| Network Working Group | M. Westerlund |
| Internet-Draft | Ericsson |
| Intended status: Informational | T. Zeng |
| Expires: October 22, 2015 | April 20, 2015 |
The Comparison of Different Network Address Translator (NAT) Traversal Techniques for Media Controlled by Real-time Streaming Protocol (RTSP)
draft-ietf-mmusic-rtsp-nat-evaluation-15
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 of 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 relate to firewalls and how each technique can be applied in different use cases. These findings were used when selecting the NAT traversal for RTSP 2.0, which is specified in a separate document.
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at http://datatracker.ietf.org/drafts/current/.
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."
This Internet-Draft will expire on October 22, 2015.
Copyright (c) 2015 IETF Trust and the persons identified as the document authors. All rights reserved.
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (http://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Simplified BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Simplified BSD License.
Today there is a proliferating deployment of different types of Network Address Translator (NAT) boxes that in many cases only loosely follow standards [RFC3022][RFC2663][RFC3424][RFC4787][RFC5382]. NATs cause discontinuity in address realms [RFC3424], therefore an application protocol, such as Real-time Streaming Protocol (RTSP) [RFC2326][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 Transmission Control Protocol (TCP) [RFC0793] 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.
Like NATs, firewalls are also middle boxes that need to be considered. Firewalls help 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 limited implementation effort. The effort needed 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 filtering behavior that appear the same to what NATs have. This type of firewall will be successfully traversed using the same solution as employed for NAT traversal, instead of relying on a RTSP ALG. Therefore firewalls will also be discussed and some important differences highlighted.
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 [RFC3424] and quoted below:
"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."
Following the guidelines spelled out in RFC 3424, we describe the required RTSP protocol extensions for each method, transition strategies, and security concerns. The transition strategies are a discussion of how and if the method encourage a move towards not having any NATs on the path.
This document is capturing the evaluation done in the process to recommend firewall/NAT traversal methods for RTSP streaming servers based on RFC 2326 [RFC2326] as well as the RTSP 2.0 core spec [I-D.ietf-mmusic-rfc2326bis]. The evaluation is focused on NAT traversal for the media streams carried over User Datagram Protocol (UDP) [RFC0768] with Real-time Transport Protocol (RTP) [RFC3550] over UDP being the main case for such usage. The findings should be applicable to other protocols as long as they have similar properties.
At the time when the bulk of work on this document was done, a single level of NAT was the dominant deployment for NATs, and multiple level of NATs, including Carrier Grade NATs (CGNs) was not considered. Thus, any characterizations or findings may not be applicable in such scenarios, unless CGN or multiple level of NATs are explicitly noted.
An ICE-based RTSP NAT traversal mechanism is specified in "A Network Address Translator (NAT) Traversal mechanism for media controlled by Real-Time Streaming Protocol (RTSP)" [I-D.ietf-mmusic-rtsp-nat].
We begin by reviewing two quotes from Section 3 in "Network Address Translation (NAT) Behavioral Requirements for Unicast UDP" [RFC4787] concerning NATs and their terminology:
"Readers are urged to refer to [RFC2663] for information on NAT taxonomy and terminology. Traditional NAT is the most common type of NAT device deployed. Readers may refer to [RFC3022] for detailed information on traditional NAT. Traditional NAT has two main varieties -- Basic NAT and Network Address/Port Translator (NAPT).
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 session 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)"."
"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."
In addition to the above quote there exists a number of address and port mapping behaviors described in more detail in Section 4.1 of "Network Address Translation (NAT) Behavioral Requirements for Unicast UDP" [RFC4787] that are highly relevant to the discussion in this document.
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 "Network Address Translation (NAT) Behavioral Requirements for Unicast UDP" [RFC4787] for more information on the different types of filtering that have been identified.
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. Internet. Many organizations use firewalls to prevent intrusions and an malicious attacks on computing resources in the private intranet [RFC2588].
A comparison between NAT and firewall is given below:
It should be noted that many NAT devices intended for Residential or small office/home office (SOHO) use include both NATs and firewall functionality.
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 [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.
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.
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 media data within a few seconds after having received the "200 OK" response to a RTSP PLAY request which starts the media delivery, it may be the result of a middlebox 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 if they are received and not also blocked. 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 mapping for the RTP stream has been removed.
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 (default 60 seconds), 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 the client is mal-functioning.
This section considers the set of requirements for the evaluation of RTSP NAT traversal solutions.
RTSP is a client-server protocol. Typically service providers deploy RTSP servers on the Internet or otherwise reachable address realm. However, there are use cases where the reverse is true: RTSP clients are connecting from any address realm to RTSP servers behind NATs, e.g. in a home. 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 primary issue to solve is the RTSP NAT traversal problem for RTSP servers deployed in a network where the server is on the external side of any NAT, i.e. server is not behind a NAT. The server behind a NAT is desirable, but of much lower priority.
An important consideration for any NAT traversal technique is whether any protocol modification needs occur, where the implementation burden occur, server, client or middlebox. If the incitement to get RTSP to work over a NAT is sufficient to motivate the owner of the server, client or middlebox to update or configure or otherwise perform changes to the device and its software to support the NAT traversal. Thus, the question of who this burden falls on and how big it is is highly relevant.
The list of feature requirements for RTSP NAT solutions are given below:
The last requirement addresses the Distributed Denial-of-Service (DDoS) threat, which relates to NAT traversal as explained below.
During NAT traversal, when the RTSP server determines the media destination (address and port) for the client, the result may be that the IP address of the RTP receiver host is different than the 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 can occur if the attacker can fake the messages in the NAT traversal mechanism to trick the RTSP server into believing that the client's RTP receiver is located on a host to be attacked. 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 mitigation is for the RTSP server to disallow the cases where the client's RTP receiver has a different IP address than that of the RTSP client. This is recommended behavior in RTSP 2.0 unless other solutions to prevent this attack is present, See 21.2.1 in [I-D.ietf-mmusic-rfc2326bis]. With the increased deployment of NAT middleboxes by operators, i.e. carrier grade NAT (CGN), the reuse of an IP address on the NAT's external side by 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.
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.
The survey of traversal techniques was done prior to 2007 and is based on what was available then. This section includes NAT traversal techniques that have not been formally specified anywhere else. 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 techniques. Some techniques used as part of some of the traversal solutions 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. For example RTP No-Op [I-D.ietf-avt-rtp-no-op] was a proposed RTP payload format that failed to be specified, thus it is not available for use today. In each such case, the missing parts will be noted and some basic reasons will be given.
Session Traversal Utilities for NAT (STUN) [RFC5389] 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. Then using the knowledge of these NAT mappings use the external addresses to directly connect to the independent RTSP server. However, this is only possible if the NAT mapping behavior is such that the STUN server and RTSP server will see the same external address and port for the same internal address and port.
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. Indications are used by the client for keep-alive messages towards the server and requires no response from the server.
The first version of STUN [RFC3489] included categorization and parameterization of NATs. This was abandoned in the updated version [RFC5389] due to it being unreliable and brittle. This particular traversal method uses the removed RFC3489 functionality to detect the NAT type to give an early failure indication when the NAT is showing the behavior that this method can't support. This method also suggest using the RTP NO-OP payload format [I-D.ietf-avt-rtp-no-op] for key-alives of the RTP traffic in the client to server direction. This can be replaced with another form of UDP packet as will be further discussed below.
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).
Limitations:
Method:
A RTSP client using RTP transport over UDP can use STUN to traverse a NAT(s) in the following way:
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.
To allow UDP packets to arrive from the server to a client behind a "Address Dependent" or "Address and Port 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 on each mapping, to the IP address and port given as the server's source address and port. For a NAT that only is "Address Dependent" filtering, the hole-punching packet could be sent to the server's discard port (port number 9). For "Address and Port Dependent" filtering NATs the hole-punching packet must go to the port used for sending UDP packets to the client. To be able to do that the server need to include the "src_addr" in the "Transport" header (which is the "source" transport parameter in RFC2326). 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.
As hole-punching and keep-alive messages, one could have used the RTP No-Op packet [I-D.ietf-avt-rtp-no-op] had they been defined. That would have ensured that the traffic would look like RTP and thus likely have the least risk of being dropped by any firewall. The drawback of using RTP No-Op is that the payload type number must be dynamically assigned through RTSP first. Other options are STUN, a RTP packet without any payload, or an UDP packet without any payload. For RTCP it is most suitable to use correctly generated RTCP packets. In general sending unsolicited traffic to the RTSP server may trigger security functions resulting in blocking of the keep-alive messages or termination of the RTSP session itself.
This method is further brittle as it doesn't support address and port dependent mappings. Thus, it proposes to use the old STUN methods to classify the NAT behavior, thus enabling early error indication. This is strictly not required but will lead to failures during setup when the NAT has the wrong behavior. This failure can also occur If the NAT changes the properties of the existing mapping and filtering state or between the classification message exchange and the actual RTSP session setup. for example due to load.
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 NAT external 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:
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.
For the Stand-Alone usage of STUN the following applies:
Advantages:
Disadvantages:
Transition:
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.
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.
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.
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.
This Section describes an alternative to the stand-alone STUN usage in the previous section that has quite significantly different behavior.
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.
This NAT traversal solution has limitations:
Deviations from STUN as defined in RFC 5389:
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).
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.
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 [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.
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.
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. For more discussion of this, please see Section 5.1.2 of [RFC5764].
The same ALG traversal considerations as for Stand-Alone STUN applies [sec-stun-alg].
For the "Embedded STUN" method the following applies:
Advantages:
Disadvantages:
Transition:
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, i.e. no NATs or firewalls.
See Stand-Alone STUN [sec-stun-sec].
ICE (Interactive Connectivity Establishment) [RFC5245] 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.
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.
End-point A then sends a request to establish connectivity with end-point B, which includes all possible "destinations" [RFC5245] 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.
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.
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.
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.
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 not behind a NAT in regards to the client, and the client is either having an address in the same address domain, or is behind NAT(s) which can address the address domain of the server. This process is only intended to support PLAY mode, i.e. media traffic flows from server to client.
To keep media paths alive the client needs to periodically send data to the server. This will be realized with STUN. RTCP sent by the client should be able to keep RTCP open but STUN will also be used based on the same motivations as for ICE for SIP.
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 among all the NAT/firewall traversal methods in this document.
RTSP server implementation requirements are:
RTSP client implementation requirements are:
If there is an RTSP ALG that doesn't support the NAT traversal method, it may interfere with the NAT traversal. As the usage of ICE for the traversal manifest itself in the RTSP message primarily as new transport specification, an ALG that passes through unknown will not prevent the traversal. An ALG that discards unknown specifications will however prevent the NAT traversal. These issues can be avoided by preventing the ALG to interfere with the signalling by using TLS for the RTSP message transport.
An ALG that supports this traversal method, can on the most basic level just pass the transport specifications through. ALGs in NATs and Firewalls could use the ICE candidates to establish filtering state that would allow incoming STUN messages prior to any outgoing hole-punching packets, and in that way speed up the connectivity checks and reduce the risk of failures.
Advantages:
Disadvantages:
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. An important factor is to secure the signalling, i.e. use TLS between RTSP client and server. 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.
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. Symmetric RTP [RFC4961]. There exist a description for latching of SIP negotiated media streams in Session Border Controllers [RFC7362].
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. The format of the latching packet will have to be defined.
Latching is very vulnerable to both hijacking and becoming a tool in Distributed Denial of Service (DDoS) attacks (See Security Considerations of [RFC7362]), 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.
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 among 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.
The latching packet as discussed above should have field which can contain an client and RTP session identifier to correctly associate the latching packet with the correct context. If an RTP packet is to be used, there would have been a benefit to use a well defined RTP payload format for this purpose as the No-Op payload format proposed [I-D.ietf-avt-rtp-no-op]. However, in the absence of such a specification an RTP packet without a payload could be used. Using SSRC has the benefit that RTP and RTCP both would work as is. However, also other packet formats could be used that carry the necessary identification of the context, and such a solution is discussed in Section 4.5.
An RTSP ALG not supporting this method could interfer with the methods used to indicate that latching is to be done, as well as the SSRC signalling. Thus preventing the method from working. However, if the RTSP ALG instead opens the corresponding pinholes and create the necessary mapping in the NAT, traversal will still work. Securing the RTSP message transport using TLS will avoid this issue.
An RTSP ALG that support this traversal method can for basic functionality simply pass the related signalling parameters transparently. Due to the security considerations for latching it might exist a benefit for an RTSP ALG that will enable NAT traversal to negotiate with the path and turn off the latching procedures when the ALG handles this. However, this opens up to failure modes when there are multiple levels of NAT and only one supports an RTSP ALG.
Advantages:
Disadvantages:
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).
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. Even allowing only a limited address range containing the signalling address from where latching is allowed opens up a significant vulnerability as it is difficult to determine the address usage for the network the client connects from.
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.
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 and port to the target one desires to attack.
One can fend off the snooping based attack by applying encryption to the RTSP signaling transport. However, if the attacker is a man in the middle modifying latching packets, the attack is impossible to defend against other than through address restrictions. As a NAT re-writes the source IP and (possibly) port this cannot be authenticated, but authentication is required in order to protect against this type of DoS attack.
Yet another issue is that these attacks also can be used to deny the client the service it desires from the RTSP server completely. The attacker modifies or originates its own latching packets with another port than what the legit latching packets uses, which results in that the media server sends the RTP/RTCP traffic to ports the client isn't listening for RTP/RTCP on.
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 off the client to server network path, i.e. lacks the capability to see the client's latching packets. The proposal above 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.
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.
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.
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 establish a bi-directional transport flow for RTP streams. There is currently no appropriate RTP packet format for this purpose, although the RTP No-Op format was a proposal to fix the problem [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 [RFC6263] and its findings are very relevant to the format of the Latching packet.
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.
In this variation of Latching, the Latching packet is a small UDP packet that does not contain an RTP header. In response to the 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.
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.
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.
RTSP signaling can be added to do the following:
See Latching ALG consideration Section 4.4.3.
This approach has the following advantages when compared with the Latching approach [sec-latching]:
This approach has the following disadvantages when compared with the Latching approach:
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 unlikely 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.
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.
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.
Uses the same RTSP extensions as the alternative latching method [sym-rtp-var] uses. The extensions for this variant are only in the format and transmission of the Latching packets.
The client to server latching packet is similar to the Alternative Latching [sym-rtp-var], 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 message 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.
See Latching ALG consideration Section 4.4.3.
A solution with a 3-way handshake and its own Latching packets can be compared with the ICE-based solution [sec-ice] and have the following differences:
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.
The three way latching is significantly securer than its simpler versions discussed above. The client to server nonce which is included in signalling and also can be bigger than the 32-bits of random data that the SSRC field supports makes it very difficult for an off-path attacker to perform an denial of service attack by diverting the media.
The client to server nonce and its echoing back does not protect against on-patch attacker, including malicious clients. However, the server to client nonce and its echoing back prevents malicious clients to divert the media stream by spoofing the source address and port, as it can't echo back the nonce in these cases.
Three way latching is really only vulnerable to an on-path attacker that is quite capable. First the attacker can either learn the client to server nonce, by intercepting the signalling, or modifying the source information (target destination) of a client's latching packet. Secondly, it is also on-path between the server and target destination and can generate a response using the server's nonce. An adversary that has these capabilities are commonly capable of causing significantly worse damage than this using other methods.
Three-way latching do results in that the server to client packet is bigger than the client to server packet, due to the inclusion of the server to client nonce in addition to the client to server nonce. Thus an amplification effect do exist, however, to achieve this amplification effect the attacker has to create a session state on the RTSP server. The RTSP server can also limit the number of response it will generate before considering the latching to be failed.
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:
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.
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.
Otherwise continue to the next step.
Advantage:
Disadvantage:
Transition:
An RTSP ALG will not be phased out in any automatic way. It must be removed, probably through the removal or update of the NAT it is associated with.
An ALG will not work with deployment of end-to-end RTSP signaling security, however it will work with the hop-by-hop security method defined in Section 19.3 of RTSP 2.0 [I-D.ietf-mmusic-rfc2326bis]. Therefore deployment of ALG may result in clients located behind NATs not using end-to-end security, or more likely the selection a NAT traversal solution that allow for security.
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.
A NAT with RTSP ALG that assigns mappings based on SETUP requests could potentially become victim of a 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.
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 work.
The RTSP core specification [I-D.ietf-mmusic-rfc2326bis] supports interleaving of media data on the TCP connection that carries RTSP signaling. See section 14 in [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 [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 [I-D.ietf-mmusic-rfc2326bis]. This is based on the framing of RTP over the TCP connection as described in RFC 4571 [RFC4571].
An RTSP ALG will face a different issue with TCP tunneling, at least the Interleaved version. Now the full data stream will flow can end up flowing through the ALG implementation. Thus it is important that the ALG is efficient in dealing with the interleaved media data frames to avoid consuming to much resource and thus creating performance issues.
The RTSP ALG can also effect the transport specifications that indicate that TCP tunneling can be done and its priortization, including removing the transport specification, thus preventing TCP tunneling.
Advantage:
Disadvantage:
Transition:
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.
The TCP tunneling of RTP has no known significant security problems besides those already presented in the RTSP specification. It is difficult to get any amplification effect for denial of service attacks due to TCP's flow control. The RTSP server TCP socket, independently if used for media tunneling or only RTSP messages can be used for a redirected syn attack. By spoofing the source address of any TCP init packets, the TCP SYNs from the server can be directed towards a target.
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.
Traversal Using Relay NAT (TURN) [RFC5766] 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.
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.
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. Alternatively, both the RTSP TCP connection as well as the RTP media is relayed through the same TURN server.
To use a TURN server for NAT traversal, the following steps should be performed.
As the RTSP client inserts the address information of the TURN relay's external allocations in the SETUP messages, and ALG that replaces the address, without considering that the address do not belong to the internal address realm of the NAT, will prevent this mechanism from working. This can be prevented by securing the RTSP signalling.
Advantages:
Disadvantage:
Transition:
TURN is not intended to be phased-out completely, see Section 19 of [RFC5766]. However the usage of TURN could be reduced when the demand for having NAT traversal is reduced.
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.
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.
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.
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.
The above possibilities for firewalls to inspect and respond to the signalling are prevented if end-to-end confidentiality protection is used for the RTSP signalling, e.g. using the specified RTSP over TLS. This results in that firewalls can't be actively opening pinholes for the media streams based on the signalling. To enable an RTSP ALG in firewall to correctly function the hop-by-hop signalling security (See Section 19.3) in RTSP 2.0 [I-D.ietf-mmusic-rfc2326bis] can be used. If not, other methods have to be used to enable the transport flows for the media.
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.
This section evaluates the techniques described above against the requirements listed in Section 3.
In the following table, the columns correspond to the numbered requirements. For instance, the column under R1 corresponds to the first requirement in Section 3: 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 Section 4.5. "3-W Latch" is short for the Three Way Latching described in Section 4.6.
A Summary of the requirements are:
The following considerations are also added to requirements:
------------+------+------+------+------+------+------+------+
| 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 |
------------+------+------+------+------+------+------+------+
Figure 1: Comparison of fulfillment of requirements
Looking at Figure 1 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.
The ICE based RTSP NAT traversal solution is specified in "A Network Address Translator (NAT) Traversal mechanism for media controlled by Real-Time Streaming Protocol (RTSP)" [I-D.ietf-mmusic-rtsp-nat].
This document makes no request of IANA.
Note to RFC Editor: this section may be removed on publication as an RFC.
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.
Each of the proposed solutions has security implications. Using STUN will provide the same level of security as RTSP without 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 Latching [sym-rtp-var] with authentication features. Still both those variants of Latching are 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 Section 4.6.
ICE resolves the binding vulnerability of latching by using signed STUN messages, as well as requiring that both sides perform connectivity checks to verify that the target IP address in the candidate pair is both reachable and willing to respond. ICE can however create a significant amount of traffic if the number of candidate pairs are large. Thus pacing is required and implementations should attempt to limit their number of candidates to reduce the number of packets.
If the signalling between the ICE peers (RTSP client and Server) is not confidentiality and integrity protected ICE is vulnerable to attacks where the candidate list is manipulated. Lack of signalling security will also simplify spoofing of STUN binding messages by revealing the secret used in signing.
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.
The latching and variant of latching have so big security issues that they should not be used at all. The three way latching as well as ICE mitigates these security issues and performs the important return-routability check that prevents spoofed source addresses, and should be recommended for that reason. RTP ALG's is a security risk as they can create an incitement against using secure RTSP signalling. That can be avoided as ALGs requires trust in the middlebox, and that trust becomes explicit if one uses the hop-by-hop security solution as specified in Section 19.3 of RTSP 2.0. [I-D.ietf-mmusic-rfc2326bis]. The remaining methods can be considered safe enough, assuming that the appropriate security mechanisms are used and not ignored.
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, Bill Atwood, Alissa Cooper, Colin Perkins, Sarah Banks and David Black. Thomas Zeng would also like to give special thanks to Greg Sherwood of PacketVideo for his input into this memo.
Section 1.1 contains text originally written for RFC 4787 by Francois Audet and Cullen Jennings.