible to the attacks described in
Section 12.2.3, 12.2.4, 12.2.5 or 12.2.6 of <xref
target="I-D.ietf-behave-rfc3489bis"></xref> [[TODO: Update section
number references to 3489bis]]. These attacks are based on the fact that
a STUN server mirrors the source IP address, which cannot be
authenticated. STUN does not use the source address of the Allocate
Request in providing the RELAY-ADDRESS, and therefore, those attacks do
not apply.</t>
<t>TURN cannot be used by clients for subverting firewall policies. TURN
has fairly limited applicability, requiring a user to explicitly
authorize permission to receive data from a peer, one IP address at a
time. Thus, it does not provide a general technique for externalizing
sockets. Rather, it has similar security properties to the placement of
an address-restricted NAT in the network, allowing messaging in from a
peer only if the internal client has sent a packet out towards the IP
address of that peer. This limitation means that TURN cannot be used to
run web servers, email servers, SIP servers, or other network servers
that service a large number of clients. Rather, it facilitates
rendezvous of NATted clients that use some other protocol, such as SIP,
to communicate IP addresses and ports for communications.</t>
<t>Confidentiality of the transport addresses learned through Allocate
transactions does not appear to be that important. If required, it can
be provided by running TURN over TLS.</t>
<t>TURN does not and cannot guarantee that UDP data is delivered in
sequence or to the correct address. As most TURN clients will only
communicate with a single peer, the use of a single channel number will
be very common. Consider an enterprise where Alice and Bob are involved
in separate calls through the enterprise NAT to their corporate TURN
server. If the corporate NAT reboots, it is possible that Bob will
obtain the exact NAT binding originally used by Alice. If Alice and Bob
were using identical channel numbers, Bob will receive unencapsulated
data intended for Alice and will send data accidentally to Alice's peer.
This is not a problem with TURN. This is precisely what would happen if
there was no TURN server and Bob and Alice instead provided a (STUN)
reflexive transport address to their peers. If detecting this
misdelivery is a problem, the client and its peer need to use message
integrity on their data.</t>
<t>One TURN-specific DoS attack bears extra discussion. An attacker who
can corrupt, drop, or cause the loss of a Send or Data indication sent
over UDP, and then forge a Channel Confirmation indication for the
corresponding channel number, can cause a TURN client (server) to start
sending unencapsulated data that the server (client) will discard. Since
indications are not integrity protected, this attack is not prevented by
cryptographic means. However, any attacker who can generate this level
of network disruption could simply prevent a large fraction of the data
from arriving at its destination, and therefore protecting against this
attack does not seem important. The ChannelConfirmation forging attack
is not possible when the client to server communication is over TCP or
TLS over TCP.</t>
<t>Relay servers are useful even for users not behind a NAT. They can
provide a way for truly anonymous communications. A user can cause a
call to have its media routed through a TURN server, so that the user's
IP addresses are never revealed.</t>
<t>Any relay addresses learned through an Allocate request will not
operate properly with <xref target="RFC4302">IPSec Authentication Header
(AH)</xref> in transport or tunnel mode. However, tunnel-mode <xref
target="RFC4303">IPSec ESP</xref> should still operate.</t>
</section>
<!-- Security -->
<section title="IANA Considerations">
<t>This specification defines several new STUN methods, STUN attributes,
and STUN response codes. This section directs IANA to add these new
protocol elements to the IANA registry of STUN protocol elements.</t>
<section title="New STUN Methods">
<figure>
<preamble></preamble>
<artwork><![CDATA[
Request/Response Transactions
0x003 : Allocate
0x004 : Refresh
Indications
0x006 : Send
0x007 : Data
0x009 : Channel Confirmation
]]></artwork>
</figure>
</section>
<section title="New STUN Attributes">
<figure>
<preamble></preamble>
<artwork><![CDATA[
0x000C: CHANNEL-NUMBER
0x000D: LIFETIME
0x0010: BANDWIDTH
0x0012: PEER-ADDRESS
0x0013: DATA
0x0016: RELAY-ADDRESS
0x0018: REQUESTED-PORT-PROPS
0x0019: REQUESTED-TRANSPORT
0x0022: REQUESTED-IP
]]></artwork>
</figure>
</section>
<section title="New STUN Response Codes">
<figure>
<artwork><![CDATA[
437 Allocation Mismatch
442 Unsupported Transport Protocol
443 Invalid IP Address
444 Invalid Port
486 Allocation Quota Reached
507 Insufficient Capacity
]]></artwork>
</figure>
</section>
</section>
<section title="IAB Considerations">
<t>The IAB has studied the problem of "Unilateral Self Address Fixing",
which is the general process by which a client attempts to determine its
address in another realm on the other side of a NAT through a
collaborative protocol reflection mechanism <xref target="RFC3424">RFC
3424</xref>. The TURN extension is an example of a protocol that
performs this type of function. The IAB has mandated that any protocols
developed for this purpose document a specific set of
considerations.</t>
<t>TURN is an extension of the STUN protocol. As such, the specific
usages of STUN that use the TURN extensions need to specifically address
these considerations. Currently the only STUN usage that uses TURN is
<xref target="I-D.ietf-mmusic-ice">ICE</xref>.</t>
</section>
<!-- IAB Considerations -->
<section title="Example">
<t>In this example, a TURN client is behind a NAT. This TURN client is
running SIP. The client has a private address of 10.0.1.1. The TURN
server is on the public side of the NAT, and is listening for TURN
requests on 192.0.2.3:8776. The public side of the NAT has an IP address
of 192.0.2.1. The client is attempting to send a SIP INVITE to a peer,
and wishes to allocate an IP address and port for inclusion in the SDP
of the INVITE. Normally, TURN would be used in conjunction with ICE when
applied to SIP. However, to keep the example simple, TURN is shown
without ICE.</t>
<t>The client communicates with a SIP user agent on the public network.
This user agent uses a 192.0.2.17:12734 for receipt of its RTP
packets.</t>
<figure anchor="fig-example">
<artwork><![CDATA[
10.0.1.1 192.0.2.1 192.0.2.3 192.0.2.17
Client NAT TURN Server Peer
| | | |
|(1) Allocate |(2) Allocate | |
|S=10.0.1.1:4334 |S=192.0.2.1:63346 | |
|D=192.0.2.3:8776 |D=192.0.2.3:8776 | |
|------------------>|------------------>| |
| | | |
|(4) Error |(3) Error | |
|S=192.0.2.3:8776 |S=192.0.2.3:8776 | |
|D=10.0.1.1:4334 |D=192.0.2.1:63346 | |
|<------------------|<------------------| |
| | | |
|(5) Allocate |(6) Allocate | |
|S=10.0.1.1:4334 |S=192.0.2.1:63346 | |
|D=192.0.2.3:8776 |D=192.0.2.3:8776 | |
|------------------>|------------------>| |
| | | |
| | (allocates port 32766) |
| | | |
| | | |
|(8) Response |(7) Response | |
|RA=192.0.2.3:32766 |RA=192.0.2.3:32766 | |
|MA=192.0.2.1:63346 |MA=192.0.2.1:63346 | |
|S=192.0.2.3:8776 |S=192.0.2.3:8776 | |
|D=10.0.1.1:4334 |D=192.0.2.1:63346 | |
|<------------------|<------------------| |
| | | |
|(9) SIP INVITE | | |
|SDP=192.0.2.3:32766| | |
|---------------------------------------------------------->|
| | | |
|(10) SIP 200 OK | | |
|SDP=192.0.2.17:12734 | |
|<----------------------------------------------------------|
| | | |
| | |(11) RTP |
| | |S=192.0.2.17:12734 |
| | |D=192.0.2.3:32766 |
| | |<------------------|
| | | |
| | (no permission; packet dropped) |
| | | |
|(12) SIP ACK | | |
|---------------------------------------------------------->|
| | | |
|(13) Send Indic. |(14) Send Indic. | |
|TURN Channel=0 |TURN Channel=0 | |
|STUN DATA=RTP |STUN DATA=RTP | |
|CHANNEL-NUMER=77 |CHANNEL-NUMBER=77 | |
|PA=192.0.2.17:12734|PA=192.0.2.17:12734| |
|S=10.0.1.1:4334 |S=192.0.2.1:63346 | |
|D=192.0.2.3:8776 |D=192.0.2.3:8776 | |
|------------------>|------------------>| |
| | | |
| | permission created |
| | | |
| | |(15) RTP |
| | |S=192.0.2.3:32766 |
| | |D=192.0.2.17:12734 |
| | |------------------>|
| | | |
|(17) ChannelConf |(16) ChannelConf | |
|TURN Channel=0 |TURN Channel=0 | |
|CHANNEL-NUMBER=77 |CHANNEL-NUMBER=77 | |
|PA=192.0.2.17:12734|PA=192.0.2.17:12734| |
|S=192.0.2.3:8776 |S=192.0.2.3:8776 | |
|D=10.0.1.1:4334 |D=192.0.2.1:63346 | |
|<------------------|<------------------| |
| | | |
|(18) TURN Framed |(19) TURN Framed | |
|TURN Channel=77 |TURN Channel=77 |(20) RTP |
|S=10.0.1.1:4334 |S=192.0.2.1:63346 |S=192.0.2.3:32766 |
|D=192.0.2.3:8776 |D=192.0.2.3:8776 |D=192.0.2.17:12734 |
|------------------>|------------------>|------------------>|
| | | |
|(23) Data Indic. |(22) Data Indic. | |
|TURN Channel=0 |TURN Channel=0 | |
|CHANNEL-NUMBER=33 |CHANNEL-NUMBER=33 |(21) RTP |
|S=192.0.2.3:8776 |S=192.0.2.3:8776 |S=192.0.2.17:12734 |
|D=10.0.1.1:4334 |D=192.0.2.1:63346 |D=192.0.2.3:32766 |
|<------------------|<------------------|<------------------|
| | | |
|(24) ChannelConf |(25) ChannelConf | |
|TURN Channel=0 |TURN Channel=0 | |
|CHANNEL-NUMBER=33 |CHANNEL-NUMBER=33 | |
|S=10.0.0.1:4334 |S=192.0.2.3:8776 | |
|D=192.0.2.3:8776 |D=192.0.2.3:8776 | |
|------------------>|------------------>| |
| | | |
|(28) TURN Framed |(27) TURN Framed | |
|TURN Channel=33 |TURN Channel=33 |(26) RTP |
|S=192.0.2.3:8776 |S=192.0.2.3:8776 |S=192.0.2.17:12734 |
|D=10.0.1.1:4334 |D=192.0.2.1:63346 |D=192.0.2.3:32766 |
|<------------------|<------------------|<------------------|
| | | |
]]></artwork>
</figure>
<t>The message flow is shown in <xref target="fig-example"></xref>. In
step 1-2, the client allocates a UDP port from the local operating
system on its private interface, obtaining 4334. It then attempts to
obtain a port for RTP traffic. RTCP processing is not shown in the
example.</t>
<t>In step 1, the client sends an Allocate Request (1) with a source
address (denoted by S) of 10.0.1.1:4334 and a destination (denoted by D)
of 192.0.2.3:8776. This passes through the NAT (2), which allocates a
new UDP port (63346) on the NAT's public interface (192.0.2.1), and
creates an internal mapping between the internal address 10.0.1.1:4334
and that external address 192.0.2.1:63346. The NAT sends this request to
the TURN server (3). The TURN server challenges the request, requesting
credentials by sending a STUN error and including the NONCE and REALM
attributes. Message 3 is relayed, by the NAT, to the TURN client (4).
The client sends a new request (from the same UDP port), including its
credentials (5, 6). The TURN server authenticates the request. The TURN
server allocates a new UDP port on one of its interfaces,
192.0.2.3:32766. The TURN server puts 192.0.2.3:32766 into the
RELAY-ADDRESS (denoted by RA) attribute of the response, and puts the
source IP address and UDP port of the request (as seen by the TURN
server) into the XOR-MAPPED-ADDRESS attribute (denoted by MA). In step
7, this message is sent back to the TURN client and relayed by the NAT
in step 8.</t>
<t>The client now proceeds to perform a basic SIP call setup. In message
9, the TURN client includes the TURN server's address (which it learned
in message 8) in the SDP of its INVITE (e.g., using syntax described
in<xref target="I-D.ietf-mmusic-ice"></xref>). The called party responds
with its SDP in a provisional response (18x) or a final response (200
Ok). The called party's SDP includes its IP address and UDP port,
192.0.2.17:12734. Immediately after sending its 200 Ok, the called party
sends an RTP packet to the TURN server's IP address (11). This RTP
packet is dropped by the TURN server, because the TURN server has not
been given permission to relay that data. Incoming packets are dropped
until a permission is created. The SIP exchange completes with an SIP
200 Ok message (12).</t>
<t>Steps 13-20 show the client performing a channel allocation. The TURN
client needs to send an RTP packet. Since no channels and no permissions
have been created, the TURN client sends the RTP packet inside of a Send
Indication, using channel number 0, with the CHANNEL-NUMBER attribute
set to the channel number the TURN client wants to use for subsequent
communication with this TURN peer (77 is shown in the example). The TURN
peer's IP address and UDP port (which were learned from the SDP answer
received in step 10) are placed in the PEER-ADDRESS attribute (denoted
by PA). In message 13, the TURN client sends this Send Indication, and
it is relayed by the NAT to the TURN server (14). Upon receipt of that
message, the TURN server creates a permission, which allows subsequent
traffic from that same peer address to be relayed to that TURN client's
IP address and UDP port. The TURN server sends the contents of the Send
Indication's DATA attribute towards the PEER-ADDRESS (15); this will
typically be an RTP packet. Note that the source address and port of
message 15 is the TURN server's address, 192.0.2.3:32766, which is the
allocated transport address communicated to the TURN client in messages
7 and 8.</t>
<t>In step 16, the TURN server sends a channel confirmation message to
the TURN client. Once the TURN client receives this message, it can
forgo using the Send Indication for that channel. Instead, it can
utilize the channel number in the TURN framing header. Steps 18 and 19
show the TURN client sending a message to TURN server using the TURN
framing header, with channel=1. Step 20 shows the TURN server removing
the TURN framing and sending the RTP packet to the TURN peer.</t>
<t>Steps 21-28 show an RTP packet from the TURN peer, which causes a
channel allocation by the TURN server. In packet 21, an RTP packet is
sent by the TURN peer to the TURN server. There is an existing
permission (created in step 14), so the TURN server accepts this
incoming RTP packet. The TURN server knows the TURN client to send this
packet to, but does not yet have a channel assigned for traffic in this
direction. The TURN server chooses a channel number (33 in the example),
and sends a Data Indication to the TURN client (message 22). The NAT
relays this to the TURN client (message 23). The TURN client sends an
Channel Confirmation message (24) which is relayed by the NAT (25). When
the TURN server receives the Channel Confirmation, it no longer needs to
use a Send Indication for traffic from that remote peer; instead, it can
use TURN framing with its chosen channel number (33). The next RTP
packet that arrives from that peer (26) is sent by the TURN server using
TURN framing indicating the channel number (message 27) and relayed by
the NAT to the TURN client (28).</t>
</section>
<!-- Example -->
<section title="Changes since version -04">
<t>This section lists the major changes between thiis document and
draft-ietf-behave-turn-04:<list style="symbols">
<t>Removed the ability to allocate addresses for TCP relaying. This
is now covered in a separate document. However, communication
between the client and the server can still run over TCP or TLS/TCP.
This resulted in the removal of the Connect method and the TIMER-VAL
and CONNECT-STAT attributes.</t>
<t>Added the concept of channels. All communication between the
client and the server flows on a channel. Channels are numbered
0..65535. Channel 0 is used for TURN messages, while the remaining
channels are used for sending unencapsulated data to/from a remote
peer. This concept adds a new Channel Confirmation method and a new
CHANNEL-NUMBER attribute. The new attribute is also used in the Send
and Data methods.</t>
<t>The framing mechanism formally used just for stream-oriented
transports is now also used for UDP, and the former Type and
Reserved fields in the header have been replaced by a Channel Number
field. The length field is zero when running over UDP.</t>
<t>TURN now runs on its own port, rather than using the STUN port.
The use of channels requires this.</t>
<t>Removed the SetActiveDestination concept. This has been replaced
by the concept of channels.</t>
<t>Changed the allocation refresh mechanism. The new mechanism uses
a new Refresh method, rather than repeating the Allocation
transaction.</t>
<t>Changed the syntax of SRV requests for secure transport. The new
syntax is "_turns._tcp" rather than the old "_turn._tls". This
change mirrors the corresponding change in STUN SRV syntax.</t>
<t>Renamed the old REMOTE-ADDRESS attribute to PEER-ADDRESS, and
changed it to use the XOR-MAPPED-ADDRESS format.</t>
<t>Changed the RELAY-ADDRESS attribute to use the XOR-MAPPED-ADDRESS
format (instead of the MAPPED-ADDRESS format)).</t>
<t>Renamed the 437 error code from "No Binding" to "Allocation
Mismatch".</t>
<t>Added a discussion of what happens if a client's public binding
on its outermost NAT changes.</t>
<t>The document now consistently uses the term "peer" as the name of
a remote endpoint with which the client wishes to communicate.</t>
<t>Rewrote much of the document to describe the new concepts. At the
same time, tried to make the presentation clearer and less
repetitive.</t>
</list></t>
</section>
<section title="Acknowledgements">
<t>The authors would like to thank Marc Petit-Huguenin for his comments
and suggestions.</t>
</section>
</middle>
<back>
<references title="Normative References">
<?rfc include="reference.I-D.ietf-behave-rfc3489bis"?>
<?rfc include="reference.RFC.2119"?>
</references>
<references title="Informative References">
<?rfc include="reference.RFC.3550"?>
<?rfc include="reference.RFC.3264"?>
<?rfc include="reference.RFC.4302"?>
<?rfc include="reference.RFC.4303"?>
<?rfc include="reference.RFC.3424"?>
<?rfc include="reference.I-D.ietf-mmusic-ice"?>
<?rfc include="reference.RFC.4787"?>
<?rfc include="reference.I-D.ietf-behave-turn-tcp"?>
</references>
</back>
</rfc>
<!--
<?rfc include="reference.RFC.2782"?>
<?rfc include="reference.RFC.2617"?>
<?rfc include="reference.RFC.2327"?>
<?rfc include="reference.RFC.2326"?>
<?rfc include="reference.RFC.3235"?>
-->
<!-- Notes for Rohan from conversation with JDR
1. The Connect Request doesn't make sense. If the peer could receive a TCP connection request,
the client would just open a connection to the peer directly.
2. Still need the concept of a door if we want TCP to TCP case to work behind most NATs/FWs.
Took serveral steps to make sure door concept does not allow you to run a real server.
Can't ask for a door and specific port number. Can't get a door with a well known port.
Can't open a door twice.
(Also, the door can eliminate the use of two TURN servers if the caller and callee are both
behind bad NATs, as long as thier is no forking. With ICE, everything works fine.
You try one TURN server and if that doesn't work you try both.)
3. Introduce lightweight TCP framing. Send any data with magic cookie over UDP in a Data/Send
4. Disallow UDP (client to relay) to TCP (relay to peer) case.
5. Add connection status notification as an indication. This works well since the client to
server connection is always reliable now anytime you would get this indication.
6. Add "move my flow over here" feature.
-->
<!-- from my notes while doing turn-00-->
<!-- need to align the turn draft with the long term direction of stun
and turn for control of nat. In particular, this means ultimately
allowing multiple mapped address attributes returned from a stun query
(one from each nat processing the stun) and also having the binding
lifetime make its way into turn somehow -->
<!-- need to fix the big disambiguation problem, of how to know whether
this stun request is for me as a turn server, or for a downstream stun server
as in the ice case -->
<!-- terminology check - the draft uses the terms binding and allocation
too loosely. Allocation is the transport address allocated to the client,
and permissions needs to be used consistently. Also the term external
5-tuples is used a bunch and this probably needs to be removed. Other
terminology things: use correspondent instead of external client; it
sound sbetter. Also suggest allocated transport address instead of
external local, and well-known transport address instead of internal
local. Suggest reflexive for internal remote. -->
<!-- Draft needs to be a bit clearer on lifecycle management of
allocations. -->
<!-- Encoding of addresses - should these be using xor form or regular
form? in places like RELAY-ADDRESs and DESTINATION-ADDRESS -->
<!-- IANA registry needs to be added. THis includes registering the
new methods and attributes, and creating registries for the few things
in here, like port properties -->
<!-- Check to make sure all response codes mentioned in the back of
the document are used, and vice a versa -->
<!-- need to clean up terminology around shared secrets. There are two
types, long term and short term, with a relationship between them
(been using derived-from). That needs to be explained. Also, need to
be clearer on which credentials are reall required in a turn request -
needs to be equal to or cousin to one from original allocate -->
<!-- rewrite IAB consideratoins -->
<!-- major work on security considerations - what about data packets
and indications in particular? -->
<!-- scrub references, make sure all are used -->
<!-- spec says you need to include mapped address. But, more like, use
mapped address if magic cookie is not there, and xor-mapped if it is
-->
<!-- remove active 5-tuple terminology -->
<!-- I've used the term 'linkage' to try and include the traditional
tcp connection and udp connection. Need to decide if i like it and use
it or not -->
<!-- should i make turn into a mobile ip replacement? can be done, by
allowing internal address to be updated in an allocate. requires
changing keying structures around. today, incoming allocate uses
incoming 'linkage' or connection to refer to allocation. Would need to
be identified some other way. Would be cool though. Thats an OPEN
ISSUE -->
<!-- add table of mandatory/optional attributes here -->
<!-- maybe add a discussion about how alternate server can be used,
with a front end load balancer -->
<!-- Add text about using multiple virtual turn servers to deal with the
case where the turn server is effectively or virtually mutli-homed. -->
<!-- is message-integrity needed for the error responses? If so, it
belongs in rfc3489bis too -->
<!-- ISSUE: how to allow the server whether to know to start tls
procedures when it receives a tcp connection? Is this assuming
start-tls or are we using a separate port?? -->
<!-- ISSUE: once udp destination is set, server looks for TURN packets by magic cookie. But what
if packet being send is e2e stun connectivity check which
also has magic cookie! This is a real use case. Need to move to model where
control transitions, like for tcp.
-->
<!-- xtunnel 1: udp to tcp, specify qos parameters for leaky bucket -->
<!-- xteunnel 2: tcp one direction, udp the other?? -->
<!-- xtunnel 3: forking issue - requires creation of a new allocation. Alternative is to always
encapsulate with something more lightweight -->
<!-- xtunnel 4: some issue with requiring a third changed address to deal with running out of
local ports on the client. Rob says he'll send me some text
on the situation -->
<!-- add example -->
<!-- turn encapsulation issue raised by Justin Umberti - what if the data
that is sent unencapsulated is less than 7 bytes? Server has to buffer
7 (is it 7) bytes to look for the cookie. -->
<!-- from ietf65: add note that contiguous port requests is for legacy
interop -->
<!-- new framing proposal: use my 32 bit framing word for tcp. Note that,
you can put multiple tcp application frames within a data frame. Length
only indicates frequency of stun signaling. For UDP, don't use this.
However, packets inwards towards stun relay, stun relay looks for
magic cookie in the bytes 5-7 and uses data indication if they are
there. Similarly, client should look for cookie in those bytes before
sending and if present, use Send. This means you don't have the
problem of figuring out whether data is destined for this turn server
or not - those data are always done with Send. Nice. -->
<!-- agree to have refreshes come from different source/dest IP -->
<!-- from adam's note on behave 3/28, remove capability for udp from client
to turn server and then tcp outwards. No way this works -->| PAFTECH AB 2003-2026 | 2026-04-24 01:20:46 |