RFC 5245
Interactive Connectivity Establishment (ICE): A Protocol for Network Address Translator (NAT) Traversal for Offer/Answer Protocols
Appendix A. Lite and Full Implementations
ICE allows for two types of implementations. A full implementation supports the controlling and controlled roles in a session, and can also perform address gathering. In contrast, a lite implementation is a minimalist implementation that does little but respond to STUN checks. Because ICE requires both endpoints to support it in order to bring benefits to either endpoint, incremental deployment of ICE in a network is more complicated. Many sessions involve an endpoint that is, by itself, not behind a NAT and not one that would worry about NAT traversal. A very common case is to have one endpoint that requires NAT traversal (such as a VoIP hard phone or soft phone) make a call to one of these devices. Even if the phone supports a full ICE implementation, ICE won't be used at all if the other device doesn't support it. The lite implementation allows for a low-cost entry point for these devices. Once they support the lite implementation, full implementations can connect to them and get the full benefits of ICE. Consequently, a lite implementation is only appropriate for devices that will *always* be connected to the public Internet and have a public IP address at which it can receive packets from any correspondent. ICE will not function when a lite implementation is placed behind a NAT. ICE allows a lite implementation to have a single IPv4 host candidate and several IPv6 addresses. In that case, candidate pairs are selected by the controlling agent using a static algorithm, such as the one in RFC 3484, which is recommended by this specification. However, static mechanisms for address selection are always prone to error, since they cannot ever reflect the actual topology and can never provide actual guarantees on connectivity. They are always heuristics. Consequently, if an agent is implementing ICE just to select between its IPv4 and IPv6 addresses, and none of its IP addresses are behind NAT, usage of full ICE is still RECOMMENDED in order to provide the most robust form of address selection possible. It is important to note that the lite implementation was added to this specification to provide a stepping stone to full implementation. Even for devices that are always connected to the public Internet with just a single IPv4 address, a full implementation is preferable if achievable. A full implementation will reduce call setup times, since ICE's aggressive mode can be used. Full implementations also obtain the security benefits of ICE unrelated to NAT traversal; in particular, the voice hammer attack described in Section 18 is prevented only for full implementations,
not lite. Finally, it is often the case that a device that finds itself with a public address today will be placed in a network tomorrow where it will be behind a NAT. It is difficult to definitively know, over the lifetime of a device or product, that it will always be used on the public Internet. Full implementation provides assurance that communications will always work.Appendix B. Design Motivations
ICE contains a number of normative behaviors that may themselves be simple, but derive from complicated or non-obvious thinking or use cases that merit further discussion. Since these design motivations are not neccesary to understand for purposes of implementation, they are discussed here in an appendix to the specification. This section is non-normative.B.1. Pacing of STUN Transactions
STUN transactions used to gather candidates and to verify connectivity are paced out at an approximate rate of one new transaction every Ta milliseconds. Each transaction, in turn, has a retransmission timer RTO that is a function of Ta as well. Why are these transactions paced, and why are these formulas used? Sending of these STUN requests will often have the effect of creating bindings on NAT devices between the client and the STUN servers. Experience has shown that many NAT devices have upper limits on the rate at which they will create new bindings. Experiments have shown that once every 20 ms is well supported, but not much lower than that. This is why Ta has a lower bound of 20 ms. Furthermore, transmission of these packets on the network makes use of bandwidth and needs to be rate limited by the agent. Deployments based on earlier draft versions of this document tended to overload rate- constrained access links and perform poorly overall, in addition to negatively impacting the network. As a consequence, the pacing ensures that the NAT device does not get overloaded and that traffic is kept at a reasonable rate. The definition of a "reasonable" rate is that STUN should not use more bandwidth than the RTP itself will use, once media starts flowing. The formula for Ta is designed so that, if a STUN packet were sent every Ta seconds, it would consume the same amount of bandwidth as RTP packets, summed across all media streams. Of course, STUN has retransmits, and the desire is to pace those as well. For this reason, RTO is set such that the first retransmit on the first transaction happens just as the first STUN request on the last transaction occurs. Pictorially:
First Packets Retransmits
| |
| |
-------+------ -------+------
/ \ / \
/ \ / \
+--+ +--+ +--+ +--+ +--+ +--+
|A1| |B1| |C1| |A2| |B2| |C2|
+--+ +--+ +--+ +--+ +--+ +--+
---+-------+-------+-------+-------+-------+------------ Time
0 Ta 2Ta 3Ta 4Ta 5Ta
In this picture, there are three transactions that will be sent (for
example, in the case of candidate gathering, there are three host
candidate/STUN server pairs). These are transactions A, B, and C.
The retransmit timer is set so that the first retransmission on the
first transaction (packet A2) is sent at time 3Ta.
Subsequent retransmits after the first will occur even less
frequently than Ta milliseconds apart, since STUN uses an exponential
back-off on its retransmissions.
B.2. Candidates with Multiple Bases
Section 4.1.3 talks about eliminating candidates that have the same
transport address and base. However, candidates with the same
transport addresses but different bases are not redundant. When can
an agent have two candidates that have the same IP address and port,
but different bases? Consider the topology of Figure 10:
+----------+ | STUN Srvr| +----------+ | | ----- // \\ | | | B:net10 | | | \\ // ----- | | +----------+ | NAT | +----------+ | | ----- // \\ | A | |192.168/16 | | | \\ // ----- | | |192.168.1.100 ----- +----------+ // \\ +----------+ | | | | | | | Offerer |---------| C:net10 |-----------| Answerer | | |10.0.1.100| | 10.0.1.101 | | +----------+ \\ // +----------+ ----- Figure 10: Identical Candidates with Different Bases In this case, the offerer is multihomed. It has one IP address, 10.0.1.100, on network C, which is a net 10 private network. The answerer is on this same network. The offerer is also connected to network A, which is 192.168/16. The offerer has an IP address of 192.168.1.100 on this network. There is a NAT on this network, natting into network B, which is another net 10 private network, but not connected to network C. There is a STUN server on network B. The offerer obtains a host candidate on its IP address on network C (10.0.1.100:2498) and a host candidate on its IP address on network A
(192.168.1.100:3344). It performs a STUN query to its configured STUN server from 192.168.1.100:3344. This query passes through the NAT, which happens to assign the binding 10.0.1.100:2498. The STUN server reflects this in the STUN Binding response. Now, the offerer has obtained a server reflexive candidate with a transport address that is identical to a host candidate (10.0.1.100:2498). However, the server reflexive candidate has a base of 192.168.1.100:3344, and the host candidate has a base of 10.0.1.100:2498.B.3. Purpose of the <rel-addr> and <rel-port> Attributes
The candidate attribute contains two values that are not used at all by ICE itself -- <rel-addr> and <rel-port>. Why is it present? There are two motivations for its inclusion. The first is diagnostic. It is very useful to know the relationship between the different types of candidates. By including it, an agent can know which relayed candidate is associated with which reflexive candidate, which in turn is associated with a specific host candidate. When checks for one candidate succeed and not for others, this provides useful diagnostics on what is going on in the network. The second reason has to do with off-path Quality of Service (QoS) mechanisms. When ICE is used in environments such as PacketCable 2.0, proxies will, in addition to performing normal SIP operations, inspect the SDP in SIP messages, and extract the IP address and port for media traffic. They can then interact, through policy servers, with access routers in the network, to establish guaranteed QoS for the media flows. This QoS is provided by classifying the RTP traffic based on 5-tuple, and then providing it a guaranteed rate, or marking its Diffserv codepoints appropriately. When a residential NAT is present, and a relayed candidate gets selected for media, this relayed candidate will be a transport address on an actual TURN server. That address says nothing about the actual transport address in the access router that would be used to classify packets for QoS treatment. Rather, the server reflexive candidate towards the TURN server is needed. By carrying the translation in the SDP, the proxy can use that transport address to request QoS from the access router.B.4. Importance of the STUN Username
ICE requires the usage of message integrity with STUN using its short-term credential functionality. The actual short-term credential is formed by exchanging username fragments in the SDP offer/answer exchange. The need for this mechanism goes beyond just security; it is actually required for correct operation of ICE in the first place.
Consider agents L, R, and Z. L and R are within private enterprise 1, which is using 10.0.0.0/8. Z is within private enterprise 2, which is also using 10.0.0.0/8. As it turns out, R and Z both have IP address 10.0.1.1. L sends an offer to Z. Z, in its answer, provides L with its host candidates. In this case, those candidates are 10.0.1.1:8866 and 10.0.1.1:8877. As it turns out, R is in a session at that same time, and is also using 10.0.1.1:8866 and 10.0.1.1:8877 as host candidates. This means that R is prepared to accept STUN messages on those ports, just as Z is. L will send a STUN request to 10.0.1.1:8866 and another to 10.0.1.1:8877. However, these do not go to Z as expected. Instead, they go to R! If R just replied to them, L would believe it has connectivity to Z, when in fact it has connectivity to a completely different user, R. To fix this, the STUN short-term credential mechanisms are used. The username fragments are sufficiently random that it is highly unlikely that R would be using the same values as Z. Consequently, R would reject the STUN request since the credentials were invalid. In essence, the STUN username fragments provide a form of transient host identifiers, bound to a particular offer/answer session. An unfortunate consequence of the non-uniqueness of IP addresses is that, in the above example, R might not even be an ICE agent. It could be any host, and the port to which the STUN packet is directed could be any ephemeral port on that host. If there is an application listening on this socket for packets, and it is not prepared to handle malformed packets for whatever protocol is in use, the operation of that application could be affected. Fortunately, since the ports exchanged in SDP are ephemeral and usually drawn from the dynamic or registered range, the odds are good that the port is not used to run a server on host R, but rather is the agent side of some protocol. This decreases the probability of hitting an allocated port, due to the transient nature of port usage in this range. However, the possibility of a problem does exist, and network deployers should be prepared for it. Note that this is not a problem specific to ICE; stray packets can arrive at a port at any time for any type of protocol, especially ones on the public Internet. As such, this requirement is just restating a general design guideline for Internet applications -- be prepared for unknown packets on any port.
B.5. The Candidate Pair Priority Formula
The priority for a candidate pair has an odd form. It is: pair priority = 2^32*MIN(G,D) + 2*MAX(G,D) + (G>D?1:0) Why is this? When the candidate pairs are sorted based on this value, the resulting sorting has the MAX/MIN property. This means that the pairs are first sorted based on decreasing value of the minimum of the two priorities. For pairs that have the same value of the minimum priority, the maximum priority is used to sort amongst them. If the max and the min priorities are the same, the controlling agent's priority is used as the tie-breaker in the last part of the expression. The factor of 2*32 is used since the priority of a single candidate is always less than 2*32, resulting in the pair priority being a "concatenation" of the two component priorities. This creates the MAX/MIN sorting. MAX/MIN ensures that, for a particular agent, a lower-priority candidate is never used until all higher-priority candidates have been tried.B.6. The remote-candidates Attribute
The a=remote-candidates attribute exists to eliminate a race condition between the updated offer and the response to the STUN Binding request that moved a candidate into the Valid list. This race condition is shown in Figure 11. On receipt of message 4, agent L adds a candidate pair to the valid list. If there was only a single media stream with a single component, agent L could now send an updated offer. However, the check from agent R has not yet generated a response, and agent R receives the updated offer (message 7) before getting the response (message 9). Thus, it does not yet know that this particular pair is valid. To eliminate this condition, the actual candidates at R that were selected by the offerer (the remote candidates) are included in the offer itself, and the answerer delays its answer until those pairs validate.
Agent A Network Agent B |(1) Offer | | |------------------------------------------>| |(2) Answer | | |<------------------------------------------| |(3) STUN Req. | | |------------------------------------------>| |(4) STUN Res. | | |<------------------------------------------| |(5) STUN Req. | | |<------------------------------------------| |(6) STUN Res. | | |-------------------->| | | |Lost | |(7) Offer | | |------------------------------------------>| |(8) STUN Req. | | |<------------------------------------------| |(9) STUN Res. | | |------------------------------------------>| |(10) Answer | | |<------------------------------------------| Figure 11: Race Condition FlowB.7. Why Are Keepalives Needed?
Once media begins flowing on a candidate pair, it is still necessary to keep the bindings alive at intermediate NATs for the duration of the session. Normally, the media stream packets themselves (e.g., RTP) meet this objective. However, several cases merit further discussion. Firstly, in some RTP usages, such as SIP, the media streams can be "put on hold". This is accomplished by using the SDP "sendonly" or "inactive" attributes, as defined in RFC 3264 [RFC3264]. RFC 3264 directs implementations to cease transmission of media in these cases. However, doing so may cause NAT bindings to timeout, and media won't be able to come off hold. Secondly, some RTP payload formats, such as the payload format for text conversation [RFC4103], may send packets so infrequently that the interval exceeds the NAT binding timeouts. Thirdly, if silence suppression is in use, long periods of silence may cause media transmission to cease sufficiently long for NAT bindings to time out.
For these reasons, the media packets themselves cannot be relied upon. ICE defines a simple periodic keepalive utilizing STUN Binding indications. This makes its bandwidth requirements highly predictable, and thus amenable to QoS reservations.B.8. Why Prefer Peer Reflexive Candidates?
Section 4.1.2 describes procedures for computing the priority of candidate based on its type and local preferences. That section requires that the type preference for peer reflexive candidates always be higher than server reflexive. Why is that? The reason has to do with the security considerations in Section 18. It is much easier for an attacker to cause an agent to use a false server reflexive candidate than it is for an attacker to cause an agent to use a false peer reflexive candidate. Consequently, attacks against address gathering with Binding requests are thwarted by ICE by preferring the peer reflexive candidates.B.9. Why Send an Updated Offer?
Section 11.1 describes rules for sending media. Both agents can send media once ICE checks complete, without waiting for an updated offer. Indeed, the only purpose of the updated offer is to "correct" the SDP so that the default destination for media matches where media is being sent based on ICE procedures (which will be the highest- priority nominated candidate pair). This begs the question -- why is the updated offer/answer exchange needed at all? Indeed, in a pure offer/answer environment, it would not be. The offerer and answerer will agree on the candidates to use through ICE, and then can begin using them. As far as the agents themselves are concerned, the updated offer/answer provides no new information. However, in practice, numerous components along the signaling path look at the SDP information. These include entities performing off-path QoS reservations, NAT traversal components such as ALGs and Session Border Controllers (SBCs), and diagnostic tools that passively monitor the network. For these tools to continue to function without change, the core property of SDP -- that the existing, pre-ICE definitions of the addresses used for media -- the m and c lines and the rtcp attribute -- must be retained. For this reason, an updated offer must be sent.B.10. Why Are Binding Indications Used for Keepalives?
Media keepalives are described in Section 10. These keepalives make use of STUN when both endpoints are ICE capable. However, rather than using a Binding request transaction (which generates a response), the keepalives use an Indication. Why is that?
The primary reason has to do with network QoS mechanisms. Once media begins flowing, network elements will assume that the media stream has a fairly regular structure, making use of periodic packets at fixed intervals, with the possibility of jitter. If an agent is sending media packets, and then receives a Binding request, it would need to generate a response packet along with its media packets. This will increase the actual bandwidth requirements for the 5-tuple carrying the media packets, and introduce jitter in the delivery of those packets. Analysis has shown that this is a concern in certain layer 2 access networks that use fairly tight packet schedulers for media. Additionally, using a Binding Indication allows integrity to be disabled, allowing for better performance. This is useful for large- scale endpoints, such as PSTN gateways and SBCs.B.11. Why Is the Conflict Resolution Mechanism Needed?
When ICE runs between two peers, one agent acts as controlled, and the other as controlling. Rules are defined as a function of implementation type and offerer/answerer to determine who is controlling and who is controlled. However, the specification mentions that, in some cases, both sides might believe they are controlling, or both sides might believe they are controlled. How can this happen? The condition when both agents believe they are controlled shows up in third party call control cases. Consider the following flow: A Controller B |(1) INV() | | |<-------------| | |(2) 200(SDP1) | | |------------->| | | |(3) INV() | | |------------->| | |(4) 200(SDP2) | | |<-------------| |(5) ACK(SDP2) | | |<-------------| | | |(6) ACK(SDP1) | | |------------->| Figure 12: Role Conflict Flow This flow is a variation on flow III of RFC 3725 [RFC3725]. In fact, it works better than flow III since it produces fewer messages. In this flow, the controller sends an offerless INVITE to agent A, which
responds with its offer, SDP1. The agent then sends an offerless INVITE to agent B, which it responds to with its offer, SDP2. The controller then uses the offer from each agent to generate the answers. When this flow is used, ICE will run between agents A and B, but both will believe they are in the controlling role. With the role conflict resolution procedures, this flow will function properly when ICE is used. At this time, there are no documented flows that can result in the case where both agents believe they are controlled. However, the conflict resolution procedures allow for this case, should a flow arise that would fit into this category.Author's Address
Jonathan Rosenberg jdrosen.net Monmouth, NJ US Email: jdrosen@jdrosen.net URI: http://www.jdrosen.net