RFC6250

Internet Architecture Board (IAB) D. Thaler Request for Comments: 6250 May 2011 Category: Informational ISSN: 2070-1721

                   Evolution of the IP Model

Abstract

This RFC attempts to document various aspects of the IP service model and how it has evolved over time. In particular, it attempts to document the properties of the IP layer as they are seen by upper- layer protocols and applications, especially properties that were (and, at times, still are) incorrectly perceived to exist as well as properties that would cause problems if changed. The discussion of these properties is organized around evaluating a set of claims, or misconceptions. Finally, this document provides some guidance to protocol designers and implementers.

Status of This Memo

This document is not an Internet Standards Track specification; it is published for informational purposes.

This document is a product of the Internet Architecture Board (IAB) and represents information that the IAB has deemed valuable to provide for permanent record. Documents approved for publication by the IAB are not a candidate for any level of Internet Standard; see Section 2 of RFC 5741.

Information about the current status of this document, any errata, and how to provide feedback on it may be obtained at http://www.rfc-editor.org/info/rfc6250.

Copyright Notice

Copyright (c) 2011 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.

       3.1.3. Claim: Error messages can be received in

       3.1.6. Claim: Multicast/broadcast is less expensive

       3.1.7. Claim: The end-to-end latency of the first

       3.1.9. Claim: Loss is rare and probabilistic, not

       3.1.10. Claim: An end-to-end path exists at a

       3.2.1. Claim: Addresses are stable over long

       3.2.3. Claim: A host has only one address on one interface 12
       3.2.4. Claim: A non-multicast/broadcast address
              identifies a single host over a long period of time 13
       3.2.5. Claim: An address can be used as an

       3.2.6. Claim: An address used by an application is

       3.2.8. Claim: Selecting a local address selects

       3.2.9. Claim: An address is part of an on-link

       3.3.1. Claim: New transport-layer protocols can

       3.3.2. Claim: If one stream between a pair of

Introduction

Since the Internet Protocol was first published as [IEN028] in 1978, IP has provided a network-layer connectivity service to upper-layer protocols and applications. The basic IP service model was documented in the original IENs (and subsequently in the RFCs that obsolete them). However, since the mantra has been "Everything Over IP", the IP service model has evolved significantly over the past 30 years to enable new behaviors that the original definition did not envision. For example, by 1989 there was already some confusion and so RFC1122 clarified many things and extended the model. In 2004, RFC3819 advised link-layer protocol designers on a number of issues that affect upper layers and is the closest in intent to this document. Today's IP service model is not well documented in a single place, but is either implicit or discussed piecemeal in many different RFCs. As a result, today's IP service model is actually not well known, or at least is often misunderstood.

In the early days of IP, changing or extending the basic IP service model was easier since it was not as widely deployed and there were fewer implementations. Today, the ossification of the Internet makes evolving the IP model even more difficult. Thus, it is important to understand the evolution of the IP model for two reasons:

1. To clarify what properties can and cannot be depended upon by

   upper-layer protocols and applications.  There are many
   misconceptions on which applications may be based and which are
   problematic.

2. To document lessons for future evolution to take into account.

   It is important that the service model remain consistent, rather
   than evolving in two opposing directions.  It is sometimes the
   case in IETF Working Groups today that directions are considered
   or even taken that would change the IP service model.  Doing this
   without understanding the implications on applications can be
   dangerous.

This RFC attempts to document various aspects of the IP service model and how it has evolved over time. In particular, it attempts to document the properties of the IP layer, as seen by upper-layer protocols and applications, especially properties that were (and at times still are) incorrectly perceived to exist. It also highlights properties that would cause problems if changed.

The IP Service Model

In this document, we use the term "IP service model" to refer to the model exposed by IP to higher-layer protocols and applications. This is depicted in Figure 1 by the horizontal line.

+-------------+                                  +-------------+
| Application |                                  | Application |
+------+------+                                  +------+------+
       |                                                |
+------+------+                                  +------+------+
| Upper-Layer |                                  | Upper-Layer |
|  Protocol   |                                  |  Protocol   |
+------+------+                                  +------+------+
       |                                                |

---

       |                                                |
    +--+--+                  +-----+                 +--+--+
    | IP  |                  | IP  |                 | IP  |
    +--+--+                  +--+--+                 +--+--+
       |                        |                       |
 +-----+------+           +-----+------+          +-----+------+
 | Link Layer |           | Link Layer |          | Link Layer |
 +-----+------+           +--+-----+---+          +-----+------+
       |                     |     |                    |
       +---------------------+     +--------------------+

     Source                                        Destination

                         IP Service Model

                             Figure 1

The foundation of the IP service model today is documented in Section 2.2 of RFC0791. Generally speaking, IP provides a connectionless delivery service for variable size packets, which does not guarantee ordering, delivery, or lack of duplication, but is merely best effort (although some packets may get better service than others). Senders can send to a destination address without signaling a priori, and receivers just listen on an already provisioned address, without signaling a priori.

Architectural principles of the IP model are further discussed in RFC1958 and in Sections 5 and 6 of [NEWARCH].

Links and Subnets

Section 2.1 of RFC4903 discusses the terms "link" and "subnet" with respect to the IP model.

A "link" in the IP service model refers to the topological area within which a packet with an IPv4 Time to Live (TTL) or IPv6 Hop Limit of 1 can be delivered. That is, where no IP-layer forwarding (which entails a TTL/Hop Limit decrement) occurs between two nodes.

A "subnet" in the IP service model refers to the topological area within which addresses from the same subnet prefix are assigned to interfaces.

Common Application Misconceptions

Below is a list of properties that are often assumed by applications and upper-layer protocols, but which have become less true over time.

Misconceptions about Routing

Claim: Reachability is symmetric

Many applications assume that if a host A can contact a host B, then the reverse is also true. Examples of this behavior include request- response patterns, which require reverse reachability only after forward reachability, as well as callbacks (e.g., as used by the File Transfer Protocol (FTP) RFC0959).

Originally, it was the case that reachability was symmetric (although the path taken may not be), both within a link and across the Internet. With the advent of technologies such as Network Address Translators (NATs) and firewalls (as in the following example figure), this can no longer be assumed. Today, host-to-host connectivity is challenging if not impossible in general. It is relatively easy to initiate communication from hosts (A-E in the example diagram) to servers (S), but not vice versa, nor between hosts A-E. For a longer discussion on peer-to-peer connectivity, see Appendix A of RFC5694.

       __________                                 ___       ___
      /          \             ___        ___    /   \ ____|FW |__A
     /            \    ___    /   \ _____|NAT|__|     |    |___|
    |              |__|NAT|__|     |     |___|  |     |__B
    |              |  |___|  |     |__C          \___/
    |              |          \___/               ___
 S__|   Internet   |           ___        ___    /   \
    |              |   ___    /   \ _____|NAT|__|     |__D
    |              |__|FW |__|     |     |___|  |     |
    |              |  |___|  |     |__E          \___/
     \            /           \___/
      \__________/

                             Figure 2

However, it is still the case that if a request can be sent, then a reply to that request can generally be received, but an unsolicited request in the other direction may not be received. RFC2993 discusses this in more detail.

There are also links (e.g., satellite) that were defined as unidirectional links and hence an address on such a link results in asymmetric reachability. RFC3077 explicitly addresses this problem for multihomed hosts by tunneling packets over another interface in order to restore symmetric reachability.

Finally, even with common wireless networks such as 802.11, this assumption may not be true, as discussed in Section 5.5 of [WIRELESS].

Claim: Reachability is transitive

Many applications assume that if a host A can contact host B, and B can contact C, then host A can contact C. Examples of this behavior include applications and protocols that use referrals.

Originally, it was the case that reachability was transitive, both within a link and across the Internet. With the advent of technologies such as NATs and firewalls and various routing policies, this can no longer be assumed across the Internet, but it is often still true within a link. As a result, upper-layer protocols and applications may be relying on transitivity within a link. However, some radio technologies, such as 802.11 ad hoc mode, violate this assumption within a link.

Claim: Error messages can be received in response to data

    packets

Some upper-layer protocols and applications assume that ICMP error messages will be received in response to packets sent that cannot be delivered. Examples of this include the use of Path MTU Discovery RFC1191 RFC1981 relying on messages indicating packets were too big, and traceroute and the use of expanding ring search RFC1812 relying on messages indicating packets reached their TTL/Hop Limit.

Originally, this assumption largely held, but many ICMP senders then chose to rate-limit responses in order to mitigate denial-of-service attacks, and many firewalls now block ICMP messages entirely. For a longer discussion, see Section 2.1 of RFC2923.

This led to an alternate mechanism for Path MTU Discovery that does not rely on this assumption being true RFC4821 and guidance to firewall administrators (RFC4890 and Section 3.1.1 of RFC2979).

Claim: Multicast is supported within a link

RFC1112 introduced multicast to the IP service model. In this evolution, senders still just send to a destination address without signaling a priori, but in contrast to the original IP model, receivers must signal to the network before they can receive traffic to a multicast address.

Today, many applications and protocols use multicast addresses, including protocols for address configuration, service discovery, etc. (See [MCAST4] and [MCAST6] for those that use well-known addresses.)

Most of these only assume that multicast works within a link and may or may not function across a wider area. While network-layer multicast works over most link types, there are Non-Broadcast Multi- Access (NBMA) links over which multicast does not work (e.g., X.25, ATM, frame relay, 6to4, Intra-Site Automatic Tunnel Addressing Protocol (ISATAP), Teredo) and this can interfere with some protocols and applications. Similarly, there are links such as 802.11 ad hoc mode where multicast packets may not get delivered to all receivers on the link. RFC4861 states:

  Note that all link types (including NBMA) are expected to provide
  multicast service for applications that need it (e.g., using
  multicast servers).

and its predecessor RFC2461 contained similar wording.

However, not all link types today meet this expectation.

Claim: IPv4 broadcast is supported

IPv4 broadcast support was originally defined on a link, across a network, and for subnet-directed broadcast, and it is used by many applications and protocols. For security reasons, however, RFC2644 deprecated the forwarding of broadcast packets. Thus, since 1999, broadcast can only be relied on within a link. Still, there exist NBMA links over which broadcast does not work, and there exist some "semi-broadcast" links (e.g., 802.11 ad hoc mode) where broadcast packets may not get delivered to all nodes on the link. Another case where broadcast fails to work is when a /32 or /31 is assigned to a point-to-point interface (e.g., RFC3021), leaving no broadcast address available.

To a large extent, the addition of link-scoped multicast to the IP service model obsoleted the need for broadcast. It is also worth noting that the broadcast API model used by most platforms allows receivers to just listen on an already provisioned address, without signaling a priori, but in contrast to the unicast API model, senders must signal to the local IP stack (SO_BROADCAST) before they can send traffic to a broadcast address. However, from the network's perspective, the host still sends without signaling a priori.

Claim: Multicast/broadcast is less expensive than replicated

    unicast

Some applications and upper-layer protocols that use multicast or broadcast do so not because they do not know the addresses of receivers, but simply to avoid sending multiple copies of the same packet over the same link.

In wired networks, sending a single multicast packet on a link is generally less expensive than sending multiple unicast packets. This may not be true for wireless networks, where implementations can only send multicast at the basic rate, regardless of the negotiated rates of potential receivers. As a result, replicated unicast may achieve much higher throughput across such links than multicast/broadcast traffic.

Claim: The end-to-end latency of the first packet to a

    destination is typical

Many applications and protocols choose a destination address by sending a message to each of a number of candidates, picking the first one to respond, and then using that destination for subsequent communication. If the end-to-end latency of the first packet to each

destination is atypical, this can result in a highly non-optimal destination being chosen, with much longer paths (and hence higher load on the Internet) and lower throughput.

Today, there are a number of reasons this is not true. First, when sending to a new destination there may be some startup latency resulting from the link-layer or network-layer mechanism in use, such as the Address Resolution Protocol (ARP), for instance. In addition, the first packet may follow a different path from subsequent packets. For example, protocols such as Mobile IPv6 RFC3775, Protocol Independent Multicast - Sparse Mode (PIM-SM) RFC4601, and the Multicast Source Discovery Protocol (MSDP) RFC3618 send packets on one path, and then allow immediately switching to a shorter path, resulting in a large latency difference. There are various proposals currently being evaluated by the IETF Routing Research Group that result in similar path switching.

Claim: Reordering is rare

As discussed in [REORDER], RFC2991, and Section 15 of RFC3819, there are a number of effects of reordering. For example, reordering increases buffering requirements (and jitter) in many applications and in devices that do packet reassembly. In particular, TCP RFC0793 is adversely affected by reordering since it enters fast- retransmit when three packets are received before a late packet, which drastically lowers throughput. Finally, some NATs and firewalls assume that the initial fragment arrives first, resulting in packet loss when this is not the case.

Today, there are a number of things that cause reordering. For example, some routers do per-packet, round-robin load balancing, which, depending on the topology, can result in a great deal of reordering. As another example, when a packet is fragmented at the sender, some hosts send the last fragment first. Finally, as discussed in Section 3.1.7, protocols that do path switching after the first packet result in deterministic reordering within the first burst of packets.

Claim: Loss is rare and probabilistic, not deterministic

In the original IP model, senders just send, without signaling the network a priori. This works to a degree. In practice, the last hop (and in rare cases, other hops) of the path needs to resolve next hop information (e.g., the link-layer address of the destination) on demand, which results in queuing traffic, and if the queue fills up, some traffic gets dropped. This means that bursty sources can be problematic (and indeed a single large packet that gets fragmented becomes such a burst). The problem is rarely observed in practice

today, either because the resolution within the last hop happens very quickly, or because bursty applications are rarer. However, any protocol that significantly increases such delays or adds new resolutions would be a change to the classic IP model and may adversely impact upper-layer protocols and applications that result in bursts of packets.

In addition, mechanisms that simply drop the first packet, rather than queuing it, also break this assumption. Similar to the result of reordering, they can result in a highly non-optimal destination being chosen by applications that use the first one to respond. Two examples of mechanisms that appear to do this are network interface cards that support a "Wake-on-LAN" capability where any packet that matches a specified pattern will wake up a machine in a power- conserving mode, but only after dropping the matching packet, and MSDP, where encapsulating data packets is optional, but doing so enables bursty sources to be accommodated while a multicast tree is built back to the source's domain.

3.1.10. Claim: An end-to-end path exists at a single point in time

In classic IP, applications assume that either an end-to-end path exists to a destination or that the packet will be dropped. In addition, IP today tends to assume that the packet delay is relatively short (since the "Time"-to-Live is just a hop count). In IP's earlier history, the TTL field was expected to also be decremented each second (not just each hop).

In general, this assumption is still true today. However, the IRTF Delay Tolerant Networking Research Group is investigating ways for applications to use IP in networks where this assumption is not true, such as store-and-forward networks (e.g., packets carried by vehicles or animals).

3.1.11. Discussion

The reasons why the assumptions listed above are increasingly less true can be divided into two categories: effects caused by attributes of link-layer technologies and effects caused by network-layer technologies.

RFC 3819 RFC3819 advises link-layer protocol designers to minimize these effects. Generally, the link-layer causes are not intentionally trying to break IP, but rather adding IP over the technology introduces the problem. Hence, where the link-layer protocol itself does not do so, when specifying how IP is defined over such a link protocol, designers should compensate to the maximum extent possible. As examples, RFC3077 and RFC2491 compensate for

the lack of symmetric reachability and the lack of link-layer multicast, respectively. That is, when IP is defined over a link type, the protocol designers should attempt to restore the assumptions listed in this document. For example, since an implementation can distinguish between 802.11 ad hoc mode versus infrastructure mode, it may be possible to define a mechanism below IP to compensate for the lack of transitivity over such links.

At the network layer, as a general principle, we believe that reachability is good. For security reasons (RFC4948), however, it is desirable to restrict reachability by unauthorized parties; indeed IPsec, an integral part of the IP model, provides one means to do so. Where there are issues with asymmetry, non-transitivity, and so forth, which are not direct results of restricting reachability to only authorized parties (for some definition of authorized), the IETF should attempt to avoid or solve such issues. Similar to the principle outlined in Section 3.9 of RFC1958, the general theme when defining a protocol is to be liberal in what effects you accept, and conservative in what effects you cause.

However, in being liberal in what effects you accept, it is also important to remember that diagnostics are important, and being too liberal can mask problems. Thus, a tussle exists between the desire to provide a better experience to one's own users or applications and thus be more successful (RFC5218), versus the desire to put pressure on getting problems fixed. One solution is to provide a separate "pedantic mode" that can be enabled to see the problems rather than mask them.

Misconceptions about Addressing

Claim: Addresses are stable over long periods of time

Originally, addresses were manually configured on fixed machines, and hence addresses were very stable. With the advent of technologies such as DHCP, roaming, and wireless, addresses can no longer be assumed to be stable for long periods of time. However, the APIs provided to applications today typically still assume stable addresses (e.g., address lifetimes are not exposed to applications that get addresses). This can cause problems when addresses become stale.

For example, many applications resolve names to addresses and then cache them without any notion of lifetime. In fact, the classic name resolution APIs do not even provide applications with the lifetime of entries.

Proxy Mobile IPv6 RFC5213 tries to restore this assumption to some extent by preserving the same address while roaming around a local area. The issue of roaming between different networks has been known since at least 1980 when [IEN135] proposed a mobility solution that attempted to restore this assumption by adding an additional address that can be used by applications, which is stable while roaming anywhere with Internet connectivity. More recent protocols such as Mobile IPv6 (MIP6) RFC3775 and the Host Identity Protocol (HIP) RFC4423 follow in this same vein.

Claim: An address is four bytes long

Many applications and protocols were designed to only support addresses that are four bytes long. Although this was sufficient for IPv4, the advent of IPv6 made this assumption invalid and with the exhaustion of IPv4 address space this assumption will become increasingly less true. There have been some attempts to try to mitigate this problem with limited degrees of success in constrained cases. For example, "Bump-In-the-Stack" RFC2767 and "Bump-in-the- API" RFC3338 attempt to provide four-byte "IPv4" addresses for IPv6 destinations, but have many limitations including (among a number of others) all the problems of NATs.

Claim: A host has only one address on one interface

Although many applications assume this (e.g., by calling a name resolution function such as gethostbyname and then just using the first address returned), it was never really true to begin with, even if it was the common case. Even RFC0791 states:

  ... provision must be made for a host to have several physical
  interfaces to the network with each having several logical
  Internet addresses.

However, this assumption is increasingly less true today, with the advent of multiple interfaces (e.g., wired and wireless), dual-IPv4/ IPv6 nodes, multiple IPv6 addresses on the same interface (e.g., link-local and global), etc. Similarly, many protocol specifications such as DHCP only describe operations for a single interface, whereas obtaining host-wide configuration from multiple interfaces presents a merging problem for nodes in practice. Too often, this problem is simply ignored by Working Groups, and applications and users suffer as a result from poor merging algorithms.

One use of protocols such as MIP6 and HIP is to make this assumption somewhat more true by adding an additional "address" that can be the one used by such applications, and the protocol will deal with the complexity of multiple physical interfaces and addresses.

Claim: A non-multicast/broadcast address identifies a single

    host over a long period of time

Many applications and upper-layer protocols maintain a communication session with a destination over some period of time. If that address is reassigned to another host, or if that address is assigned to multiple hosts and the host at which packets arrive changes, such applications can have problems.

In addition, many security mechanisms and configurations assume that one can block traffic by IP address, implying that a single attacker can be identified by IP address. If that IP address can also identify many legitimate hosts, applying such a block can result in denial of service.

RFC1546 introduced the notion of anycast to the IP service model. It states:

  Because anycasting is stateless and does not guarantee delivery of
  multiple anycast datagrams to the same system, an application
  cannot be sure that it is communicating with the same peer in two
  successive UDP transmissions or in two successive TCP connections
  to the same anycast address.

  The obvious solutions to these issues are to require applications
  which wish to maintain state to learn the unicast address of their
  peer on the first exchange of UDP datagrams or during the first
  TCP connection and use the unicast address in future
  conversations.

The issues with anycast are further discussed in RFC4786 and [ANYCAST].

Another mechanism by which multiple hosts use the same address is as a result of scoped addresses, as defined for both IPv4 RFC1918 RFC3927 and IPv6 RFC4007. Because such addresses can be reused within multiple networks, hosts in different networks can use the same address. As a result, a host that is multihomed to two such networks cannot use the destination address to uniquely identify a peer. For example, a host can no longer use a 5-tuple to uniquely identify a TCP connection. This is why IPv6 added the concept of a "zone index".

Yet another example is that, in some high-availability solutions, one host takes over the IP address of another failed host.

See RFC2101, RFC2775, and [SHARED-ADDRESSING] for additional discussion on address uniqueness.

Claim: An address can be used as an indication of physical

    location

Some applications attempt to use an address to infer some information about the physical location of the host with that address. For example, geo-location services are often used to provide targeted content or ads.

Various forms of tunneling have made this assumption less true, and this will become increasingly less true as the use of IPv4 NATs for large networks continues to increase. See Section 7 of [SHARED-ADDRESSING] for a longer discussion.

Claim: An address used by an application is the same as the

    address used for routing

Some applications assume that the address the application uses is the same as that used by routing. For example, some applications use raw sockets to read/write packet headers, including the source and destination addresses in the IP header. As another example, some applications make assumptions about locality (e.g., whether the destination is on the same subnet) by comparing addresses.

Protocols such as Mobile IPv6 and HIP specifically break this assumption (in an attempt to restore other assumptions as discussed above). Recently, the IRTF Routing Research Group has been evaluating a number of possible mechanisms, some of which would also break this assumption, while others preserve this assumption near the edges of the network and only break it in the core of the Internet.

Breaking this assumption is sometimes referred to as an "identifier/ locator" split. However, as originally defined in 1978 ([IEN019], [IEN023]), an address was originally defined as only a locator, whereas names were defined to be the identifiers. However, the TCP protocol then used addresses as identifiers.

Finally, in a liberal sense, any tunneling mechanism might be said to break this assumption, although, in practice, applications that make this assumption will continue to work, since the address of the inside of the tunnel is still used for routing as expected.

Claim: A subnet is smaller than a link

In the classic IP model, a "subnet" is smaller than, or equal to, a "link". Destinations with addresses in the same on-link subnet prefix can be reached with TTL (or Hop Count) = 1. Link-scoped multicast packets, and all-ones broadcast packets will be delivered (in a best-effort fashion) to all listening nodes on the link.

Subnet broadcast packets will be delivered (in a best effort fashion) to all listening nodes in the subnet. There have been some efforts in the past (e.g., RFC0925, RFC3069) to allow multi-link subnets and change the above service model, but the adverse impact on applications that have such assumptions recommend against changing this assumption.

RFC4903 discusses this topic in more detail and surveys a number of protocols and applications that depend on this assumption. Specifically, some applications assume that, if a destination address is in the same on-link subnet prefix as the local machine, then therefore packets can be sent with TTL=1, or that packets can be received with TTL=255, or link-scoped multicast or broadcast can be used to reach the destination.

Claim: Selecting a local address selects the interface

Some applications assume that binding to a given local address constrains traffic reception to the interface with that address, and that traffic from that address will go out on that address's interface. However, Section 3.3.4.2 of RFC1122 defines two models: the Strong End System (or strong host) model where this is true, and the Weak End System (or weak host) model where this is not true. In fact, any router is inherently a weak host implementation, since packets can be forwarded between interfaces.

Claim: An address is part of an on-link subnet prefix

To some extent, this was never true, in that there were cases in IPv4 where the "mask" was 255.255.255.255, such as on a point-to-point link where the two endpoints had addresses out of unrelated address spaces, and no on-link subnet prefix existed on the link. However, this didn't stop many platforms and applications from assuming that every address had a "mask" (or prefix) that was on-link. The assumption of whether a subnet is on-link (in which case one can send directly to the destination after using ARP/ND) or off-link (in which case one just sends to a router) has evolved over the years, and it can no longer be assumed that an address has an on-link prefix. In 1998, RFC2461 introduced the distinction as part of the core IPv6 protocol suite. This topic is discussed further in [ON-OFF-LINK], and RFC4903 also touches on this topic with respect to the service model seen by applications.

3.2.10. Discussion

Section 4.1 of RFC 1958 RFC1958 states: "In general, user applications should use names rather than addresses".

We emphasize the above point, which is too often ignored. Many commonly used APIs unnecessarily expose addresses to applications that already use names. Similarly, some protocols are defined to carry addresses, rather than carrying names (instead of or in addition to addresses). Protocols and applications that are already dependent on a naming system should be designed in such a way that they avoid or minimize any dependence on the notion of addresses.

One challenge is that many hosts today do not have names that can be resolved. For example, a host may not have a fully qualified domain name (FQDN) or a Domain Name System (DNS) server that will host its name.

Applications that, for whatever reason, cannot use names should be IP-version agnostic.

Misconceptions about Upper-Layer Extensibility

Claim: New transport-layer protocols can work across the

    Internet

IP was originally designed to support the addition of new transport- layer protocols, and [PROTOCOLS] lists many such protocols.

However, as discussed in [WAIST-HOURGLASS], NATs and firewalls today break this assumption and often only allow UDP and TCP (or even just HTTP).

Hence, while new protocols may work from some places, they will not necessarily work from everywhere, such as from behind such NATs and firewalls.

Since even UDP and TCP may not work from everywhere, it may be necessary for applications to support "HTTP failover" modes. The use of HTTP as a "transport of last resort" has become common (e.g., [BOSH] among others) even in situations where it is sub-optimal, such as in real-time communications or where bidirectional communication is required. Also, the IETF HyBi Working Group is now in the process of designing a standards-based solution for layering other protocols on top of HTTP. As a result of having to support HTTP failover, applications may have to be engineered to sustain higher latency.

Claim: If one stream between a pair of addresses can get

    through, then so can another

Some applications and protocols use multiple upper-layer streams of data between the same pair of addresses and initiated by the same party. Passive-mode FTP RFC0959, and RTP RFC3550, are two examples of such protocols, which use separate streams for data versus control channels.

Today, there are many reasons why this may not be true. Firewalls, for example, may selectively allow/block specific protocol numbers and/or values in upper-layer protocol fields (such as port numbers). Similarly, middleboxes such as NATs that create per-stream state may cause other streams to fail once they run out of space to store additional stream state.

Discussion

Section 5.1 of [NEWARCH] discusses the primary requirements of the original Internet architecture, including Service Generality. It states:

  This goal was to support the widest possible range of
  applications, by supporting a variety of types of service at the
  transport level.  Services might be distinguished by speed,
  latency, or reliability, for example.  Service types might include
  virtual circuit service, which provides reliable, full-duplex byte
  streams, and also datagram service, which delivers individual
  packets with no guarantees of reliability or ordering.  The
  requirement for datagram service was motivated by early ARPAnet
  experiments with packet speech (using IMP Type 3 messages).

The reasons that the assumptions in this section are becoming less true are due to network-layer (or higher-layer) techniques being introduced that interfere with the original requirement. Generally, these are done either in the name of security or as a side effect of solving some other problem such as address shortage. Work is needed to investigate ways to restore the original behavior while still meeting today's security requirements.

Misconceptions about Security

Claim: Packets are unmodified in transit

Some applications and upper-layer protocols assume that a packet is unmodified in transit, except for a few well-defined fields (e.g., TTL). Examples of this behavior include protocols that define their own integrity-protection mechanism such as a checksum.

This assumption is broken by NATs as discussed in RFC2993 and other middleboxes that modify the contents of packets. There are many tunneling technologies (e.g., RFC4380) that attempt to restore this assumption to some extent.

The IPsec architecture RFC4301 added security to the IP model, providing a way to address this problem without changing applications, although transport-mode IPsec is not currently widely used over the Internet.

Claim: Packets are private

The assumption that data is private has never really been true. However, many old applications and protocols (e.g., FTP) transmit passwords or other sensitive data in the clear.

IPsec provides a way to address this problem without changing applications, although it is not yet widely deployed, and doing encryption/decryption for all packets can be computationally expensive.

Claim: Source addresses are not forged

Most applications and protocols use the source address of some incoming packet when generating a response, and hence assume that it has not been forged (and as a result can often be vulnerable to various types of attacks such as reflection attacks).

Various mechanisms that restore this assumption include, for example, IPsec and Cryptographically Generated Addresses (CGAs) RFC3972.

Discussion

A good discussion of threat models and common tools can be found in RFC3552. Protocol designers and applications developers are encouraged to be familiar with that document.

Security Considerations

This document discusses assumptions about the IP service model made by many applications and upper-layer protocols. Whenever these assumptions are broken, if the application or upper-layer protocol has some security-related behavior that is based on the assumption, then security can be affected.

For example, if an application assumes that binding to the IP address of a "trusted" interface means that it will never receive traffic from an "untrusted" interface, and that assumption is broken (as discussed in Section 3.2.8), then an attacker could get access to private information.

As a result, great care should be taken when expanding the extent to which an assumption is false. On the other hand, application and upper-layer protocol developers should carefully consider the impact of basing their security on any of the assumptions enumerated in this document.

It is also worth noting that many of the changes that have occurred over time (e.g., firewalls, dropping directed broadcasts, etc.) that are discussed in this document were done in the interest of improving security at the expense of breaking some applications.

Conclusion

Because a huge number of applications already exist that use TCP/IP for business-critical operations, any changes to the service model need to be done with extreme care. Extensions that merely add additional optional functionality without impacting any existing applications are much safer than extensions that change one or more of the core assumptions discussed above. Any changes to the above assumptions should only be done in accordance with some mechanism to minimize or mitigate the risks of breaking mission-critical applications. Historically, changes have been done without regard to such considerations and, as a result, the situation for applications today is already problematic. The key to maintaining an interoperable Internet is documenting and maintaining invariants that higher layers can depend on, and being very judicious with changes.

In general, lower-layer protocols should document the contract they provide to higher layers; that is, what assumptions the upper layer can rely on (sometimes this is done in the form of an applicability statement). Conversely, higher-layer protocols should document the assumptions they rely on from the lower layer (sometimes this is done in the form of requirements).

We must also recognize that a successful architecture often evolves as success brings growth and as technology moves forward. As a result, the various assumptions made should be periodically reviewed when updating protocols.

Acknowledgements

Chris Hopps, Dow Street, Phil Hallam-Baker, and others provided helpful discussion on various points that led to this document. Iain Calder, Brian Carpenter, Jonathan Rosenberg, Erik Nordmark, Alain Durand, and Iljitsch van Beijnum also provided valuable feedback.

IAB Members at the Time of This Writing

Loa Andersson Gonzalo Camarillo Stuart Cheshire Russ Housley Olaf Kolkman Gregory Lebovitz Barry Leiba Kurtis Lindqvist Andrew Malis Danny McPherson David Oran Dave Thaler Lixia Zhang

IAB Members at the Time of Approval

Bernard Aboba Marcelo Bagnulo Ross Callon Spencer Dawkins Russ Housley John Klensin Olaf Kolkman Danny McPherson Jon Peterson Andrei Robachevsky Dave Thaler Hannes Tschofenig

References

Normative References

RFC0791 Postel, J., "Internet Protocol", STD 5, RFC 791,

          September 1981.

RFC1112 Deering, S., "Host extensions for IP multicasting", STD 5,

          RFC 1112, August 1989.

RFC1122 Braden, R., "Requirements for Internet Hosts -

          Communication Layers", STD 3, RFC 1122, October 1989.

RFC1546 Partridge, C., Mendez, T., and W. Milliken, "Host

          Anycasting Service", RFC 1546, November 1993.

RFC2461 Narten, T., Nordmark, E., and W. Simpson, "Neighbor

          Discovery for IP Version 6 (IPv6)", RFC 2461,
          December 1998.

RFC2644 Senie, D., "Changing the Default for Directed Broadcasts

          in Routers", BCP 34, RFC 2644, August 1999.

RFC4301 Kent, S. and K. Seo, "Security Architecture for the

          Internet Protocol", RFC 4301, December 2005.

RFC4861 Narten, T., Nordmark, E., Simpson, W., and H. Soliman,

          "Neighbor Discovery for IP version 6 (IPv6)", RFC 4861,
          September 2007.

Informative References

[ANYCAST] McPherson, D. and D. Oran, "Architectural Considerations

          of IP Anycast", Work in Progress, February 2010.

[BOSH] Paterson, I., Smith, D., Saint-Andre, P., and J. Moffitt,

          "Bidirectional-streams Over Synchronous HTTP (BOSH)",
          XEP 0124, 2010,
          <http://xmpp.org/extensions/xep-0124.html>.

[IEN019] Shoch, J., "A note on Inter-Network Naming, Addressing,

          and Routing", IEN 19, January 1978,
          <http://www.rfc-editor.org/ien/ien19.txt>.

[IEN023] Cohen, D., "On Names, Addresses and Routings", IEN 23,

          January 1978, <http://www.rfc-editor.org/ien/ien23.txt>.

[IEN028] Postel, J., "Draft Internetwork Protocol Specification",

          IEN 28, February 1978,
          <http://www.rfc-editor.org/ien/ien28.pdf>.

[IEN135] Sunshine, C. and J. Postel, "Addressing Mobile Hosts in

          the ARPA Internet Environment", IEN 135, March 1980,
          <http://www.rfc-editor.org/ien/ien135.txt>.

[MCAST4] Internet Assigned Numbers Authority, "IPv4 Multicast

          Addresses",
          <http://www.iana.org/assignments/multicast-addresses>.

[MCAST6] Internet Assigned Numbers Authority, "INTERNET PROTOCOL

          VERSION 6 MULTICAST ADDRESSES",
          <http://www.iana.org/assignments/
          ipv6-multicast-addresses>.

[NEWARCH] Clark, D., et al., "New Arch: Future Generation Internet

          Architecture", Air Force Research Laboratory Technical
          Report AFRL-IF-RS-TR-2004-235, August 2004, <http://
          www.dtic.mil/cgi-bin/
          GetTRDoc?AD=ADA426770&Location=U2&doc=GetTRDoc.pdf>.

[ON-OFF-LINK]

          Singh, H., Beebee, W., and E. Nordmark, "IPv6 Subnet
          Model", Work in Progress, February 2008.

[PROTOCOLS]

          Internet Assigned Numbers Authority, "Protocol Numbers",
          <http://www.iana.org/assignments/protocol-numbers>.

[REORDER] Bennett, J., Partridge, C., and N. Shectman, "Packet

          reordering is not pathological network behavior", IEEE/ACM
          Transactions on Networking, Vol. 7, No. 6, December 1999.

RFC0793 Postel, J., "Transmission Control Protocol", STD 7,

          RFC 793, September 1981.

RFC0925 Postel, J., "Multi-LAN address resolution", RFC 925,

          October 1984.

RFC0959 Postel, J. and J. Reynolds, "File Transfer Protocol",

          STD 9, RFC 959, October 1985.

RFC1191 Mogul, J. and S. Deering, "Path MTU discovery", RFC 1191,

          November 1990.

RFC1812 Baker, F., "Requirements for IP Version 4 Routers",

          RFC 1812, June 1995.

RFC1918 Rekhter, Y., Moskowitz, R., Karrenberg, D., Groot, G., and

          E. Lear, "Address Allocation for Private Internets",
          BCP 5, RFC 1918, February 1996.

RFC1958 Carpenter, B., "Architectural Principles of the Internet",

          RFC 1958, June 1996.

RFC1981 McCann, J., Deering, S., and J. Mogul, "Path MTU Discovery

          for IP version 6", RFC 1981, August 1996.

RFC2101 Carpenter, B., Crowcroft, J., and Y. Rekhter, "IPv4

          Address Behaviour Today", RFC 2101, February 1997.

RFC2491 Armitage, G., Schulter, P., Jork, M., and G. Harter, "IPv6

          over Non-Broadcast Multiple Access (NBMA) networks",
          RFC 2491, January 1999.

RFC2767 Tsuchiya, K., HIGUCHI, H., and Y. Atarashi, "Dual Stack

          Hosts using the "Bump-In-the-Stack" Technique (BIS)",
          RFC 2767, February 2000.

RFC2775 Carpenter, B., "Internet Transparency", RFC 2775,

          February 2000.

RFC2923 Lahey, K., "TCP Problems with Path MTU Discovery",

          RFC 2923, September 2000.

RFC2979 Freed, N., "Behavior of and Requirements for Internet

          Firewalls", RFC 2979, October 2000.

RFC2991 Thaler, D. and C. Hopps, "Multipath Issues in Unicast and

          Multicast Next-Hop Selection", RFC 2991, November 2000.

RFC2993 Hain, T., "Architectural Implications of NAT", RFC 2993,

          November 2000.

RFC3021 Retana, A., White, R., Fuller, V., and D. McPherson,

          "Using 31-Bit Prefixes on IPv4 Point-to-Point Links",
          RFC 3021, December 2000.

RFC3069 McPherson, D. and B. Dykes, "VLAN Aggregation for

          Efficient IP Address Allocation", RFC 3069, February 2001.

RFC3077 Duros, E., Dabbous, W., Izumiyama, H., Fujii, N., and Y.

          Zhang, "A Link-Layer Tunneling Mechanism for
          Unidirectional Links", RFC 3077, March 2001.

RFC3338 Lee, S., Shin, M-K., Kim, Y-J., Nordmark, E., and A.

          Durand, "Dual Stack Hosts Using "Bump-in-the-API" (BIA)",
          RFC 3338, October 2002.

RFC3550 Schulzrinne, H., Casner, S., Frederick, R., and V.

          Jacobson, "RTP: A Transport Protocol for Real-Time
          Applications", STD 64, RFC 3550, July 2003.

RFC3552 Rescorla, E. and B. Korver, "Guidelines for Writing RFC

          Text on Security Considerations", BCP 72, RFC 3552,
          July 2003.

RFC3618 Fenner, B. and D. Meyer, "Multicast Source Discovery

          Protocol (MSDP)", RFC 3618, October 2003.

RFC3775 Johnson, D., Perkins, C., and J. Arkko, "Mobility Support

          in IPv6", RFC 3775, June 2004.

RFC3819 Karn, P., Bormann, C., Fairhurst, G., Grossman, D.,

          Ludwig, R., Mahdavi, J., Montenegro, G., Touch, J., and L.
          Wood, "Advice for Internet Subnetwork Designers", BCP 89,
          RFC 3819, July 2004.

RFC3927 Cheshire, S., Aboba, B., and E. Guttman, "Dynamic

          Configuration of IPv4 Link-Local Addresses", RFC 3927,
          May 2005.

RFC3972 Aura, T., "Cryptographically Generated Addresses (CGA)",

          RFC 3972, March 2005.

RFC4007 Deering, S., Haberman, B., Jinmei, T., Nordmark, E., and

          B. Zill, "IPv6 Scoped Address Architecture", RFC 4007,
          March 2005.

RFC4380 Huitema, C., "Teredo: Tunneling IPv6 over UDP through

          Network Address Translations (NATs)", RFC 4380,
          February 2006.

RFC4423 Moskowitz, R. and P. Nikander, "Host Identity Protocol

          (HIP) Architecture", RFC 4423, May 2006.

RFC4601 Fenner, B., Handley, M., Holbrook, H., and I. Kouvelas,

          "Protocol Independent Multicast - Sparse Mode (PIM-SM):
          Protocol Specification (Revised)", RFC 4601, August 2006.

RFC4786 Abley, J. and K. Lindqvist, "Operation of Anycast

          Services", BCP 126, RFC 4786, December 2006.

RFC4821 Mathis, M. and J. Heffner, "Packetization Layer Path MTU

          Discovery", RFC 4821, March 2007.

RFC4890 Davies, E. and J. Mohacsi, "Recommendations for Filtering

          ICMPv6 Messages in Firewalls", RFC 4890, May 2007.

RFC4903 Thaler, D., "Multi-Link Subnet Issues", RFC 4903,

          June 2007.

RFC4948 Andersson, L., Davies, E., and L. Zhang, "Report from the

          IAB workshop on Unwanted Traffic March 9-10, 2006",
          RFC 4948, August 2007.

RFC5213 Gundavelli, S., Leung, K., Devarapalli, V., Chowdhury, K.,

          and B. Patil, "Proxy Mobile IPv6", RFC 5213, August 2008.

RFC5218 Thaler, D. and B. Aboba, "What Makes For a Successful

          Protocol?", RFC 5218, July 2008.

RFC5694 Camarillo, G. and IAB, "Peer-to-Peer (P2P) Architecture:

          Definition, Taxonomies, Examples, and Applicability",
          RFC 5694, November 2009.

[SHARED-ADDRESSING]

          Ford, M., Boucadair, M., Durand, A., Levis, P., and P.
          Roberts, "Issues with IP Address Sharing", Work
          in Progress, March 2011.

[WAIST-HOURGLASS]

          Rosenberg, J., "UDP and TCP as the New Waist of the
          Internet Hourglass", Work in Progress, February 2008.

[WIRELESS]

          Kotz, D., Newport, C., and C. Elliott, "The mistaken
          axioms of wireless-network research", Dartmouth College
          Computer Science Technical Report TR2003-467, July 2003, <
          http://www.cs.dartmouth.edu/cms_file/SYS_techReport/337/
          TR2003-467.pdf>.

Authors' Addresses

Dave Thaler One Microsoft Way Redmond, WA 98052 USA

Phone: +1 425 703 8835 EMail: [email protected]

Internet Architecture Board

EMail: [email protected]