One document matched: draft-ietf-mmusic-rtsp-nat-evaluation-09.xml


<?xml version="1.0" encoding="US-ASCII"?>
<!DOCTYPE rfc SYSTEM "rfc2629.dtd">
<?rfc toc="yes"?>
<?rfc tocompact="yes"?>
<?rfc tocdepth="3"?>
<?rfc tocindent="yes"?>
<?rfc symrefs="yes"?>
<?rfc sortrefs="yes"?>
<?rfc comments="yes"?>
<?rfc inline="yes"?>
<?rfc compact="yes"?>
<?rfc subcompact="no"?>
<rfc category="info" docName="draft-ietf-mmusic-rtsp-nat-evaluation-09"
     ipr="pre5378Trust200902">
  <front>
    <title abbrev="Evaluation of NAT Traversal for RTSP">The Evaluation of
    Different Network Address Translator (NAT) Traversal Techniques for Media
    Controlled by Real-time Streaming Protocol (RTSP)</title>

    <author fullname="Magnus Westerlund" initials="M." surname="Westerlund">
      <organization>Ericsson</organization>

      <address>
        <postal>
          <street>Farogatan 6</street>

          <city>Stockholm</city>

          <region/>

          <code>SE-164 80</code>

          <country>Sweden</country>
        </postal>

        <phone>+46 8 719 0000</phone>

        <facsimile/>

        <email>magnus.westerlund@ericsson.com</email>

        <uri/>
      </address>
    </author>

    <author fullname="Thomas Zeng" initials="T." surname="Zeng">
      <organization/>

      <address>
        <postal>
          <street/>

          <city/>

          <region/>

          <code/>

          <country/>
        </postal>

        <phone/>

        <facsimile/>

        <email>thomas.zeng@gmail.com</email>

        <uri/>
      </address>
    </author>

    <date day="29" month="May" year="2013"/>

    <abstract>
      <t>This document describes several Network Address Translator (NAT)
      traversal techniques that were considered to be used for establishing
      the RTP media flows controlled by the Real-time Streaming Protocol
      (RTSP). Each technique includes a description on how it would be used,
      the security implications of using it and any other deployment
      considerations it has. There are also discussions on how NAT traversal
      techniques relates to firewalls and how each technique can be applied in
      different use cases. These findings where used when selecting the NAT
      traversal for RTSP 2.0.</t>
    </abstract>
  </front>

  <middle>
    <section title="Introduction">
      <t>Today there is a proliferate deployment of different flavors of
      Network Address Translator (NAT) boxes that in many cases only loosely
      follow <xref target="RFC3022">standards</xref><xref
      target="RFC2663"/><xref target="RFC3424"/><xref target="RFC4787"/><xref
      target="RFC5382"/>. NATs cause discontinuity in <xref
      target="RFC3424">address realms</xref>, therefore an application
      protocol, such as <xref target="RFC2326">Real-time Streaming Protocol
      (RTSP)</xref><xref target="I-D.ietf-mmusic-rfc2326bis"/>, needs to deal
      with such discontinuities caused by NATs. The problem is that, being a
      media control protocol managing one or more media streams, RTSP carries
      network address and port information within its protocol messages.
      Because of this, even if RTSP itself, when carried over <xref
      target="RFC0793">Transmission Control Protocol (TCP)</xref> for example,
      is not blocked by NATs, its media streams may be blocked by NATs. This
      will occur unless special protocol provisions are added to support
      NAT-traversal.</t>

      <t>Like NATs, Firewalls are also middle boxes that need to be
      considered. Firewalls helps prevent unwanted traffic from getting in or
      out of the protected network. RTSP is designed such that a firewall can
      be configured to let RTSP controlled media streams go through with
      minimal implementation effort. The minimal effort is to implement an
      Application Level Gateway (ALG) to interpret RTSP parameters. There is
      also a large class of firewalls, commonly home firewalls, that uses a
      similar filtering behavior to what NAT has. This type of firewalls can
      be handled using the same solution as employed for NAT traversal instead
      of relying on ALGs.</t>

      <t>This document describes several NAT-traversal mechanisms for RTSP
      controlled media streaming. Many of these NAT solutions fall into the
      category of "UNilateral Self-Address Fixing (UNSAF)" as defined in <xref
      target="RFC3424"/> and quoted below:</t>

      <t>"UNSAF is a process whereby some originating process attempts to
      determine or fix the address (and port) by which it is known - e.g. to
      be able to use address data in the protocol exchange, or to advertise a
      public address from which it will receive connections."</t>

      <t>Following the guidelines spelled out in RFC 3424, we describe the
      required RTSP protocol extensions for each method, transition
      strategies, and security concerns.</t>

      <t>This document is capturing the evaluation done in the process to
      recommend Firewall/NAT traversal methods for RTSP streaming servers
      based on <xref target="RFC2326">RFC 2326</xref> as well as the <xref
      target="I-D.ietf-mmusic-rfc2326bis">RTSP 2.0 core spec</xref>. The
      evaluation is focused on NAT traversal for the media streams carried
      over <xref target="RFC0768">User Datagram Protocol (UDP)</xref> with
      <xref target="RFC3550">Real-time Transport Protocol (RTP)</xref> over
      UDP being the main case for such usage. The findings should be
      applicable to other protocols as long as they have similar
      properties.</t>

      <t>The resulting ICE-based RTSP NAT traversal mechanism is specified in
      <xref target="I-D.ietf-mmusic-rtsp-nat">"A Network Address Translator
      (NAT) Traversal mechanism for media controlled by Real-Time Streaming
      Protocol (RTSP)"</xref>.</t>

      <section anchor="sec-nat-intro" title="Network Address Translators">
        <t>Readers are urged to refer to <xref target="RFC2663">"IP Network
        Address Translator (NAT) Terminology and Considerations"</xref> for
        information on NAT taxonomy and terminology. Traditional NAT is the
        most common type of NAT device deployed. Readers may refer to <xref
        target="RFC3022">"Traditional IP Network Address Translator
        (Traditional NAT)"</xref> for detailed information on traditional NAT.
        Traditional NAT has two main varieties -- Basic NAT and Network
        Address/Port Translator (NAPT).</t>

        <t>NAPT is by far the most commonly deployed NAT device. NAPT allows
        multiple internal hosts to share a single public IP address
        simultaneously. When an internal host opens an outgoing TCP or UDP
        session through a NAPT, the NAPT assigns the session a public IP
        address and port number, so that subsequent response packets from the
        external endpoint can be received by the NAPT, translated, and
        forwarded to the internal host. The effect is that the NAPT
        establishes a NAT mapping to translate the (private IP address,
        private port number) tuple to a (public IP address, public port
        number) tuple, and vice versa, for the duration of the session. An
        issue of relevance to peer-to-peer applications is how the NAT behaves
        when an internal host initiates multiple simultaneous sessions from a
        single (private IP, private port) endpoint to multiple distinct
        endpoints on the external network. In this specification, the term
        "NAT" refers to both "Basic NAT" and "Network Address/Port Translator
        (NAPT)".</t>

        <t>This document uses the term "address and port mapping" as the
        translation between an external address and port and an internal
        address and port. Note that this is not the same as an "address
        binding" as defined in RFC 2663. There exist a number of address and
        port mapping behaviors described in more detail in Section 4.1 of
        <xref target="RFC4787">"Network Address Translation (NAT) Behavioral
        Requirements for Unicast UDP"</xref>.</t>

        <t>NATs also have a filtering behavior on traffic arriving on the
        external side. Such behavior affects how well different methods for
        NAT traversal works through these NATs. See Section 5 of <xref
        target="RFC4787">"Network Address Translation (NAT) Behavioral
        Requirements for Unicast UDP"</xref> for more information on the
        different types of filtering that have been identified.</t>
      </section>

      <section title="Firewalls">
        <t>A firewall is a security gateway that enforces certain access
        control policies between two network administrative domains: a private
        domain (intranet) and a external domain, e.g. public Internet. Many
        organizations use firewalls to prevent privacy intrusions and
        malicious attacks to corporate computing resources in the private
        intranet <xref target="RFC2588"/>.</t>

        <t>A comparison between NAT and Firewall is given below:</t>

        <t><list style="numbers">
            <t>A firewall must sit between two network administrative domains,
            while NAT does not have to sit between two domains.</t>

            <t>NAT does not in itself provide security, although some access
            control policies can be implemented using address translation
            schemes. The inherent filtering behaviours are commonly mistaken
            for real security policies.</t>
          </list></t>

        <t>It should be noted that many NAT devices intended for Residential
        or small office/home office (SOHO) use include both NATs and firewall
        functionality.</t>

        <t>In the rest of this memo we use the phrase "NAT traversal"
        interchangeably with "Firewall traversal", and "NAT/Firewall
        traversal".</t>
      </section>

      <section title="Glossary">
        <t><list hangIndent="6" style="hanging">
            <t hangText="Address-Dependent Mapping:">The NAT reuses the port
            mapping for subsequent packets sent from the same internal IP
            address and port to the same external IP address, regardless of
            the external port. See <xref target="RFC4787"/>.</t>

            <t hangText="Address and Port-Dependent Mapping:">The NAT reuses
            the port mapping for subsequent packets sent from the same
            internal IP address and port to the same external IP address and
            port while the mapping is still active. See <xref
            target="RFC4787"/>.</t>

            <t hangText="ALG:">Application Level Gateway, an entity that can
            be embedded in a NAT or other middlebox to perform the application
            layer functions required for a particular protocol to traverse the
            NAT/middlebox.</t>

            <t hangText="Endpoint-Independent Mapping:">The NAT reuses the
            port mapping for subsequent packets sent from the same internal IP
            address and port to any external IP address and port. See <xref
            target="RFC4787"/>.</t>

            <t hangText="ICE:">Interactive Connectivity Establishment, see
            <xref target="RFC5245"/>.</t>

            <t hangText="DNS:">Domain Name Service</t>

            <t hangText="DoS:">Denial of Service</t>

            <t hangText="DDoS:">Distributed Denial of Service</t>

            <t hangText="NAT:">Network Address Translator, see <xref
            target="RFC3022"/>.</t>

            <t hangText="NAPT:">Network Address/Port Translator, see <xref
            target="RFC3022"/>.</t>

            <t hangText="RTP:">Real-time Transport Protocol, see <xref
            target="RFC3550"/>.</t>

            <t hangText="RTSP:">Real-Time Streaming Protocol, see <xref
            target="RFC2326"/> and <xref
            target="I-D.ietf-mmusic-rfc2326bis"/>.</t>

            <t hangText="RTT:">Round Trip Times.</t>

            <t hangText="SDP:">Session Description Protocol, see <xref
            target="RFC4566"/>.</t>

            <t hangText="SSRC:">Synchronization source in RTP, see <xref
            target="RFC3550"/>.</t>
          </list></t>
      </section>
    </section>

    <section title="Detecting the loss of NAT mappings">
      <t>Several NAT traversal techniques in the next chapter make use of the
      fact that the NAT UDP mapping's external address and port can be
      discovered. This information is then utilized to traverse the NAT box.
      However any such information is only good while the mapping is still
      valid. As the IAB's UNSAF document <xref target="RFC3424"/> points out,
      the mapping can either timeout or change its properties. It is therefore
      important for the NAT traversal solutions to handle the loss or change
      of NAT mappings, according to RFC3424.</t>

      <t>First, since NATs may also dynamically reclaim or readjust
      address/port translations, "keep-alive" and periodic re-polling may be
      required according to RFC 3424. Secondly, it is possible to detect and
      recover from the situation where the mapping has been changed or
      removed. The loss of a mapping can be detected when no traffic arrives
      for a while. Below we will give some recommendation on how to detect
      loss of NAT mappings when using RTP/RTCP under RTSP control.</t>

      <t>A RTP session normally has both RTP and RTCP streams. The loss of a
      RTP mapping can only be detected when expected traffic does not arrive.
      If a client does not receive data within a few seconds after having
      received the "200 OK" response to a PLAY request, there are likely some
      middleboxes blocking the traffic. However, for a receiver to be more
      certain to detect the case where no RTP traffic was delivered due to NAT
      trouble, one should monitor the RTCP Sender reports. The sender report
      carries a field telling how many packets the server has sent. If that
      has increased and no RTP packets has arrived for a few seconds it is
      likely the RTP mapping has been removed.</t>

      <t>The loss of mapping for RTCP is simpler to detect. RTCP is normally
      sent periodically in each direction, even during the RTSP ready state.
      If RTCP packets are missing for several RTCP intervals, the mapping is
      likely lost. Note that if neither RTCP packets nor RTSP messages are
      received by the RTSP server for a while, the RTSP server has the option
      to delete the corresponding RTP session, SSRC and RTSP session ID,
      because either the client can not get through a middle box NAT/Firewall,
      or that the client is mal-functioning.</t>
    </section>

    <section anchor="req-section" title="Requirements on Solutions">
      <t>This section considers the set of requirements for the evaluation of
      RTSP NAT traversal solutions.</t>

      <t>RTSP is a client-server protocol. Typically service providers deploy
      RTSP servers in the public address realm. However, there are use cases
      where the reverse is true: RTSP clients are connecting from public
      address realm to RTSP servers behind home NATs. This is the case for
      instance when home surveillance cameras running as RTSP servers intend
      to stream video to cell phone users in the public address realm through
      a home NAT. In terms of requirements, the first requirement should be to
      solve the RTSP NAT traversal problem for RTSP servers deployed in a
      public network, i.e. no NAT at the server side.</t>

      <t>The list of feature requirements for RTSP NAT solutions are given
      below:</t>

      <t><list style="numbers">
          <t>Must work for all flavors of NATs, including NATs with Address
          and Port-Dependent Filtering.</t>

          <t>Must work for firewalls (subject to pertinent firewall
          administrative policies), including those with ALGs.</t>

          <t>Should have minimal impact on clients in the open and not
          dual-hosted. RTSP dual-hosting means that the RTSP signalling
          protocol and the media protocol (e.g. RTP) are implemented on
          different computers with different IP addresses.<list
              style="symbols">
              <t>For instance, no extra delay from RTSP connection till
              arrival of media</t>
            </list></t>

          <t>Should be simple to use/implement/administer so people actually
          turn them on<list style="symbols">
              <t>Otherwise people will resort to TCP tunneling through
              NATs</t>

              <t>Discovery of the address(es) assigned by NAT should happen
              automatically, if possible</t>
            </list></t>

          <t>Should authenticate dual-hosted client transport handler to
          prevent DDoS attacks.</t>
        </list>The last requirement addresses the Distributed
      Denial-of-Service (DDoS) threat, which relates to NAT traversal as
      explained below.</t>

      <t>During NAT traversal, when the RTSP server determines the media
      destination (address and port) for the client, the result may be that
      the public IP address of the RTP receiver host is different than the
      public IP address of the RTSP client host. This posts a DDoS threat that
      has significant amplification potentials because the RTP media streams
      in general consist of large number of IP packets. DDoS attacks occur if
      the attacker fakes the messages in the NAT traversal mechanism to trick
      the RTSP server into believing that the client's RTP receiver is located
      on a separate host. For example, user A may use his RTSP client to
      direct the RTSP server to send video RTP streams to target.example.com
      in order to degrade the services provided by target.example.com. Note a
      simple preventative measure commonly deployed is for the RTSP server to
      disallow the cases where the client's RTP receiver has a different
      public IP address than that of the RTSP client. With the increased
      deployment of NAT middleboxes by operators, i.e. carrier grade NAT
      (CGN), the reusing of a public IP address for many customers reduces the
      protection provided. Also in some applications (e.g., centralized
      conferencing), dual-hosted RTSP/RTP clients have valid use cases. The
      key is how to authenticate the messages exchanged during the NAT
      traversal process.</t>
    </section>

    <section anchor="NAT-trav-tech" title="NAT Traversal Techniques">
      <t>There exists a number of potential NAT traversal techniques that can
      be used to allow RTSP to traverse NATs. They have different features and
      are applicable to different topologies; their costs are also different.
      They also vary in security levels. In the following sections, each
      technique is outlined with discussions on the corresponding advantages
      and disadvantages.</t>

      <t>The main evaluation was done prior to 2007 and are based on what was
      available then. This section includes NAT traversal techniques that have
      not been formally specified anywhere else. The overview section of this
      document may be the only publicly available specification of some of the
      NAT traversal techniques. However that is not a real barrier against
      doing an evaluation of the NAT traversal technique. Some other
      techniques have been recommended against or are no longer possible due
      to standardization works' outcome or their failure to progress within
      IETF after the initial evaluation in this document, e.g. RTP No-Op <xref
      target="I-D.ietf-avt-rtp-no-op"/>.</t>

      <section title="Stand-Alone STUN">
        <t/>

        <section title="Introduction">
          <t><xref target="RFC5389">Session Traversal Utilities for NAT
          (STUN)</xref> is a standardized protocol that allows a client to use
          secure means to discover the presence of a NAT between itself and
          the STUN server. The client uses the STUN server to discover the
          address mappings assigned by the NAT. STUN is a client-server
          protocol. The STUN client sends a request to a STUN server and the
          server returns a response. There are two types of STUN messages -
          Binding Requests and Indications. Binding requests are used when
          determining a client's external address and solicits a response from
          the STUN server with the seen address.</t>

          <t>The first version of <xref target="RFC3489">STUN</xref> included
          categorization and parameterization of NATs. This was abandoned in
          the <xref target="RFC5389">updated version</xref> due to it being
          unreliable and brittle. Some of the below discussed methods are
          based on RFC3489 functionality which will be called out and the
          downside of that will be part of the characterization.</t>
        </section>

        <section title="Using STUN to traverse NAT without server modifications">
          <t>This section describes how a client can use STUN to traverse NATs
          to RTSP servers without requiring server modifications. Note that
          this method has limited applicability and requires the server to be
          available in the external/public address realm in regards to the
          client located behind a NAT(s).</t>

          <t>Limitations:</t>

          <t><list style="symbols">
              <t>The server must be located in either a public address realm
              or the next hop external address realm in regards to the
              client.</t>

              <t>The client may only be located behind NATs that perform
              "Endpoint-Independent" or "Address-Dependent" Mappings. Clients
              behind NATs that do "Address and Port-Dependent" Mappings cannot
              use this method. See <xref target="RFC4787"/> for full
              definition of these terms.</t>

              <t>Based on the discontinued middlebox classification of the
              <xref target="RFC3489">replaced STUN specification</xref>. Thus
              brittle and unreliable.</t>
            </list>Method:</t>

          <t>A RTSP client using RTP transport over UDP can use STUN to
          traverse a NAT(s) in the following way:</t>

          <t><list style="numbers">
              <t>Use STUN to try to discover the type of NAT, and the timeout
              period for any UDP mapping on the NAT. This is recommend to be
              performed in the background as soon as IP connectivity is
              established. If this is performed prior to establishing a
              streaming session the delays in the session establishment will
              be reduced. If no NAT is detected, normal SETUP should be
              used.</t>

              <t>The RTSP client determines the number of UDP ports needed by
              counting the number of needed media transport protocols sessions
              in the multi-media presentation. This information is available
              in the media description protocol, e.g. SDP <xref
              target="RFC4566"/>. For example, each RTP session will in
              general require two UDP ports, one for RTP, and one for
              RTCP.</t>

              <t>For each UDP port required, establish a mapping and discover
              the public/external IP address and port number with the help of
              the STUN server. A successful mapping looks like: client's local
              address/port <-> public address/port.</t>

              <t>Perform the RTSP SETUP for each media. In the transport
              header the following parameter should be included with the given
              values: <xref
              target="I-D.ietf-mmusic-rfc2326bis">"dest_addr"</xref> or
              "destination" + <xref target="RFC2326">"client_port"</xref> with
              the public/external IP address and port pair for both RTP and
              RTCP. To be certain that this works servers must allow a client
              to setup the RTP stream on any port, not only even ports and
              with non-contiguous port numbers for RTP and RTCP. This requires
              the new feature provided in the <xref
              target="I-D.ietf-mmusic-rfc2326bis">update to RFC2326</xref>.
              The server should respond with a transport header containing an
              "src_addr" or "source" + "server_port" parameters with the RTP
              and RTCP source IP address and port of the media stream.</t>

              <t>To keep the mappings alive, the client should periodically
              send UDP traffic over all mappings needed for the session. For
              the mapping carrying RTCP traffic the periodic RTCP traffic are
              likely enough. For mappings carrying RTP traffic and for
              mappings carrying RTCP packets at too low a frequency,
              keep-alive messages should be sent. As keep alive messages, one
              could use the <xref target="I-D.ietf-avt-rtp-no-op">RTP No-Op
              packet</xref> to the streaming server's discard port (port
              number 9). The drawback of using RTP No-Op is that the payload
              type number must be dynamically assigned through RTSP first.
              Otherwise STUN could be used for the keep-alive as well as empty
              UDP packets.</t>
            </list>If a UDP mapping is lost, the above discovery process must
          be repeated. The media stream also needs to be SETUP again to change
          the transport parameters to the new ones. This will cause a glitch
          in media playback.</t>

          <t>To allow UDP packets to arrive from the server to a client behind
          a "Address Dependent" filtering NAT, the client must first send a
          UDP packet to establish filtering state in the NAT. The client,
          before sending a RTSP PLAY request, must send a so called
          hole-punching packet (such as a RTP No-Op packet) on each mapping,
          to the IP address given as the servers source address. To create
          minimum problems for the server these UDP packets should be sent to
          the server's discard port (port number 9). Since UDP packets are
          inherently unreliable, to ensure that at least one UDP message
          passes the NAT, hole-punching packets should be retransmitted a
          reasonable number of times.</t>

          <t>For an "Address and Port Dependent" filtering NAT the client must
          send messages to the exact ports used by the server to send UDP
          packets before sending a RTSP PLAY request. This makes it possible
          to use the above described process with the following additional
          restrictions: for each port mapping, hole-punching packets need to
          be sent first to the server's source address/port. To minimize
          potential effects on the server from these messages the following
          type of hole punching packets must be sent. RTP: an empty or less
          than 12 bytes UDP packet. RTCP: A correctly formatted RTCP RR or SR
          message. The above described adaptations for restricted NATs will
          not work unless the server includes the "src_addr" in the
          "Transport" header (which is the "source" transport parameter in
          RFC2326).</t>

          <t>This method is brittle because it assumes one can use STUN to
          classify the NAT behavior, which was found to be <xref
          target="RFC5389">problematic</xref>. If the NAT changes the
          properties of the existing mapping and filtering state for example
          due to load, then the methods will fail.</t>
        </section>

        <section anchor="sec-stun-alg" title="ALG considerations">
          <t>If a NAT supports RTSP ALG (Application Level Gateway) and is not
          aware of the STUN traversal option, service failure may happen,
          because a client discovers its public IP address and port numbers,
          and inserts them in its SETUP requests. When the RTSP ALG processes
          the SETUP request it may change the destination and port number,
          resulting in unpredictable behavior. An ALG should not update
          address fields which contains addresses other than the NATs internal
          address domain. In cases where the ALG modifies fields unnecessarily
          two alternatives exist:<list style="numbers">
              <t>Use TLS to encrypt the RTSP TCP connection to prevent the ALG
              from reading and modifying the RTSP messages.</t>

              <t>Turn off the STUN based NAT traversal mechanism</t>
            </list>As it may be difficult to determine why the failure occurs,
          the usage of TLS protected RTSP message exchange at all times would
          avoid this issue.</t>
        </section>

        <section title="Deployment Considerations">
          <t>For the Stand-Alone usage of STUN the following applies:</t>

          <t>Advantages:</t>

          <t><list style="symbols">
              <t>STUN is a solution first used by SIP applications. As shown
              above, with little or no changes, the RTSP application can
              re-use STUN as a NAT traversal solution, avoiding the pit-fall
              of solving a problem twice.</t>

              <t>Using STUN does not require RTSP server modifications; it
              only affects the client implementation.</t>
            </list>Disadvantages:</t>

          <t><list style="symbols">
              <t>Requires a STUN server deployed in the public address
              space.</t>

              <t>Only works with NATs that perform endpoint independent and
              address dependent mappings. Address and Port-Dependent filtering
              NATs create some issues.</t>

              <t>Brittle to NATs changing the properties of the NAT mapping
              and filtering.</t>

              <t>Does not work with port and address dependent mapping NATs
              without server modifications.</t>

              <t>Will mostly not work if a NAT uses multiple IP addresses,
              since RTSP servers generally require all media streams to use
              the same IP as used in the RTSP connection to prevent becoming a
              DDoS tool.</t>

              <t>Interaction problems exist when a RTSP-aware ALG interferes
              with the use of STUN for NAT traversal unless TLS secured RTSP
              message exchange is used.</t>

              <t>Using STUN requires that RTSP servers and clients support the
              <xref target="I-D.ietf-mmusic-rfc2326bis">updated RTSP
              specification</xref>, because it is no longer possible to
              guarantee that RTP and RTCP ports are adjacent to each other, as
              required by the "client_port" and "server_port" parameters in
              RFC2326.</t>
            </list>Transition:</t>

          <t>The usage of STUN can be phased out gradually as the first step
          of a STUN capable server or client should be to check the presence
          of NATs. The removal of STUN capability in the client
          implementations will have to wait until there is absolutely no need
          to use STUN.</t>
        </section>

        <section anchor="sec-stun-sec" title="Security Considerations">
          <t>To prevent the RTSP server from being used as Denial of Service
          (DoS) attack tools the RTSP Transport header parameter "destination"
          and "dest_addr" are generally not allowed to point to any IP address
          other than the one the RTSP message originates from. The RTSP server
          is only prepared to make an exception to this rule when the client
          is trusted (e.g., through the use of a secure authentication
          process, or through some secure method of challenging the
          destination to verify its willingness to accept the RTP traffic).
          Such a restriction means that STUN in general does not work for use
          cases where RTSP and media transport go to different addresses.</t>

          <t>STUN combined with destination address restricted RTSP has the
          same security properties as the core RTSP. It is protected from
          being used as a DoS attack tool unless the attacker has the ability
          to spoof the TCP connection carrying RTSP messages.</t>

          <t>Using STUN's support for message authentication and secure
          transport of RTSP messages, attackers cannot modify STUN responses
          or RTSP messages (TLS) to change media destination. This protects
          against hijacking, however as a client can be the initiator of an
          attack, these mechanisms cannot securely prevent RTSP servers being
          used as DoS attack tools.</t>
        </section>
      </section>

      <section title="Server Embedded STUN">
        <t/>

        <section title="Introduction">
          <t>This Section describes an alternative to the stand-alone STUN
          usage in the previous section that has quite significantly different
          behavior.</t>
        </section>

        <section title="Embedding STUN in RTSP">
          <t>This section outlines the adaptation and embedding of STUN within
          RTSP. This enables STUN to be used to traverse any type of NAT,
          including address and Port-Dependent mapping NATs. This would
          require RTSP level protocol changes.</t>

          <t>This NAT traversal solution has limitations:</t>

          <t><list style="numbers">
              <t>It does not work if both RTSP client and RTSP server are
              behind separate NATs.</t>

              <t>The RTSP server may, for security reasons, refuse to send
              media streams to an IP different from the IP in the client RTSP
              requests.</t>
            </list></t>

          <t>Deviations from STUN as defined in RFC 5389:</t>

          <t><list style="numbers">
              <t>The RTSP application must provision the client with an
              identity and shared secret to use in the STUN
              authentication;</t>

              <t>We require STUN server to be co-located on RTSP server's
              media source ports.</t>
            </list></t>

          <t>If STUN server is co-located with RTSP server's media source
          port, an RTSP client using RTP transport over UDP can use STUN to
          traverse ALL types of NATs. In the case of port and address
          dependent mapping NATs, the party on the inside of the NAT must
          initiate UDP traffic. The STUN Binding Request, being a UDP packet
          itself, can serve as the traffic initiating packet. Subsequently,
          both the STUN Binding Response packets and the RTP/RTCP packets can
          traverse the NAT, regardless of whether the RTSP server or the RTSP
          client is behind NAT (however only one of the can be behind a
          NAT).</t>

          <t>Likewise, if an RTSP server is behind a NAT, then an embedded
          STUN server must be co-located on the RTSP client's RTCP port. Also
          it will become the client that needs to disclose his destination
          address rather than the server, so the server can correctly
          determine its NAT external source address for the media streams. In
          this case, we assume that the client has some means of establishing
          TCP connection to the RTSP server behind NAT so as to exchange RTSP
          messages with the RTSP server, potentially using a proxy or static
          rules.</t>

          <t>To minimize delay, we require that the RTSP server supporting
          this option must inform the client about the RTP and RTCP ports from
          where the server will send out RTP and RTCP packets, respectively.
          This can be done by using the "server_port" parameter in RFC2326,
          and the "src_addr" parameter in <xref
          target="I-D.ietf-mmusic-rfc2326bis"/>. Both are in the RTSP
          Transport header. But in general this strategy will require that one
          first do one SETUP request per media to learn the server ports, then
          perform the STUN checks, followed by a subsequent SETUP to change
          the client port and destination address to what was learned during
          the STUN checks.</t>

          <t>To be certain that RTCP works correctly the RTSP end-point
          (server or client) will be required to send and receive RTCP packets
          from the same port.</t>
        </section>

        <section title="Discussion On Co-located STUN Server">
          <t>In order to use STUN to traverse "address and port dependent"
          filtering or mapping NATs the STUN server needs to be co-located
          with the streaming server media output ports. This creates a
          de-multiplexing problem: we must be able to differentiate a STUN
          packet from a media packet. This will be done based on heuristics.
          The existing STUN heuristics is the first byte in the packet and the
          Magic Cookie field (added in RFC5389), which works fine between STUN
          and RTP or RTCP where the first byte happens to be different. Thanks
          to the magic cookie field it is unlikely that other protocols would
          be mistaken for a STUN packet, but not assured.</t>
        </section>

        <section title="ALG considerations">
          <t>The same ALG traversal considerations as for <xref
          target="sec-stun-alg">Stand-Alone STUN applies</xref>.</t>
        </section>

        <section title="Deployment Considerations">
          <t>For the "Embedded STUN" method the following applies:</t>

          <t>Advantages:</t>

          <t><list style="symbols">
              <t>STUN is a solution first used by SIP applications. As shown
              above, with little or no changes, RTSP application can re-use
              STUN as a NAT traversal solution, avoiding the pit-fall of
              solving a problem twice.</t>

              <t>STUN has built-in message authentication features, which
              makes it more secure against hi-jacking attacks. See next
              section for an in-depth security discussion.</t>

              <t>This solution works as long as there is only one RTSP
              endpoint in the private address realm, regardless of the NAT's
              type. There may even be multiple NATs (see Figure 1 in <xref
              target="RFC5389"/>).</t>

              <t>Compared to other UDP based NAT traversal methods in this
              document, STUN requires little new protocol development (since
              STUN is already a IETF standard), and most likely less
              implementation effort, since open source STUN server and client
              implementations are available <xref target="STUN-IMPL"/><xref
              target="PJNATH"/>. There is the need to embed STUN in RTSP
              server and client, which require a de-multiplexer between STUN
              packets and RTP/RTCP packets. There is also a need to register
              the proper feature tags.</t>
            </list>Disadvantages:</t>

          <t><list style="symbols">
              <t>Some extensions to the RTSP core protocol, likely signaled by
              RTSP feature tags, must be introduced.</t>

              <t>Requires an embedded STUN server to be co-located on each of
              the RTSP server's media protocol's ports (e.g. RTP and RTCP
              ports), which means more processing is required to de-multiplex
              STUN packets from media packets. For example, the de-multiplexer
              must be able to differentiate a RTCP RR packet from a STUN
              packet, and forward the former to the streaming server, and the
              latter to the STUN server.</t>

              <t>Does not support use cases that require the RTSP connection
              and the media reception to happen at different addresses, unless
              the server's security policy is relaxed.</t>

              <t>Interaction problems exist when a RTSP ALG is not aware of
              STUN unless TLS is used to protect the RTSP messages.</t>

              <t>Using STUN requires that RTSP servers and clients support the
              <xref target="I-D.ietf-mmusic-rfc2326bis">updated RTSP
              specification</xref>, and they both agree to support the NAT
              traversal feature.</t>

              <t>Increases the setup delay with at least the amount of time it
              takes to perform STUN message exchanges. Most likely an extra
              SETUP sequence will be required.</t>
            </list>Transition:</t>

          <t>The usage of STUN can be phased out gradually as the first step
          of a STUN capable machine can be to check the presence of NATs for
          the presently used network connection. The removal of STUN
          capability in the client implementations will have to wait until
          there is absolutely no need to use STUN.</t>
        </section>

        <section title="Security Considerations">
          <t>See <xref target="sec-stun-sec">Stand-Alone STUN</xref>.</t>
        </section>
      </section>

      <section anchor="sec-ice" title="ICE">
        <section title="Introduction">
          <t><xref target="RFC5245">ICE (Interactive Connectivity
          Establishment)</xref> is a methodology for NAT traversal that has
          been developed for SIP using SDP offer/answer. The basic idea is to
          try, in a staggered parallel fashion, all possible connection
          addresses that an endpoint may be reachable by. This allows the
          endpoint to use the best available UDP "connection" (meaning two UDP
          end-points capable of reaching each other). The methodology has very
          nice properties in that basically all NAT topologies are possible to
          traverse.</t>

          <t>Here is how ICE works at a high level. End point A collects all
          possible addresses that can be used, including local IP addresses,
          STUN derived addresses, TURN addresses, etc. On each local port that
          any of these address and port pairs lead to, a STUN server is
          installed. This STUN server only accepts STUN requests using the
          correct authentication through the use of a username and
          password.</t>

          <t>End-point A then sends a request to establish connectivity with
          end-point B, which includes all possible <xref
          target="RFC5245">"destinations"</xref> to get the media through to
          A. Note that each of A's local address/port pairs (host candidates
          and server reflexive base) has a STUN server co-located. B in turn
          provides A with all its possible destinations for the different
          media streams. A and B then uses a STUN client to try to reach all
          the address and port pairs specified by A from its corresponding
          destination ports. The destinations for which the STUN requests
          successfully complete are then indicated and one is selected.</t>

          <t>If B fails to get any STUN response from A, all hope is not lost.
          Certain NAT topologies require multiple tries from both ends before
          successful connectivity is accomplished and therefore requests are
          retransmitted multiple times. The STUN requests may also result in
          that more connectivity alternatives (destinations) are discovered
          and conveyed in the STUN responses.</t>
        </section>

        <section title="Using ICE in RTSP">
          <t>The usage of ICE for RTSP requires that both client and server be
          updated to include the ICE functionality. If both parties implement
          the necessary functionality the following steps could provide ICE
          support for RTSP.</t>

          <t>This assumes that it is possible to establish a TCP connection
          for the RTSP messages between the client and the server. This is not
          trivial in scenarios where the server is located behind a NAT, and
          may require some TCP ports be opened, or the deployment of proxies,
          etc.</t>

          <t>The negotiation of ICE in RTSP of necessity will work different
          than in SIP with SDP offer/answer. The protocol interactions are
          different and thus the possibilities for transfer of states are also
          somewhat different. The goal is also to avoid introducing extra
          delay in the setup process at least for when the server is using a
          public address and the client is either having a public address or
          is behind NAT(s). This process is only intended to support PLAY
          mode, i.e. media traffic flows from server to client.</t>

          <t><list style="numbers">
              <t>The ICE usage begins in the SDP. The SDP for the service
              indicates that ICE is supported at the server. No candidates can
              be given here as that would not work with the on demand, DNS
              load balancing, etc., that have the SDP indicate a resource on a
              server park rather than a specific machine.</t>

              <t>The client gathers addresses and puts together its candidates
              for each media stream indicated in the session description.</t>

              <t>In each SETUP request the client includes its candidates in
              an ICE specific transport specification. This indicates for the
              server the ICE support by the client. One candidate is the most
              prioritized candidate and here the prioritization for this
              address should be somewhat different compared to SIP. High
              performance rather than always successful is recommended, as it
              is most likely to be a server in the public.</t>

              <t>The server responds to the SETUP (200 OK) for each media
              stream with its candidates. A server with a public address
              usually only provides a single ICE candidate. Also here one
              candidate is the server primary address.</t>

              <t>The connectivity checks are performed. For the server the
              connectivity checks from the server to the clients have an
              additional usage. They verify that there is someone willingly to
              receive the media, thus protecting itself from performing
              unknowingly an DoS attack.</t>

              <t>Connectivity checks from the client promoting a candidate
              pair were successful. Thus no further SETUP requests are
              necessary and processing can proceed with step 7. If another
              address than the primary has been verified by the client to
              work, that address may then be promoted for usage in a SETUP
              request (Go to 7). If the checks for the available candidates
              failed and if further candidates have been derived during the
              connectivity checks, then those can be signalled in new
              candidate lines in a SETUP request updating the list (Go to
              5).</t>

              <t>Client issues PLAY request. If the server also has completed
              its connectivity checks for the promoted candidate pair (based
              on username as it may be derived addresses if the client was
              behind NAT) then it can directly answer 200 OK (Go to 8). If the
              connectivity check has not yet completed it responds with a 1xx
              code to indicate that it is verifying the connectivity. If that
              fails within the set timeout, an error is reported back. Client
              needs to go back to 6.</t>

              <t>Process completed and media can be delivered. ICE candidates
              not used may be released.</t>
            </list></t>

          <t>To keep media paths alive the client needs to periodically send
          data to the server. This will be realized with STUN. RTCP sent by
          client should be able to keep RTCP open but STUN will also be used
          based on the same motivations as for ICE for SIP.</t>
        </section>

        <section title="Implementation burden of ICE">
          <t>The usage of ICE will require that a number of new protocols and
          new RTSP/SDP features be implemented. This makes ICE the solution
          that has the largest impact on client and server implementations
          amongst all the NAT/Firewall traversal methods in this document.</t>

          <t>RTSP server implementation requirements are:</t>

          <t><list style="symbols">
              <t>STUN server features</t>

              <t>limited STUN client features</t>

              <t>SDP generation with more parameters.</t>

              <t>RTSP error code for ICE extension</t>
            </list>RTSP client implementation requirements are:</t>

          <t><list style="symbols">
              <t>Limited STUN server features</t>

              <t>Limited STUN client features</t>

              <t>RTSP error code and ICE extension</t>
            </list></t>
        </section>

        <section title="Deployment Considerations">
          <t>Advantages:</t>

          <t><list style="symbols">
              <t>Solves NAT connectivity discovery for basically all cases as
              long as a TCP connection between them can be established. This
              includes servers behind NATs. (Note that a proxy between address
              domains may be required to get TCP through).</t>

              <t>Improves defenses against DDoS attacks, as media receiving
              client requires authentications, via STUN on its media reception
              ports.</t>
            </list>Disadvantages:</t>

          <t><list style="symbols">
              <t>Increases the setup delay with at least the amount of time it
              takes for the server to perform its STUN requests.</t>

              <t>Assumes that it is possible to de-multiplex between the
              packets of the media protocol and STUN packets.</t>

              <t>Has fairly high implementation burden put on both RTSP server
              and client. However, several Open Source ICE implementations do
              exist, such as <xref target="NICE"/><xref target="PJNATH"/>.</t>
            </list></t>
        </section>

        <section title="Security Consideration">
          <t>One should review the security consideration section of ICE and
          STUN to understand that ICE contains some potential issues. However
          these can be avoided by correctly using ICE in RTSP. In fact ICE
          does help avoid the DDoS attack issue with RTSP substantially as it
          reduces the possibility for a DDoS using RTSP servers to attackers
          that are on-path between the RTSP server and the target and capable
          of intercepting the STUN connectivity check packets and correctly
          send a response to the server.</t>
        </section>
      </section>

      <section anchor="sec-latching" title="Latching">
        <t/>

        <section title="Introduction">
          <t>Latching is a NAT traversal solution that is based on requiring
          RTSP clients to send UDP packets to the server's media output ports.
          Conventionally, RTSP servers send RTP packets in one direction: from
          server to client. Latching is similar to connection-oriented
          traffic, where one side (e.g., the RTSP client) first "connects" by
          sending a RTP packet to the other side's RTP port, the recipient
          then replies to the originating IP and port. This method is also
          referred to as "Late binding". It requires that all RTP/RTCP
          transport is done symmetrical, i.e. <xref target="RFC4961">Symmetric
          RTP</xref>.</t>

          <t>Specifically, when the RTSP server receives the latching packet
          (a.k.a. hole-punching packet, since it is used to punch a hole in
          the Firewall/NAT and to aid the server for port binding and address
          mapping) from its client, it copies the source IP and Port number
          and uses them as delivery address for media packets. By having the
          server send media traffic back the same way as the client's packet
          are sent to the server, address mappings will be honored. Therefore
          this technique works for all types of NATs, given that the server is
          not behind a NAT. However, it does require server modifications.
          Unless there is built-in protection mechanism, latching is very
          vulnerable to DDoS attacks, because attackers can simply forge the
          source IP & Port of the latching packet. Using the rule for
          restricting IP address to the one of the signaling connection will
          need to be applied here also. However, that does not protect against
          hijacking from another client behind the same NAT. This can become a
          serious issue in deployments with CGNs.</t>
        </section>

        <section title="Necessary RTSP extensions">
          <t>To support Latching, the RTSP signaling must be extended to allow
          the RTSP client to indicate that it will use Latching. The client
          also needs to be able to signal its RTP SSRC to the server in its
          SETUP request. The RTP SSRC is used to establish some basic level of
          security against hijacking attacks or simply avoid mis-association
          when multiple clients are behind the same NAT. Care must be taken in
          choosing clients´ RTP SSRC. First, it must be unique within
          all the RTP sessions belonging to the same RTSP session. Secondly,
          if the RTSP server is sending out media packets to multiple clients
          from the same send port, the RTP SSRC needs to be unique amongst
          those clients' RTP sessions. Recognizing that there is a potential
          that RTP SSRC collisions may occur, the RTSP server must be able to
          signal to a client that a collision has occurred and that it wants
          the client to use a different RTP SSRC carried in the SETUP response
          or use unique ports per RTSP session. Using unique ports limits an
          RTSP server in the number of sessions it can simultaneously handle
          per interface IP addresses.</t>
        </section>

        <section title="Deployment Considerations">
          <t>Advantages:</t>

          <t><list style="symbols">
              <t>Works for all types of client-facing NATs. (Requirement 1 in
              <xref target="req-section"/>).</t>

              <t>Has no interaction problems with any RTSP ALG changing the
              client's information in the transport header.</t>
            </list>Disadvantages:</t>

          <t><list style="symbols">
              <t>Requires modifications to both RTSP server and client.</t>

              <t>Limited to work with servers that have an public IP
              address.</t>

              <t>The format of the RTP packet for "connection setup" (a.k.a
              Latching packet) is yet to be defined. One possibility is to use
              RTP No-Op packet format in <xref
              target="I-D.ietf-avt-rtp-no-op"/>.</t>

              <t>SSRC management if RTP is used for latching due to risk for
              mis-association of clients to RTSP sessions at the server if
              SSRC collision occurs.</t>

              <t>Has worse security situation than STUN due to lack of STUN
              message authentication and will need to use address
              restrictions.</t>
            </list></t>
        </section>

        <section title="Security Consideration">
          <t>Latching's major security issue is that RTP streams can be
          hijacked and directed towards any target that the attacker desires
          unless address restrictions are used. In the case of NATs with
          multiple clients on the inside of them, hijacking can still occur.
          This becomes a significant threat in the context of carrier grade
          NATs (CGN).</t>

          <t>The most serious security problem is the deliberate attack with
          the use of a RTSP client and Latching. The attacker uses RTSP to
          setup a media session. Then it uses Latching with a spoofed source
          address of the intended target of the attack. There is no defense
          against this attack other than restricting the possible address a
          latching packet can come from to the same as the RTSP TCP connection
          are from. This prevents Latching to be used in use cases that
          require different addresses for media destination and
          signalling.</t>

          <t>A hijack attack can also be performed in various ways. The basic
          attack is based on the ability to read the RTSP signaling packets in
          order to learn the address and port the server will send from and
          also the SSRC the client will use. Having this information the
          attacker can send its own Latching packets containing the correct
          RTP SSRC to the correct address and port on the server. The RTSP
          server will then use the source IP and port from the Latching packet
          as the destination for the media packets it sends.</t>

          <t>Another variation of this attack is for a man in the middle to
          modify the RTP latching packet being sent by a client to the server
          by simply changing the source IP to the target one desires to
          attack.</t>

          <t>One can fend off the first attack by applying encryption to the
          RTSP signaling transport. However, the second variation is
          impossible to defend against. As a NAT re-writes the source IP and
          port this cannot be authenticated, but authentication is required in
          order to protect against this type of DoS attack.</t>

          <t>Yet another issues is that these attacks also can be used to deny
          the client the service it desires from the RTSP server completely.
          For a man in the middle capable of reading the signaling traffic or
          intercepting the latching packets can completely deny the client
          service by modifying or originating latching packets of itself.</t>

          <t>The amount of random non-guessable material in the latching
          packet determines how well Latching can fend off stream-hijacking
          performed by parties that are not "man-in-the-middle". This proposal
          uses the 32-bit RTP SSRC field to this effect. Therefore it is
          important that this field is derived with a non-predictable random
          number generator. It should not be possible by knowing the algorithm
          used and a couple of basic facts, to derive what random number a
          certain client will use.</t>

          <t>An attacker not knowing the SSRC but aware of which port numbers
          that a server sends from can deploy a brute force attack on the
          server by testing a lot of different SSRCs until it finds a matching
          one. Therefore a server could implement functionality that blocks
          packets to ports or from sources that receive or send multiple
          Latching packets with different invalid SSRCs, especially when they
          are coming from the same IP/Port. Note that this mitigation in
          itself opens up a new venue for DoS attacks against legit users
          trying to latch.</t>

          <t>To improve the security against attackers the amount of random
          material could be increased. To achieve a longer random tag while
          still using RTP and RTCP, it will be necessary to develop RTP and
          RTCP payload formats for carrying the random material.</t>
        </section>
      </section>

      <section anchor="sym-rtp-var" title="A Variation to Latching">
        <section title="Introduction">
          <t>Latching as described above requires the usage of a valid RTP
          format as the Latching packet, i.e. the first packet that the client
          sends to the server to set up virtual RTP connection. There existed
          no appropriate RTP packet format for this purpose, although the
          No-Op format was a proposal to fix the problem <xref
          target="I-D.ietf-avt-rtp-no-op"/>. However, that work was abandoned.
          There exists a RFC that discusses the implication of different type
          of packets as keep-alives for RTP <xref target="RFC6263"/> and its
          findings are very relevant to the format of the Latching packet.</t>

          <t>Meanwhile, there has been NAT/Firewall traversal techniques
          deployed in the wireless streaming market place that use non-RTP
          messages as Latching packets. This section describes a variant based
          on a subset of those solutions that alters the previously described
          Latching solution.</t>
        </section>

        <section title="Necessary RTSP extensions">
          <t>In this variation of Latching, the Latching packet is a small UDP
          packet that does not contain an RTP header. In response to client's
          Latching packet, the RTSP server sends back a similar Latching
          packet as a confirmation so the client can stop the so called
          "connection phase" of this NAT traversal technique. Afterwards, the
          client only has to periodically send Latching packets as keep-alive
          messages for the NAT mappings.</t>

          <t>The server listens on its RTP-media output port, and tries to
          decode any received UDP packet as Latching packet. This is valid
          since an RTSP server is not expecting RTP traffic from the RTSP
          client. Then, it can correlate the Latching packet with the RTSP
          client's session ID or the client's SSRC, and record the NAT
          bindings accordingly. The server then sends a Latching packet as the
          response to the client.</t>

          <t>The Latching packet can contain the SSRC to identify the RTP
          stream, and care must be taken if the packet is bigger than 12
          bytes, ensuring that it is distinctively different from RTP packets,
          whose header size is 12 bytes.</t>

          <t>RTSP signaling can be added to do the following:</t>

          <t><list style="numbers">
              <t>Enable or disable such Latching message exchanges. When the
              Firewall/NAT has an RTSP-aware ALG, it is possible to disable
              Latching message exchange and let the ALG work out the address
              and port mappings.</t>

              <t>Configure the number of re-tries and the re-try interval of
              the Latching message exchanges.</t>
            </list></t>
        </section>

        <section title="Deployment Considerations">
          <t>This approach has the following advantages when compared with the
          <xref target="sec-latching">Latching approach</xref>:</t>

          <t><list style="numbers">
              <t>There is no need to define RTP payload format for Firewall
              traversal, therefore it is simple to use, implement and
              administer (Requirement 4 in <xref target="req-section"/>),
              instead a Latching protocol must be defined.</t>

              <t>When properly defined, this kind of Latching packet exchange
              can also authenticate RTP receivers, to prevent hijacking
              attacks.</t>
            </list>This approach has the following disadvantages when compared
          with the Latching approach:</t>

          <t><list style="numbers">
              <t>RTP traffic is normally accompanied by RTCP traffic. This
              approach needs to rely on RTCP RRs and SRs to enable NAT
              traversal for RTCP endpoints, use <xref
              target="RFC5761">RTP/RTCP Multiplexing</xref>, or use the same
              type of Latching packets also for RTCP endpoints.</t>

              <t>The server's sender SSRC for the RTP stream or other session
              Identity information must be signaled in RTSP's SETUP response,
              in the Transport header of the RTSP SETUP response.</t>
            </list></t>
        </section>

        <section title="Security Considerations">
          <t>Compared to the security properties of Latching this variant is
          slightly improved. First of all it allows for a larger random field
          in the Latching packets which makes it more unlikelier for an
          off-path attacker to succeed in a hi-jack attack. Secondly the
          confirmation allows the client to know when Latching works and when
          it didn't and thus restart the Latching process by updating the
          SSRC. Thirdly if an authentication mechanism are included in the
          latching packet hijacking attacks can be prevented.</t>

          <t>Still the main security issue remain that the RTSP server can't
          know that the source address in the latching packet was coming from
          a RTSP client wanting to receive media and not one that likes to
          direct the media traffic to an DoS target.</t>
        </section>
      </section>

      <section anchor="sec-3way-latch" title="Three Way Latching">
        <t/>

        <section title="Introduction">
          <t>The three way Latching is an attempt to try to resolve the most
          significant security issues for both previously discussed variants
          of Latching. By adding a server request response exchange directly
          after the initial latching the server can verify that the target
          address present in the latching packet is an active listener and
          confirm its desire to establish a media flow.</t>
        </section>

        <section title="Necessary RTSP extensions">
          <t>Uses the same RTSP extensions as the <xref
          target="sym-rtp-var">alternative latching method</xref> uses. The
          extensions for this variant are only in the format and transmission
          of the Latching packets.</t>

          <t>The client to server latching packet is similar to the <xref
          target="sym-rtp-var">Alternative Latching</xref>, i.e. an UDP packet
          with some session identifier and a random value. When the server
          responds to the Latching packet with a Latching confirmation, it
          includes a random value (Nonce) of its own in addition to echoing
          back the one the client sent. Then a third messages is added to the
          exchange. The client acknowledges the reception of the Latching
          confirmation message and echoes back the server's nonce. Thus
          confirming that the Latched address goes to a RTSP client that
          initiated the latching and is actually present at that address. The
          RTSP server will refuse to send any media until the Latching
          Acknowledgement has been received with a valid nonce.</t>
        </section>

        <section title="Deployment Considerations">
          <t>A solution with a 3-way handshake and its own Latching packets
          can be compared with the <xref target="sec-ice">ICE-based
          solution</xref> and have the following differences:</t>

          <t><list style="symbols">
              <t>Only works for servers with public IP addresses compared to
              any type of server</t>

              <t>May be simpler to implement due to the avoidance of the ICE
              prioritization and check-board mechanisms.</t>
            </list>However, a 3-way Latching protocol is very similar to using
          STUN in both directions as Latching and verification protocol. Using
          STUN would remove the need for implementing a new protocol.</t>
        </section>
      </section>

      <section title="Application Level Gateways">
        <t/>

        <section title="Introduction">
          <t>An Application Level Gateway (ALG) reads the application level
          messages and performs necessary changes to allow the protocol to
          work through the middle box. However this behavior has some problems
          in regards to RTSP:</t>

          <t><list style="numbers">
              <t>It does not work when the RTSP protocol is used with
              end-to-end security. As the ALG can't inspect and change the
              application level messages the protocol will fail due to the
              middle box.</t>

              <t>ALGs need to be updated if extensions to the protocol are
              added. Due to deployment issues with changing ALGs this may also
              break the end-to-end functionality of RTSP.</t>
            </list>Due to the above reasons it is not recommended to use an
          RTSP ALG in NATs. This is especially important for NATs targeted to
          home users and small office environments, since it is very hard to
          upgrade NATs deployed in home or SOHO (small office/home office)
          environment.</t>
        </section>

        <section title="Outline On how ALGs for RTSP work">
          <t>In this section, we provide a step-by-step outline on how one
          could go about writing an ALG to enable RTSP to traverse a NAT.</t>

          <t><list style="numbers">
              <t>Detect any SETUP request.</t>

              <t>Try to detect the usage of any of the NAT traversal methods
              that replace the address and port of the Transport header
              parameters "destination" or "dest_addr". If any of these methods
              are used, then the ALG should not change the address. Ways to
              detect that these methods are used are:<list style="symbols">
                  <t>For embedded STUN, it would be to watch for a feature
                  tag, like "nat.stun". If any of those exists in the
                  "supported", "proxy-require", or "require" headers of the
                  RTSP exchange.</t>

                  <t>For stand alone STUN and TURN based solutions: This can
                  be detected by inspecting the "destination" or "dest_addr"
                  parameter. If it contains either one of the NAT's external
                  IP addresses or a public IP address then such a solution is
                  in use. However if multiple NATs are used this detection may
                  fail. Remapping should only be done for addresses belonging
                  to the NAT's own private address space.</t>
                </list>Otherwise continue to the next step.</t>

              <t>Create UDP mappings (client given IP/port <-> external
              IP/port) where needed for all possible transport specifications
              in the transport header of the request found in (1). Enter the
              external address and port(s) of these mappings in transport
              header. Mappings shall be created with consecutive public port
              numbers starting on an even number for RTP for each media
              stream. Mappings should also be given a long timeout period, at
              least 5 minutes.</t>

              <t>When the SETUP response is received from the server, the ALG
              may remove the unused UDP mappings, i.e. the ones not present in
              the transport header. The session ID should also be bound to the
              UDP mappings part of that session.</t>

              <t>If SETUP response settles on RTP over TCP or RTP over RTSP as
              lower transport, do nothing: let TCP tunneling take care of NAT
              traversal. Otherwise go to next step.</t>

              <t>The ALG should keep the UDP mappings belonging to the RTSP
              session as long as: an RTSP messages with the session's ID has
              been sent in the last timeout interval, or a UDP messages has
              been sent on any of the UDP mappings during the last timeout
              interval.</t>

              <t>The ALG may remove a mapping as soon a TEARDOWN response has
              been received for that media stream.</t>
            </list></t>
        </section>

        <section title="Deployment Considerations">
          <t>Advantage:</t>

          <t><list style="symbols">
              <t>No impact on either client or server</t>

              <t>Can work for any type of NATs</t>
            </list>Disadvantage:</t>

          <t><list style="symbols">
              <t>When deployed they are hard to update to reflect protocol
              modifications and extensions. If not updated they will break the
              functionality.</t>

              <t>When end-to-end security is used, the ALG functionality will
              fail.</t>

              <t>Can interfere with other types of traversal mechanisms, such
              as STUN.</t>
            </list>Transition:</t>

          <t>An RTSP ALG will not be phased out in any automatic way. It must
          be removed, probably through the removal of the NAT it is associated
          with.</t>
        </section>

        <section title="Security Considerations">
          <t>An ALG will not work with deployment of end-to-end RTSP signaling
          security. Therefore deployment of ALG will likely result in clients
          located behind NATs not using end-to-end security.</t>

          <t>The creation of an UDP mapping based on the signalling message
          has some potential security implications. First of all if the RTSP
          client releases its ports and another application are assigned these
          instead it could receive RTP media as long as the mappings exist and
          the RTSP server has failed to be signalled or notice the lack of
          client response.</t>

          <t>An NAT with RTSP ALG that assigns mappings based on SETUP
          requests could potentially become victim of an resource exhaustion
          attack. If an attacker creates a lot of RTSP sessions, even without
          starting media transmission could exhaust the pool of available UDP
          ports on the NAT. Thus only a limited number of UDP mappings should
          be allowed to be created by the RTSP ALG.</t>
        </section>
      </section>

      <section title="TCP Tunneling">
        <t/>

        <section title="Introduction">
          <t>Using a TCP connection that is established from the client to the
          server ensures that the server can send data to the client. The
          connection opened from the private domain ensures that the server
          can send data back to the client. To send data originally intended
          to be transported over UDP requires the TCP connection to support
          some type of framing of the media data packets. Using TCP also
          results in the client having to accept that real-time performance
          can be impacted. TCP's problem of ensuring timely delivery was one
          of the reasons why RTP was developed. Problems that arise with TCP
          are: head-of-line blocking, delay introduced by retransmissions,
          highly varying rate due to the congestion control algorithm. If
          sufficient amount of buffering (several seconds) in the receiving
          client can be tolerated then TCP clearly can work.</t>
        </section>

        <section title="Usage of TCP tunneling in RTSP">
          <t>The RTSP core specification <xref
          target="I-D.ietf-mmusic-rfc2326bis"/> supports interleaving of media
          data on the TCP connection that carries RTSP signaling. See section
          14 in <xref target="I-D.ietf-mmusic-rfc2326bis"/> for how to perform
          this type of TCP tunneling. There also exists another way of
          transporting RTP over TCP defined in Appendix C.2 in <xref
          target="I-D.ietf-mmusic-rfc2326bis"/>. For signaling and rules on
          how to establish the TCP connection in lieu of UDP, see appendix C.2
          in <xref target="I-D.ietf-mmusic-rfc2326bis"/>. This is based on the
          framing of RTP over the TCP connection as described in <xref
          target="RFC4571">RFC 4571</xref>.</t>
        </section>

        <section title="Deployment Considerations">
          <t>Advantage: <list style="symbols">
              <t>Works through all types of NATs where the RTSP server in not
              NATed or at least reachable like it was not.</t>
            </list></t>

          <t>Disadvantage: <list style="symbols">
              <t>Functionality needs to be implemented on both server and
              client.</t>

              <t>Will not always meet multimedia stream's real-time
              requirements.</t>
            </list>Transition:</t>

          <t>The tunneling over RTSP's TCP connection is not planned to be
          phased-out. It is intended to be a fallback mechanism and for usage
          when total media reliability is desired, even at the potential price
          of loss of real-time properties.</t>
        </section>

        <section title="Security Considerations">
          <t>The TCP tunneling of RTP has no known security problems besides
          those already presented in the RTSP specification. It is not
          possible to get any amplification effect for denial of service
          attacks due to TCP's flow control. A possible security
          consideration, when session media data is interleaved with RTSP,
          would be the performance bottleneck when RTSP encryption is applied,
          since all session media data also needs to be encrypted.</t>
        </section>
      </section>

      <section title="TURN (Traversal Using Relay NAT)">
        <t/>

        <section title="Introduction">
          <t><xref target="RFC5766">Traversal Using Relay NAT (TURN)</xref> is
          a protocol for setting up traffic relays that allow clients behind
          NATs and firewalls to receive incoming traffic for both UDP and TCP.
          These relays are controlled and have limited resources. They need to
          be allocated before usage. TURN allows a client to temporarily bind
          an address/port pair on the relay (TURN server) to its local source
          address/port pair, which is used to contact the TURN server. The
          TURN server will then forward packets between the two sides of the
          relay.</t>

          <t>To prevent DoS attacks on either recipient, the packets forwarded
          are restricted to the specific source address. On the client side it
          is restricted to the source setting up the allocation. On the
          external side this is limited to the source address/port pair that
          have been given permission by the TURN client creating the
          allocation. Packets from any other source on this address will be
          discarded.</t>

          <t>Using a TURN server makes it possible for a RTSP client to
          receive media streams from even an unmodified RTSP server. However
          the problem is those RTSP servers most likely restrict media
          destinations to no other IP address than the one the RTSP message
          arrives from. This means that TURN could only be used if the server
          knows and accepts that the IP belongs to a TURN server and the TURN
          server can't be targeted at an unknown address or also the RTSP
          connection is relayed through the same TURN server.</t>
        </section>

        <section anchor="use-turn-rtsp" title="Usage of TURN with RTSP">
          <t>To use a TURN server for NAT traversal, the following steps
          should be performed. <list style="numbers">
              <t>The RTSP client connects with the RTSP server. The client
              retrieves the session description to determine the number of
              media streams. To avoid the issue with having RTSP connection
              and media traffic from different addresses also the TCP
              connection must be done through the same TURN server as the one
              in the next step. This will require the usage of <xref
              target="RFC6062">TURN for TCP</xref>.</t>

              <t>The client establishes the necessary bindings on the TURN
              server. It must choose the local RTP and RTCP ports that it
              desires to receive media packets. TURN supports requesting
              bindings of even port numbers and continuous ranges.</t>

              <t>The RTSP client uses the acquired address and port
              allocations in the RTSP SETUP request using the destination
              header.</t>

              <t>The RTSP Server sends the SETUP reply, which must include the
              transport headers src_addr parameter (source and port in RTSP
              1.0). Note that the server is required to have a mechanism to
              verify that it is allowed to send media traffic to the given
              address.</t>

              <t>The RTSP Client uses the RTSP Server's response to create
              TURN permissions for the server's media traffic.</t>

              <t>The client requests that the server starts playing. The
              server starts sending media packets to the given destination
              address and ports.</t>

              <t>The first media packet arrive at the TURN server on the
              external port; If the packet matches an established permission
              the TURN server forwards the media packet to the RTSP
              client.</t>

              <t>If the client pauses and media is not sent for about 75% of
              the mapping timeout the client should use TURN to refresh the
              bindings.</t>
            </list></t>
        </section>

        <section title="Deployment Considerations">
          <t>Advantages: <list style="symbols">
              <t>Does not require any server modifications given that the
              server includes the src_addr header in the SETUP response.</t>

              <t>Works for any type of NAT as long as the RTSP server has
              public reachable IP address.</t>
            </list></t>

          <t>Disadvantage:<list style="symbols">
              <t>Requires another network element, namely the TURN server.</t>

              <t>A TURN server for RTSP may not scale since the number of
              sessions it must forward is proportional to the number of client
              media sessions.</t>

              <t>TURN server becomes a single point of failure.</t>

              <t>Since TURN forwards media packets, it necessarily introduces
              delay.</t>

              <t>An RTSP ALG may change the necessary destinations parameter.
              This will cause the media traffic to be sent to the wrong
              address.</t>
            </list> Transition:</t>

          <t>TURN is not intended to be phased-out completely, see Section 19
          of <xref target="RFC5766"/>. However the usage of TURN could be
          reduced when the demand for having NAT traversal is reduced.</t>
        </section>

        <section title="Security Considerations">
          <t>The TURN server can become part of a denial of service attack
          towards any victim. To perform this attack the attacker must be able
          to eavesdrop on the packets from the TURN server towards a target
          for the DoS attack. The attacker uses the TURN server to setup a
          RTSP session with media flows going through the TURN server. The
          attacker is in fact creating TURN mappings towards a target by
          spoofing the source address of TURN requests. As the attacker will
          need the address of these mappings he must be able to eavesdrop or
          intercept the TURN responses going from the TURN server to the
          target. Having these addresses, he can set up a RTSP session and
          start delivery of the media. The attacker must be able to create
          these mappings. The attacker in this case may be traced by the TURN
          username in the mapping requests.</t>

          <t>This attack requires that the attacker has access to a user
          account on the TURN server to be able set up the TURN mappings. To
          prevent this attack the RTSP server needs to verify that the
          ultimate target destination accept this media stream. Which would
          require something like ICE's connectivity checks being run between
          the RTSP server and the RTSP client.</t>
        </section>
      </section>
    </section>

    <section title="Firewalls">
      <t>Firewalls exist for the purpose of protecting a network from traffic
      not desired by the firewall owner. Therefore it is a policy decision if
      a firewall will let RTSP and its media streams through or not. RTSP is
      designed to be firewall friendly in that it should be easy to design
      firewall policies to permit passage of RTSP traffic and its media
      streams.</t>

      <t>The firewall will need to allow the media streams associated with a
      RTSP session to pass through it. Therefore the firewall will need an ALG
      that reads RTSP SETUP and TEARDOWN messages. By reading the SETUP
      message the firewall can determine what type of transport and from
      where, the media stream packets will be sent. Commonly there will be the
      need to open UDP ports for RTP/RTCP. By looking at the source and
      destination addresses and ports the opening in the firewall can be
      minimized to the least necessary. The opening in the firewall can be
      closed after a TEARDOWN message for that session or the session itself
      times out.</t>

      <t>Simpler firewalls do allow a client to receive media as long as it
      has sent packets to the target. Depending on the security level this can
      have the same behavior as a NAT. The only difference is that no address
      translation is done. To use such a firewall a client would need to
      implement one of the above described NAT traversal methods that include
      sending packets to the server to open up the mappings.</t>
    </section>

    <section title="Comparison of NAT traversal techniques">
      <t>This section evaluates the techniques described above against the
      requirements listed in <xref target="req-section"/>.</t>

      <t>In the following table, the columns correspond to the numbered
      requirements. For instance, the column under R1 corresponds to the first
      requirement in <xref target="req-section"/>: must work for all flavors
      of NATs. The rows represent the different NAT/Firewall traversal
      techniques. Latch is short for Latching, "V. Latch" is short for
      "variation of Latching" as described in <xref target="sym-rtp-var"/>.
      "3-W Latch" is short for the Three Way Latching described in <xref
      target="sec-3way-latch"/>.</t>

      <t>A Summary of the requirements are:</t>

      <t><list style="hanging">
          <t hangText="R1:">Work for all flavors of NATs</t>

          <t hangText="R2:">Must work with Firewalls, including those with
          ALGs</t>

          <t hangText="R3:">Should have minimal impact on clients not behind
          NATs, counted in minimal number of additional RTTs</t>

          <t hangText="R4:">Should be simple to use, Implement and
          administer.</t>

          <t hangText="R5:">Should provide mitigation against DDoS attacks</t>
        </list></t>

      <t>The following considerations are also added to requirements:<list
          style="hanging">
          <t hangText="C1:">Will solution support both Clients and Servers
          behind NAT</t>

          <t hangText="C2:">Is the solution robust to changing NAT
          behaviors</t>
        </list></t>

      <t/>

      <figure anchor="fig-comp"
              title="Comparison of fulfillment of requirements">
        <preamble/>

        <artwork><![CDATA[------------+------+------+------+------+------+------+------+
            |  R1  |  R2  |  R3  |  R4  |  R5  |  C1  |  C2  |
------------+------+------+------+------+------+------+------+
 STUN       | No   | Yes  |  1   | Maybe| No   | No   | No   |   
------------+------+------+------+------+------+------+------+
 Emb. STUN  | Yes  | Yes  |  2   | Maybe| No   | No   | Yes  |    
------------+------+------+------+------+------+------+------+
 ICE        | Yes  | Yes  | 2.5  | No   | Yes  | Yes  | Yes  |
------------+------+------+------+------+------+------+------+
 Latch      | Yes  | Yes  |  1   | Maybe| No   | No   | Yes  |
------------+------+------+------+------+------+------+------+
 V. Latch   | Yes  | Yes  |  1   | Yes  | No   | No   | Yes  |
------------+------+------+------+------+------+------+------+
 3-W Latch  | Yes  | Yes  | 1.5  | Maybe| Yes  | No   | Yes  |
------------+------+------+------+------+------+------+------+
 ALG        |(Yes) | Yes  |  0   | No   | Yes  | No   | Yes  |
------------+------+------+------+------+------+------+------+
 TCP Tunnel | Yes  | Yes  | 1.5  | Yes  | Yes  | No   | Yes  |
------------+------+------+------+------+------+------+------+
 TURN       | Yes  | Yes  |  1   | No   | Yes  |(Yes) | Yes  |
------------+------+------+------+------+------+------+------+
 ]]></artwork>

        <postamble/>
      </figure>

      <t anchor="Comparision-table">Looking at <xref target="fig-comp"/> one
      would draw the conclusion that using TCP Tunneling or Three-Way Latching
      is the solutions that best fulfill the requirements. The different
      techniques were discussed in the MMUSIC WG. It was established that the
      WG would pursue an ICE based solution due to its generality and
      capability of handling also servers delivering media from behind NATs.
      TCP Tunneling is likely to be available as an alternative, due to its
      specification in the main RTSP specification. Thus it can be used if
      desired and the potential downsides of using TCP is acceptable in
      particular deployments. When it comes to Three-Way Latching it is a very
      competitive technique given that you don't need support for RTSP servers
      behind NATs. There were some discussion in the WG if the increased
      implementation burden of ICE is sufficiently motivated compared to a the
      Three-Way Latching solution for this generality. In the end the authors
      believe that reuse of ICE, the greater flexibility and anyway need to
      deploy a new solution was the decisive factors.</t>

      <t>The ICE based RTSP NAT traversal solution is specified in <xref
      target="I-D.ietf-mmusic-rtsp-nat">"A Network Address Translator (NAT)
      Traversal mechanism for media controlled by Real-Time Streaming Protocol
      (RTSP)"</xref>.</t>
    </section>

    <section anchor="IANA" title="IANA Considerations">
      <t>This document makes no request of IANA.</t>

      <t>Note to RFC Editor: this section may be removed on publication as an
      RFC.</t>
    </section>

    <section anchor="Security" title="Security Considerations">
      <t>In the preceding sections we have discussed security merits of the
      different NAT/Firewall traversal methods for RTSP discussed here. In
      summary, the presence of NAT(s) is a security risk, as a client cannot
      perform source authentication of its IP address. This prevents the
      deployment of any future RTSP extensions providing security against
      hijacking of sessions by a man-in-the-middle.</t>

      <t>Each of the proposed solutions has security implications. Using STUN
      will provide the same level of security as RTSP with out transport level
      security and source authentications, as long as the server does not
      allow media to be sent to a different IP-address than the RTSP client
      request was sent from. Using Latching will have a higher risk of session
      hijacking or denial of service than normal RTSP. The reason is that
      there exists a probability that an attacker is able to guess the random
      bits that the client uses to prove its identity when creating the
      address bindings. This can be solved in the variation of <xref
      target="sym-rtp-var">Latching</xref> with authentication features. Still
      both those variants of Latching is vulnerable against deliberate attack
      from the RTSP client to redirect the media stream requested to any
      target assuming it can spoof the source address. This security
      vulnerability is solved by performing a Three-way Latching procedure as
      discussed in <xref target="sec-3way-latch"/>. The usage of an RTSP ALG
      does not in itself increase the risk for session hijacking. However the
      deployment of ALGs as the sole mechanism for RTSP NAT traversal will
      prevent deployment of end-to-end encrypted RTSP signaling. The usage of
      TCP tunneling has no known security problems. However, it might provide
      a bottleneck when it comes to end-to-end RTSP signaling security if TCP
      tunneling is used on an interleaved RTSP signaling connection. The usage
      of TURN has severe risk of denial of service attacks against a client.
      The TURN server can also be used as a redirect point in a DDoS attack
      unless the server has strict enough rules for who may create
      bindings.</t>
    </section>

    <section anchor="Acknowledgements" title="Acknowledgements">
      <t>The author would also like to thank all persons on the MMUSIC working
      group's mailing list that has commented on this document. Persons having
      contributed in such way in no special order to this protocol are:
      Jonathan Rosenberg, Philippe Gentric, Tom Marshall, David Yon, Amir
      Wolf, Anders Klemets, Flemming Andreasen, Ari Keranen, and Colin
      Perkins. Thomas Zeng would also like to give special thanks to Greg
      Sherwood of PacketVideo for his input into this memo.</t>

      <t><xref target="sec-nat-intro"/> contains text originally written for
      RFC 4787 by Francois Audet and Cullen Jennings.</t>
    </section>
  </middle>

  <back>
    <references title="Informative References">
      <?rfc include='reference.RFC.0768'?>

      <?rfc include='reference.RFC.0793'?>

      <?rfc include='reference.RFC.2326'?>

      <?rfc include='reference.RFC.3550'?>

      <?rfc include='reference.RFC.3489'?>

      <?rfc include='reference.RFC.3022'?>

      <?rfc include='reference.RFC.3424'?>

      <?rfc include='reference.RFC.2588'?>

      <?rfc include='reference.RFC.2663'?>

      <?rfc include='reference.RFC.4787'?>

      <?rfc include='reference.RFC.4566'?>

      <?rfc include='reference.I-D.ietf-mmusic-rfc2326bis'?>

      <?rfc include='reference.RFC.4571'?>

      <?rfc include='reference.RFC.4961'?>

      <?rfc include='reference.I-D.ietf-avt-rtp-no-op'?>

      <?rfc include='reference.RFC.5389'?>

      <?rfc include='reference.RFC.5245'?>

      <?rfc include='reference.RFC.5382'?>

      <?rfc include='reference.RFC.5761'?>

      <?rfc include='reference.RFC.5766'?>

      <?rfc include='reference.RFC.6263'?>

      <?rfc include='reference.RFC.6062'?>

      <?rfc include='reference.I-D.ietf-mmusic-rtsp-nat'?>

      <reference anchor="STUN-IMPL">
        <front>
          <title>Open Source STUN Server and Client,
          http://sourceforge.net/projects/stun/</title>

          <author>
            <organization/>
          </author>

          <date day="23" month="May" year="2013"/>
        </front>
      </reference>

      <reference anchor="NICE">
        <front>
          <title>Libnice - The GLib ICE implementation,
          http://nice.freedesktop.org/wiki/</title>

          <author>
            <organization/>
          </author>

          <date day="23" month="May" year="2013"/>
        </front>
      </reference>

      <reference anchor="PJNATH">
        <front>
          <title>PJNATH - Open Source ICE, STUN, and TURN Library,
          http://www.pjsip.org/pjnath/docs/html/</title>

          <author>
            <organization/>
          </author>

          <date day="23" month="May" year="2013"/>
        </front>
      </reference>
    </references>
  </back>
</rfc>

PAFTECH AB 2003-20262026-04-24 05:13:37