[Docs] [txt|pdf] [Tracker] [WG] [Email] [Diff1] [Diff2] [Nits]

Versions: (draft-touch-intarea-tunnels) 00 01 02 03 04 05 06 07

Internet Area WG                                               J. Touch
Internet Draft                                                  USC/ISI
Intended status: Best Current Practice                      M. Townsley
Updates: 4459                                                     Cisco
Expires: December 2017                                     June 8, 2017

                  IP Tunnels in the Internet Architecture

Status of this Memo

   This Internet-Draft is submitted in full conformance with the
   provisions of BCP 78 and BCP 79.

   This document may contain material from IETF Documents or IETF
   Contributions published or made publicly available before November
   10, 2008. The person(s) controlling the copyright in some of this
   material may not have granted the IETF Trust the right to allow
   modifications of such material outside the IETF Standards Process.
   Without obtaining an adequate license from the person(s) controlling
   the copyright in such materials, this document may not be modified
   outside the IETF Standards Process, and derivative works of it may
   not be created outside the IETF Standards Process, except to format
   it for publication as an RFC or to translate it into languages other
   than English.

   Internet-Drafts are working documents of the Internet Engineering
   Task Force (IETF), its areas, and its working groups.  Note that
   other groups may also distribute working documents as Internet-

   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."

   The list of current Internet-Drafts can be accessed at

   The list of Internet-Draft Shadow Directories can be accessed at

   This Internet-Draft will expire on November 8, 2017.

Touch, Townsley        Expires December 8, 2017                [Page 1]

Internet-Draft         Tunnels in the Internet                June 2017

Copyright Notice

   Copyright (c) 2017 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.


   This document discusses the role of IP tunnels in the Internet
   architecture. An IP tunnel transits IP datagrams as payloads in non-
   link layer protocols. This document explains the relationship of IP
   tunnels to existing protocol layers and the challenges in supporting
   IP tunneling, based on the equivalence of tunnels to links. The
   implications of this document are used to derive recommendations that
   update MTU and fragment issues in RFC 4459.

Table of Contents

   1. Introduction...................................................3
   2. Conventions used in this document..............................6
      2.1. Key Words.................................................6
      2.2. Terminology...............................................6
   3. The Tunnel Model..............................................10
      3.1. What is a Tunnel?........................................11
      3.2. View from the Outside....................................13
      3.3. View from the Inside.....................................14
      3.4. Location of the Ingress and Egress.......................15
      3.5. Implications of This Model...............................15
      3.6. Fragmentation............................................16
         3.6.1. Outer Fragmentation.................................16
         3.6.2. Inner Fragmentation.................................18
         3.6.3. The Necessity of Outer Fragmentation................19
   4. IP Tunnel Requirements........................................20
      4.1. Encapsulation Header Issues..............................20
         4.1.1. General Principles of Header Fields Relationships...20
         4.1.2. Addressing Fields...................................21
         4.1.3. Hop Count Fields....................................21

Touch, Townsley        Expires December 8, 2017                [Page 2]

Internet-Draft         Tunnels in the Internet                June 2017

         4.1.4. IP Fragment Identification Fields...................22
         4.1.5. Checksums...........................................23
      4.2. MTU Issues...............................................24
         4.2.1. Minimum MTU Considerations..........................24
         4.2.2. Fragmentation.......................................27
         4.2.3. Path MTU Discovery..................................30
      4.3. Coordination Issues......................................32
         4.3.1. Signaling...........................................32
         4.3.2. Congestion..........................................34
         4.3.3. Multipoint Tunnels and Multicast....................34
         4.3.4. Load Balancing......................................35
         4.3.5. Recursive Tunnels...................................36
   5. Observations..................................................37
      5.1. Summary of Recommendations...............................37
      5.2. Impact on Existing Encapsulation Protocols...............37
      5.3. Tunnel Protocol Designers................................40
         5.3.1. For Future Standards................................40
         5.3.2. Diagnostics.........................................40
      5.4. Tunnel Implementers......................................41
      5.5. Tunnel Operators.........................................41
   6. Security Considerations.......................................42
   7. IANA Considerations...........................................43
   8. References....................................................43
      8.1. Normative References.....................................43
      8.2. Informative References...................................43
   9. Acknowledgments...............................................48
   APPENDIX A: Fragmentation efficiency.............................50
      A.1. Selecting fragment sizes.................................50
      A.2. Packing..................................................51

1. Introduction

   The Internet layering architecture is loosely based on the ISO seven
   layer stack, in which data units traverse the stack by being wrapped
   inside data units of the next layer down [Cl88][Zi80]. A tunnel is a
   mechanism for transmitting data units between endpoints by wrapping
   them as data units of the same or higher layers, e.g., IP in IP
   (Figure 1) or IP in UDP (Figure 2).

                        | IP'| IP |     Data     |

                           Figure 1 IP inside IP

Touch, Townsley        Expires December 8, 2017                [Page 3]

Internet-Draft         Tunnels in the Internet                June 2017

                     | IP'| UDP | IP |     Data     |

                   Figure 2 IP in UDP in IP in Ethernet

   This document focuses on tunnels that transit IP packets, i.e., in
   which an IP packet is the payload of another protocol, other than a
   typical link layer. A tunnel is a virtual link that can help decouple
   the network topology seen by transiting packets from the underlying
   physical network [To98][RFC2473]. Tunnels were critical in the
   development of multicast because not all routers were capable of
   processing multicast packets [Er94]. Tunnels allowed multicast
   packets to transit efficiently between multicast-capable routers over
   paths that did not support native link-layer multicast. Similar
   techniques have been used to support incremental deployment of other
   protocols over legacy substrates, such as IPv6 [RFC2546].

   Use of tunnels is common in the Internet. The word "tunnel" occurs in
   nearly 1,500 RFCs (of nearly 8,000 current RFCs, close to 20%), and
   is supported within numerous protocols, including:

   o  IP in IP / mobile IP - IPv4 in IPv4 tunnels using protocol 4
      [RFC2003][RFC2473][RFC5944] and its precursor called "IPIP" using
      protocol 94 [RFC1853]

   o  IP in IPv6 - IPv6 or IPv4 in IPv6 [RFC2473]

   o  IPsec - includes a tunnel mode to enable encryption or
      authentication of the an entire IP datagram inside another IP
      datagram [RFC4301]

   o  Generic Router Encapsulation (GRE) - a shim layer for tunneling
      any network layer in any other network layer, as in IP in GRE in
      IP [RFC2784][RFC7588][RFC7676], or inside UDP in IP [RFC8086]

   o  MPLS - a shim layer for tunneling IP over a circuit-like path over
      a link layer [RFC3031] or inside UDP in IP [RFC7510], in which
      identifiers are rewritten on each hop, often used for traffic

   o  LISP - a mechanism that uses multipoint IP tunnels to reduce
      routing table load within an enclave of routers at the expense of
      more complex tunnel ingress encapsulation tables [RFC6830]

Touch, Townsley        Expires December 8, 2017                [Page 4]

Internet-Draft         Tunnels in the Internet                June 2017

   o  TRILL - a mechanism that uses multipoint L2 tunnels to enable use
      of L3 routing (typically IS-IS) in an enclave of Ethernet bridges

   o  Generic UDP Encapsulation (GUE) - IP in UDP in IP [He16]

   o  Automatic Multicast Tunneling (AMT) - IP in UDP in IP for
      multicast [RFC7450]

   o  L2TP - PPP over IP, to extend a subscriber's DSL/FTTH connection
      from an access line provider to an ISP [RFC3931]

   o  L2VPNs - provides a link topology different from that provided by
      physical links [RFC4664]; many of these are not classical tunnels,
      using only tags (Ethernet VLAN tags) rather than encapsulation

   o  L3VPNs - provides a network topology different from that provided
      by ISPs [RFC4176]

   o  NVO3 - data center network sharing (to be determined, which may
      include use of GUE or other tunnels) [RFC7364]

   o  PWE3 - emulates wire-like services over packet-switched services

   o  SEAL/AERO -IP in IP tunneling with an additional shim header
      designed to overcome the limitations of RFC2003 [RFC5320][Te17]

   o  A number of legacy variants, including swIPe (an IPsec precursor),
      a GRE precursor, and the Internet Encapsulation Protocol, all of
      which included a shim layer [RFC1853]

   The variety of tunnel mechanisms raises the question of the role of
   tunnels in the Internet architecture and the potential need for these
   mechanisms to have similar and predictable behavior. In particular,
   the ways in which packet size (i.e., Maximum Transmission Unit or
   MTU) mismatches and error signals (e.g., ICMP) are handled may
   benefit from a coordinated approach.

   Regardless of the layer in which encapsulation occurs, tunnels
   emulate a link. The only difference is that a link operates over a
   physical communication channel, whereas a tunnel operates over other
   software protocol layers. Because tunnels are links, they are subject
   to the same issues as any link, e.g., MTU discovery, signaling, and
   the potential utility of native support for broadcast and multicast
   [RFC3819]. Tunnels have some advantages over native links, being
   potentially easier to reconfigure and control because they can

Touch, Townsley        Expires December 8, 2017                [Page 5]

Internet-Draft         Tunnels in the Internet                June 2017

   generally rely on existing out-of-band communication between its

   The first attempt to use large-scale tunnels was to transit multicast
   traffic across the Internet in 1988, and this resulted in 'tunnel
   collapse'. At the time, tunnels were not implemented as
   encapsulation-based virtual links, but rather as loose source routes
   on un-encapsulated IP datagrams [RFC1075]. Then, as now, routers did
   not support use of the loose source route IP option at line rate, and
   the multicast traffic caused overload of the so-called "slow path"
   processing of IP datagrams in software. Using encapsulation tunnels
   avoided that collapse by allowing the forwarding of encapsulated
   packets to use the "fast path" hardware processing [Er94].

   The remainder of this document describes the general principles of IP
   tunneling and discusses the key considerations in the design of any
   protocol that tunnels IP datagrams. It derives its conclusions from
   the equivalence of tunnels and links and from requirements of
   existing standards for supporting IPv4 and IPv6 as payloads.

2. Conventions used in this document

2.1. Key Words

   The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
   document are to be interpreted as described in RFC-2119 [RFC2119].

   In this document, these key words will appear with that
   interpretation only when in ALL CAPS. Lower case uses of these words
   are not to be interpreted as carrying RFC-2119 significance.

2.2. Terminology

   This document uses the following terminology. Optional words in the
   term are indicated in parentheses, e.g., "(link or network)
   interface" or "egress (interface)".

   Terms from existing RFCs:

   o  Messages: variable length data labeled with globally-unique
      endpoint IDs, also known as a datagram for IP messages [RFC791].

Touch, Townsley        Expires December 8, 2017                [Page 6]

Internet-Draft         Tunnels in the Internet                June 2017

   o  Node: a physical or logical network device that participates as
      either a host [RFC1122][RFC6434] or router [RFC1812]. This term
      originally referred to gateways since some very early RFCs [RFC5],
      but is currently the common way to describe a point in a network
      at which messages are processed.

   o  Host or endpoint: a node that sources or sinks messages labeled
      from/to its IDs, typically known as a host for both IP and higher-
      layer protocol messages [RFC1122].

   o  Source or sender: the node that generates a message [RFC1122].

   o  Destination or receiver: the node that consumes a message

   o  Router or gateway: a node that relays IP messages using
      destination IDs and local context [RFC1812]. Routers also act as
      hosts when they source or sink messages. Also known as a forwarder
      for IP messages. Note that the notion of router is relative to the
      layer at which message processing is considered [To16].

   o  Link: a communications medium (or emulation thereof) that
      transfers IP messages between nodes without traversing a router
      (as would require decrementing the hop count) [RFC1122][RFC1812].

   o  Link packet: a link layer message, which can carry an IP datagram
      as a payload

   o  (Link or network) Interface: a location on a link co-located with
      a node where messages depart onto that link or arrive from that
      link. On physical links, this interface formats the message for
      transmission and interprets the received signals.

   o  Path: a sequence of one or more links over which an IP message
      traverses between source and destination nodes (hosts or routers).

   o  (Link) MTU: the largest message that can transit a link [RFC791],
      also often referred to simply as "MTU". It does not include the
      size of link-layer information, e.g., link layer headers or
      trailers, i.e., it refers to the message that the link can carry
      as a payload rather than the message as it appears on the link.
      This is thus the largest network layer packet (including network
      layer headers, e.g., IP datagram) that can transit a link. Note
      that this need not be the native size of messages on the link,
      i.e., the link may internally fragment and reassemble messages.
      For IPv4, the smallest MTU must be at least 68 bytes [RFC791], and
      for IPv6 the smallest MTU must be at least 1280 bytes [RFC2460].

Touch, Townsley        Expires December 8, 2017                [Page 7]

Internet-Draft         Tunnels in the Internet                June 2017

   o  EMTU_S (effective MTU for sending): the largest message that can
      transit a link, possibly also accounting for fragmentation that
      happens before the fragments are emitted onto the link [RFC1122].
      When source fragmentation is possible, EMTU_S = EMTU_R. When
      source fragmentation is not possible, EMTU_S = (link) MTU. For
      IPv4, this is MUST be at least 68 bytes [RFC791] and for IPv6 this
      MUST be at least 1280 bytes [RFC2460].

   o  EMTU_R (effective MTU to receive): the largest payload message
      that a receiver must be able to accept. This thus also represents
      the largest message that can traverse a link, taking into account
      reassembly at the receiver that happens after the fragments are
      received [RFC1122]. For IPv4, this is MUST be at least 576 bytes
      [RFC791] and for IPv6 this MUST be at least 1500 bytes [RFC2460].

   o  Path MTU (PMTU): the largest message that can transit a path of
      links [RFC1191][RFC1981]. Typically, this is the minimum of the
      link MTUs of the links of the path, and represents the largest
      network layer message (including network layer headers) that can
      transit a path without requiring fragmentation while in transit.
      Note that this is not the largest network packet that can be sent
      between a source and destination, because that network packet
      might have been fragmented at the network layer of the source and
      reassembled at the network layer of the destination.

   o  Tunnel: a protocol mechanism that transits messages between an
      ingress interface and egress interface using encapsulation to
      allow an existing network path to appear as a single link
      [RFC1853]. Note that a protocol can be used to tunnel itself (IP
      over IP). There is essentially no difference between a tunnel and
      the conventional layering of the ISO stack (i.e., by this
      definition, Ethernet is can be considered tunnel for IP). A tunnel
      is also known as a virtual link.

   o  Ingress (interface): the virtual link interface of a tunnel that
      receives messages within a node, encapsulates them according to
      the tunnel protocol, and transmits them into the tunnel [RFC2983].
      An ingress is the tunnel equivalent of the outgoing (departing)
      network interface of a link, and its encapsulation processing is
      the tunnel equivalent of encoding a message for transmission over
      a physical link. The ingress virtual link interface can be co-
      located with the traffic source.

      The term 'ingress' in other RFCs also refers to 'network ingress',
      which is the entry point of traffic to a transit network. Because
      this document focuses on tunnels, the term "ingress" used in the
      remainder of this document implies "tunnel ingress".

Touch, Townsley        Expires December 8, 2017                [Page 8]

Internet-Draft         Tunnels in the Internet                June 2017

   o  Egress (interface): a virtual link interface of a tunnel that
      receives messages that have finished transiting a tunnel and
      presents them to a node [RFC2983]. For reasons similar to ingress,
      the term 'egress' will refer to 'tunnel egress' throughout the
      remainder of this document. An egress is the tunnel equivalent of
      the incoming (arriving) network interface of a link and its
      decapsulation processing is the tunnel equivalent of interpreting
      a signal received from a physical link. The egress decapsulates
      messages for further transit to the destination. The egress
      virtual link interface can be co-located with the traffic

   o  Ingress node: network device on which an ingress is attached as a
      virtual link interface [RFC2983]. Note that a node can act as both
      an ingress node and an egress node at the same time, but typically
      only for different tunnels.

   o  Egress node: device where an egress is attached as a virtual link
      interface [RFC2983]. Note that a device can act as both a ingress
      node and an egress node at the same time, but typically only for
      different tunnels.

   o  Inner header: the header of the message as it arrives to the
      ingress [RFC2003].

   o  Outer header(s): one or more headers added to the message by the
      ingress, as part of the encapsulation for tunnel transit

   o  Mid-tunnel fragmentation: Fragmentation of the message during the
      tunnel transit, as could occur for IPv4 datagrams with DF=0

   o  Atomic packet, datagram, or fragment: an IP packet that has not
      been fragmented and which cannot be fragmented further [RFC6864]

   The following terms are introduced by this document:

   o  (Tunnel) transit packet: the packet arriving at a node connected
      to a tunnel that enters the ingress interface and exits the egress
      interface, i.e., the packet carried over the tunnel. This is
      sometimes known as the 'tunneled packet', i.e., the packet carried
      over the tunnel. This is the tunnel equivalent of a network layer
      packet as it would traverse a link. This document focuses on IPv4
      and IPv6 transit packets.

Touch, Townsley        Expires December 8, 2017                [Page 9]

Internet-Draft         Tunnels in the Internet                June 2017

   o  (Tunnel) link packet (TLP): packets that traverse between two
      interfaces, e.g., from ingress interface to egress interface, in
      which resides all or part of a transit packet. A tunnel link
      packet is the tunnel equivalent of a link (layer) packet as it
      would traverse a link, which is why we use the same terminology.

   o  Tunnel MTU: the largest transit packet that can traverse a tunnel,
      i.e., the tunnel equivalent of a link MTU, which is why we use the
      same terminology. This is the largest transit packet which can be
      reassembled at the egress interface.

   o  Tunnel maximum atomic packet (MAP): the largest transit packet
      that can traverse a tunnel as an atomic packet, i.e., without
      requiring tunnel link packet fragmentation either at the ingress
      or on-path between the ingress and egress.

   o  Inner fragmentation: fragmentation of the transit packet that
      arrives at the ingress interface before any additional headers are
      added. This can only correctly occur for IPv4 DF=0 datagrams.

   o  Outer fragmentation: source fragmentation of the tunnel link
      packet after encapsulation; this can involve fragmenting the
      outermost header or any of the other (if any) protocol layers
      involved in encapsulation.

   o  Maximum frame size (MFS): the link-layer equivalent of the MTU,
      using the OSI term 'frame'. For Ethernet, the MTU (network packet
      size) is 1500 bytes but the MFS (link frame size) is 1518 bytes
      originally, and 1522 bytes assuming VLAN (802.1Q) tagging support.

   o  EMFS_S: the link layer equivalent of EMTU_S.

   o  EMFS_R: the link layer equivalent of EMTU_R.

   o  Path MFS: the link layer equivalent of PMTU.

3. The Tunnel Model

   A network architecture is an abstract description of a distributed
   communications system, its components and their relationships, the
   requisite properties of those components and the emergent properties
   of the system that result [To03]. Such descriptions can help explain
   behavior, as when the OSI seven-layer model is used as a teaching
   example [Zi80]. Architectures describe capabilities - and, just as
   importantly, constraints.

Touch, Townsley        Expires December 8, 2017               [Page 10]

Internet-Draft         Tunnels in the Internet                June 2017

   A network can be defined as a system of endpoints and relays
   interconnected by communication paths, abstracting away issues of
   naming in order to focus on message forwarding. To the extent that
   the Internet has a single, coherent interpretation, its architecture
   is defined by its core protocols (IP [RFC791], TCP [RFC793], UDP
   [RFC768]) whose messages are handled by hosts, routers, and links
   [Cl88][To03], as shown in Figure 3:

               +------+    ------      ------    +------+
               |      |   /      \    /      \   |      |
               | HOST |--+ ROUTER +--+ ROUTER +--| HOST |
               |      |   \      /    \      /   |      |
               +------+    ------      ------    +------+

                   Figure 3 Basic Internet architecture

   As a network architecture, the Internet is a system of hosts
   (endpoints) and routers (relays) interconnected by links that
   exchange messages when possible. "When possible" defines the
   Internet's "best effort" principle. The limited role of routers and
   links represents the End-to-End Principle [Sa84] and longest-prefix
   match enables hierarchical forwarding using compact tables.

   Although the definitions of host, router, and link seem absolute,
   they are often relative as viewed within the context of one protocol
   layer, each of which can be considered a distinct network
   architecture. An Internet gateway is an OSI Layer 3 router when it
   transits IP datagrams but it acts as an OSI Layer 2 host as it
   sources or sinks Layer 2 messages on attached links to accomplish
   this transit capability. In this way, one device (Internet gateway)
   behaves as different components (router, host) at different layers.

   Even though a single device may have multiple roles - even
   concurrently - at a given layer, each role is typically static and
   determined by context. An Internet gateway always acts as a Layer 2
   host and that behavior does not depend on where the gateway is viewed
   from within Layer 2. In the context of a single layer, a device's
   behavior is typically modeled as a single component from all
   viewpoints in that layer (with some notable exceptions, e.g., Network
   Address Translators, which appear as hosts and routers, depending on
   the direction of the viewpoint [To16]).

3.1. What is a Tunnel?

   A tunnel can be modeled as a link in another network
   [To98][To01][To03]. In Figure 4, a source host (Hsrc) and destination
   host (Hdst) communicating over a network M in which two routers (Ra

Touch, Townsley        Expires December 8, 2017               [Page 11]

Internet-Draft         Tunnels in the Internet                June 2017

   and Rd) are connected by a tunnel. Keep in mind that it is possible
   that both network N and network M can both be components of the
   Internet, i.e., there may be regular traffic as well as tunneled
   traffic over any of the routers shown.

                     --_                         --
         +------+   /  \                        /  \   +------+
         | Hsrc |--+ Ra +      --      --      + Rd +--| Hdst |
         +------+   \  //\    /  \    /  \    /\\  /   +------+
                     --/I \--+ Rb +--+ Rc +--/E \--
                       \  /   \  /    \  /   \  /
                        \/     --      --     \/
                       <------ Network N ------->
         <-------------------- Network M --------------------->

                         Figure 4 The big picture

   The tunnel consists of two interfaces - an ingress (I) and an egress
   (E) that lie along a path connected by network N. Regardless of how
   the ingress and egress interfaces are connected, the tunnel serves as
   a link between the nodes it connects (here, Ra and Rd).

   IP packets arriving at the ingress interface are encapsulated to
   traverse network N. We call these packets 'tunnel transit packets'
   (or just 'transit packets') because they will transit the tunnel
   inside one or more of what we call 'tunnel link packets'. Transit
   packets correspond to network (IP) packets traversing a conventional
   link and tunnel link packets correspond to the packets of a
   conventional link layer (which can be called just 'link packets').

   Link packets use the source address of the ingress interface and the
   destination address of the egress interface - using whatever address
   is appropriate to the Layer at which the ingress and egress
   interfaces operate (Layer 2, Layer 3, Layer 4, etc.). The egress
   interface decapsulates those messages, which then continue on network
   M as if emerging from a link. To transit packets and to the routers
   the tunnel connects (Ra and Rd), the tunnel acts as a link and the
   ingress and egress interfaces act as network interfaces to that link.

   The model of each component (ingress and egress interfaces) and the
   entire system (tunnel) depends on the layer from which they are
   viewed. From the perspective of the outermost hosts (Hsrc and Hdst),
   the tunnel appears as a link between two routers (Ra and Rd). For
   routers along the tunnel (e.g., Rb and Rc), the ingress and egress
   interfaces appear as the endpoint hosts on network N.

Touch, Townsley        Expires December 8, 2017               [Page 12]

Internet-Draft         Tunnels in the Internet                June 2017

   When the tunnel network (N) is implemented using the same protocol as
   the endpoint network (M), the picture looks flatter (Figure 5), as if
   it were running over a single network. However, this appearance is
   incorrect - nothing has changed from the previous case. From the
   perspective of the endpoints, Rb and Rc and network N don't exist and
   aren't visible, and from the perspective of the tunnel, network M
   doesn't exist. The fact that network N and M use the same protocol,
   and may traverse the same links is irrelevant.

                   --_         --      --          --
       +------+   /  \  /\    /  \    /  \    /\  /  \   +------+
       | Hsrc |--+ Ra +/I \--+ Rb +--+ Rc +--/E \+ Rd +--| Hdst |
       +------+   \  / \  /   \  /    \  /   \  / \  /   +------+
                   --   \/     --      --     \/   --
                         <---- Network N ----->
           <------------------ Network M ------------------->

                     Figure 5 IP in IP network picture

3.2. View from the Outside

   As already observed, from outside the tunnel, to network M, the
   entire tunnel acts as a link (Figure 6). Consequently all
   requirements for links supporting IP also apply to tunnels [RFC3819].

                   --_                             --
       +------+   /  \                            /  \   +------+
       | Hsrc |--+ Ra +--------------------------+ Rd +--| Hdst |
       +------+   \  /                            \  /   +------+
                   --                              --
           <------------------ Network M ------------------->

                Figure 6 Tunnels as viewed from the outside

   For example, the IP datagram hop counts (IPv4 Time-to-Live [RFC791]
   and IPv6 Hop Limit [RFC2460]) are decremented when traversing a
   router, but not when traversing a link - or thus a tunnel. Similarly,
   because the ingress and egress are interfaces on this outer network,
   they should never issue ICMP messages. A router or host would issue
   the appropriate ICMP, e.g., "packet too big" (IPv4 fragmentation
   needed and DF set [RFC792] or IPv6 packet too big [RFC4443]), when
   trying to send a packet to the egress, as it would for any interface.

   Tunnels have a tunnel MTU - the largest message that can transit that
   tunnel, just as links have a link MTU. This MTU may not reflect the
   native message size of hops within a multihop link (or tunnel) and

Touch, Townsley        Expires December 8, 2017               [Page 13]

Internet-Draft         Tunnels in the Internet                June 2017

   the same is true for a tunnel. In both cases, the MTU is defined by
   the link's (or tunnel's) effective MTU to receive (EMTU_R).

3.3. View from the Inside

   Within network N, i.e., from inside the tunnel itself, the ingress
   interface is a source of tunnel link packets and the egress interface
   is a sink - so both are viewed as hosts on network N (Figure 7).
   Consequently [RFC1122] Internet host requirements apply to ingress
   and egress interfaces when Network N uses IP (and thus the
   ingress/egress interfaces use IP encapsulation).

                   _           --      --
                        /\    /  \    /  \    /\
                       /I \--+ Rb +--+ Rc +--/E \
                       \  /   \  /    \  /   \  /
                        \/     --      --     \/
                         <---- Network N ----->

            Figure 7 Tunnels, as viewed from within the tunnel

   Viewed from within the tunnel, the outer network (M) doesn't exist.
   Tunnel link packets can be fragmented by the source (ingress
   interface) and reassembled at the destination (egress interface),
   just as at conventional hosts. The path between ingress and egress
   interfaces has a path MTU, but the endpoints can exchange messages as
   large as can be reassembled at the destination (egress interface),
   i.e., the EMTU_R of the egress interface. However, in both cases,
   these MTUs refer to the size of the message that can transit the
   links and between the hosts of network N, which represents a link
   layer to network M. I.e., the MTUs of network N represent the maximum
   frame sizes (MFSs) of the tunnel as a link in network M.

   Information about the network - i.e., regarding network N MTU sizes,
   network reachability, etc. - are relayed from the destination (egress
   interface) and intermediate routers back to the source (ingress
   interface), without regard for the external network (M). When such
   messages arrive at the ingress interface, they may affect the
   properties of that interface (e.g., its reported MTU to network M),
   but they should never directly cause new ICMPs in the outer network
   M. Again, events at interfaces don't generate ICMP messages; it would
   be the host or router at which that interface is attached that would
   generate ICMPs, e.g., upon attempting to use that interface.

Touch, Townsley        Expires December 8, 2017               [Page 14]

Internet-Draft         Tunnels in the Internet                June 2017

3.4. Location of the Ingress and Egress

   The ingress and egress interfaces are endpoints of the tunnel. Tunnel
   interfaces may be physical or virtual. The interface may be
   implemented inside the node where the tunnel attaches, e.g., inside a
   host or router. The interface may also be implemented as a "bump in
   the wire" (BITW), somewhere along a link between the two nodes the
   link interconnects. IP in IP tunnels are often implemented as
   interfaces on nodes, whereas IPsec tunnels are sometimes implemented
   as BITW. These implementation variations determine only whether
   information available at the link endpoints (ingress/egress
   interfaces) can be easily shared with the connected network nodes.

   An ingress or egress can be implemented as an integrated component,
   appearing equivalent to any other network interface, or can be more
   complex. In the simple variant, each is tightly coupled to another
   network interface, e.g., where the ingress emits encapsulated packets
   directly into another network interface, or where the egress receives
   packets to decapsulate directly from another network interface.

   The other implementation variant is more modular, but more complex to
   explain. The ingress acts like a network interface by receiving IP
   packets to transmit from an upper layer protocol (or relay mechanism
   of a router), but then acts like an upper layer protocol (or relay
   mechanism of a router) when it emits encapsulated packets back into
   the same node. The egress acts like an upper layer interface (or
   relay mechanism of a router) by receiving packets from a network
   interface, but then acts like a network interface when it emits
   decapsulated packets back in to the same node. To the existing
   network interfaces, the ingress/egress act like upper layer
   interfaces (i.e., sending or receiving application stacks), while to
   the interior of the node, the ingress/egress act like network
   interfaces. This dual nature inside the node reflects the duality of
   the tunnel as transit link and host-host channel.

3.5. Implications of This Model

   This approach highlights a few key features of a tunnel as a network
   architecture construct:

   o  To the transit packets, tunnels turn a network (Layer 3) path into
      a (Layer 2) link

   o  To nodes the tunnel traverses, the tunnel ingress and egress
      interfaces act as hosts that source and sink tunnel link packets

   The consequences of these features are as follow:

Touch, Townsley        Expires December 8, 2017               [Page 15]

Internet-Draft         Tunnels in the Internet                June 2017

   o  Like a link MTU, a tunnel MTU is defined by the effective MTU of
      the receiver (i.e., EMTU_R of the egress).

   o  The messages inside the tunnel are treated like any other link
      layer, i.e., the MTU is determined by the largest (transit)
      payload that traverses the link.

   o  The tunnel path MFS is not relevant to the transited traffic.
      There is no mechanism or protocol by which it can be determined.

   o  Because routers, not links, alter hop counts [RFC1812], hopcounts
      are not decremented solely by the transit of a tunnel. A packet
      with a hop count of zero should successfully transit a link (and
      thus a tunnel) that connects two hosts.

   o  The addresses of a tunnel ingress and egress interface correspond
      to link layer addresses to the transit packet. Like links, some
      tunnels may not have their own addresses. Like network interfaces,
      ingress and egress interfaces typically require network layer

   o  Like network interfaces, the ingress and egress interfaces are
      never a direct source of ICMP messages but may provide information
      to their attached host or router to generate those ICMP messages
      during the processing of transit packets.

   o  Like network interfaces and links, two nodes may be connected by
      any combination of tunnels and links, including multiple tunnels.
      As with multiple links, existing network layer forwarding
      determines which IP traffic uses each link or tunnel.

   These observations make it much easier to determine what a tunnel
   must do to transit IP packets, notably it must satisfy all
   requirements expected of a link [RFC1122][RFC3819]. The remainder of
   this document explores these implications in greater detail.

3.6. Fragmentation

   There are two places where fragmentation can occur in a tunnel,
   called 'outer fragmentation' and 'inner fragmentation'. This document
   assumes that only outer fragmentation is viable because it is the
   only approach that works for both IPv4 datagrams with DF=1 and IPv6.

3.6.1. Outer Fragmentation

   Outer fragmentation is shown in Figure 8. The bottom of the figure
   shows the network topology, where transit packets originate at the

Touch, Townsley        Expires December 8, 2017               [Page 16]

Internet-Draft         Tunnels in the Internet                June 2017

   source, enter the tunnel at the ingress interface for encapsulation,
   exit the tunnel at the egress interface where they are decapsulated,
   and arrive at the destination. The packet traffic is shown above the
   topology, where the transit packets are shown at the top. In this
   diagram, the ingress interface is located on router 'Ra' and the
   egress interface is located on router 'Rd'.

   When the link packet - which is the encapsulated transit packet -
   would exceed the tunnel MTU, the packet needs to be fragmented. In
   this case the packet is fragmented at the outer (link) header, with
   the fragments shown as (b1) and (b2). The outer header indicates
   fragmentation (as ' and "), the inner (transit) header occurs only in
   the first fragment, and the inner (transit) data is broken across the
   two packets. These fragments are reassembled at the egress interface
   during decapsulation in step (c), where the resulting link packet is
   reassembled and decapsulated so that the transit packet can continue
   on its way to the destination.

    Transit packet
    +----+----+                                              +----+----+
    | iH | iD |------+ -  -  -  -  -  -  -  -  -  -  +------>| iH | iD |
    +----+----+      |                               |       +----+----+
                     v Link packet                   |
              +----+----+----+               +----+----+----+
          (a) | oH | iH | iD |               | oH | iH | iD | (d)
              +----+----+----+               +----+----+----+
                     |                               ^
                     |    Link packet fragment #1    |
                     |       +----+----+-----+       |
                (b1) +----- >| oH'| iH | iD1 |-------+ (c)
                     |       +----+----+-----+       |
                     |                               |
                     |    Link packet fragment #2    |
                     |       +----+-----+            |
                (b2) +----- >| oH"| iD2 |------------+
   +-----+    +--+ +---+                           +---+ +--+    +-----+
   |     |    |  |/     \                         /     \|  |    |     |
   | Src |----|Ra|Ingress|=======================|Egress |Rd|----| Dst |
   |     |    |  |\     /                         \     /|  |    |     |
   +-----+    +--+ +---+                           +---+ +--+    +-----+

             Figure 8 Fragmentation of the (outer) link packet

   Outer fragmentation isolates the tunnel encapsulation duties to the
   ingress and egress interfaces. This can be considered a benefit in
   clean, layered network design, but also may require complex egress

Touch, Townsley        Expires December 8, 2017               [Page 17]

Internet-Draft         Tunnels in the Internet                June 2017

   interface decapsulation, especially where tunnels aggregate large
   amounts of traffic, such as may result in IP ID overload (see Sec.
   4.1.4). Outer fragmentation is valid for any tunnel link protocol
   that supports fragmentation (e.g., IPv4 or IPv6), in which the tunnel
   endpoints act as the host endpoints of that protocol.

   Along the tunnel, the inner (transit) header is contained only in the
   first fragment, which can interfere with mechanisms that 'peek' into
   lower layer headers, e.g., as for relayed ICMP (see Sec. 4.3).

3.6.2. Inner Fragmentation

   Inner fragmentation distributes the impact of tunnel fragmentation
   across both egress interface decapsulation and transit packet
   destination, as shown in Figure 9; this can be especially important
   when the tunnel would otherwise need to source (outer) fragment large
   amounts of traffic. However, this mechanism is valid only when the
   transit packets can be fragmented on-path, e.g., as when the transit
   packets are IPv4 datagrams with DF=0.

   Again, the network topology is shown at the bottom of the figure, and
   the original packets show at the top. Packets arrive at the ingress
   node (router Ra) and are fragmented there based into transit packet
   fragments #1 (a1) and #2 (a2). These fragments are encapsulated at
   the ingress interface in steps (b1) and (b2) and each resulting link
   packet traverses the tunnel. When these link packets arrive at the
   egress interface they are decapsulated in steps (c1) and (c2) and the
   egress node (router) forwards the transit packet fragments to their
   destination. This destination is then responsible for reassembling
   the transit packet fragments into the original transit packet (d).

   Along the tunnel, the inner headers are copied into each fragment,
   and so can be 'peeked at' inside the tunnel (see Sec. 4.3).
   Fragmentation shifts from the ingress interface to the ingress router
   and reassembly shifts from the egress interface to the destination.

Touch, Townsley        Expires December 8, 2017               [Page 18]

Internet-Draft         Tunnels in the Internet                June 2017

    Transit packet
   +----+----+                                               +----+----+
   | iH | iD |-+ - - - - -  -  -  -  -  -  -  -  -  -  -  - >| iH | iD |
   +----+----+ |                                             +----+----+
               v Transit packet fragment #1                         ^
            +----+-----+                           +----+-----+     |
       (a1) | iH'| iD1 |                           | iH'| iD1 |-----+(d)
            +----+-----+                           +----+-----+     ^
               |     |        Link packet #1         ^              |
               |     |       +----+----+-----        |              |
               | (b1)+----- >| oH | iH'| iD1 |-------+(c1)          |
               |             +----+----+-----+                      |
               |                                                    |
               v Transit packet fragment #2                         |
            +----+-----+                           +----+-----+     |
       (a2) | iH"| iD2 |                           | iH"| iD2 |-----+
            +----+-----+                           +----+-----+
                     |        Link packet #2         |
                     |       +----+----+-----+       |
                 (b2)+----- >| oH | iH"| iD2 |-------+(c2)
   +-----+    +--+ +---+                           +---+ +--+    +-----+
   |     |    |  |/     \                         /     \|  |    |     |
   | Src |----|Ra|Ingress|=======================|Egress |Rd|----| Dst |
   |     |    |  |\     /                         \     /|  |    |     |
   +-----+    +--+ +---+                           +---+ +--+    +-----+

           Figure 9 Fragmentation of the inner (transit) packet

3.6.3. The Necessity of Outer Fragmentation

   Fragmentation is critical for tunnels that support transit packets
   for protocols with minimum MTU requirements, while operating over
   tunnel paths using protocols that have their own MTU requirements.
   Depending on the amount of space used by encapsulation, these two
   minimums will ultimately interfere (especially when a protocol
   transits itself either directly, as with IP-in-IP, or indirectly, as
   in IP-in-GRE-in-IP), and the transit packet will need to be
   fragmented to both support a tunnel MTU while traversing tunnels with
   their own tunnel path MTUs.

   Outer fragmentation is the only solution that supports all IPv4 and
   IPv6 traffic, because inner fragmentation is allowed only for IPv4
   datagrams with DF=0.

Touch, Townsley        Expires December 8, 2017               [Page 19]

Internet-Draft         Tunnels in the Internet                June 2017

4. IP Tunnel Requirements

   The requirements of an IP tunnel are defined by the requirements of
   an IP link because both transit IP packets. A tunnel thus must
   transit the IP minimum MTU, i.e., 68 bytes for IPv4 [RFC793] and 1280
   bytes for IPv6 [RFC2460] and a tunnel must support address resolution
   when there is more than one egress interface for that tunnel.

   The requirements of the tunnel ingress and egress interfaces are
   defined by the network over which they exchange messages (link
   packets). For IP-over-IP, this means that the ingress interface MUST
   NOT exceed the IP fragment identification field uniqueness
   requirements [RFC6864]. Uniqueness is more difficult to maintain at
   high packet rates for IPv4, whose fragment ID field is only 16 bits.

   These requirements remain even though tunnels have some unique
   issues, including the need for additional space for encapsulation
   headers and the potential for tunnel MTU variation.

4.1. Encapsulation Header Issues

   Tunnel encapsulation uses a non-link protocol as a link layer. The
   encapsulation layer thus has the same requirements and expectations
   as any other IP link layer when used to transit IP packets. These
   relationships are addressed in the following subsections.

4.1.1. General Principles of Header Fields Relationships

   Some tunnel specifications attempt to relate the header fields of the
   transit packet and tunnel link packet. In some cases, this
   relationship is warranted, whereas in other cases the two protocol
   layers need to be isolated from each other. For example, the tunnel
   link header source and destination addresses are network endpoints in
   the tunnel network N, but have no meaning in the outer network M. The
   two sets of addresses are effectively independent, just as are other
   network and link addresses.

   Because the tunneled packet uses source and destination addresses
   with a separate meaning, it is inappropriate to copy or reuse the
   IPv4 Identification (ID) or IPv6 Fragment ID fields of the tunnel
   transit packet (see Section 4.1.4). Similarly, the DF field of the
   transit packet is not related to that field in the tunnel link packet
   header (presuming both are IPv4) (see Section 4.2). Most other fields
   are similarly independent between the transit packet and tunnel link
   packet. When a field value is generated in the encapsulation header,
   its meaning should be derived from what is desired in the context of
   the tunnel as a link. When feedback is received from these fields,

Touch, Townsley        Expires December 8, 2017               [Page 20]

Internet-Draft         Tunnels in the Internet                June 2017

   they should be presented to the tunnel ingress and egress as if they
   were network interfaces. The behavior of the node where these
   interfaces attach should be identical to that of a conventional link.

   There are exceptions to this rule that are explicitly intended to
   relay signals from inside the tunnel to the network outside the
   tunnel, typically relevant only when the tunnel network N and the
   outer network M use the same network. These apply only when that
   coordination is defined, as with explicit congestion notification
   (ECN) [RFC6040] (see Section 4.3.2), and differentiated services code
   points (DSCPs) [RFC2983]. Equal-cost multipath routing may also
   affect how some encapsulation fields are set, including IPv6 flow
   labels [RFC6438] and source ports for transport protocols when used
   for tunnel encapsulation [RFC8085] (see Section 4.3.4).

4.1.2. Addressing Fields

   Tunnel ingresses and egresses have addresses associated with the
   encapsulation protocol. These addresses are the source and
   destination (respectively) of the encapsulated packet while
   traversing the tunnel network.

   Tunnels may or may not have addresses in the network whose traffic
   they transit (e.g., network M in Figure 4). In some cases, the tunnel
   is an unnumbered interface to a point-to-point virtual link. When the
   tunnel has multiple egresses, tunnel interfaces require separate
   addresses in network M.

   To see the effect of tunnel interface addresses, consider traffic
   sourced at router Ra in Figure 4. Even before being encapsulated by
   the ingress, traffic needs a source IP network address that belongs
   to the router. One option is to use an address associated with one of
   the other interfaces of the router [RFC1122]. Another option is to
   assign a number to the tunnel interface itself. Regardless of which
   address is used, the resulting IP packet is then encapsulated by the
   tunnel ingress using the ingress address as a separate operation.

4.1.3. Hop Count Fields

   The Internet hop count field is used to detect and avoid forwarding
   loops that cannot be corrected without a synchronized reboot. The
   IPv4 Time-to-Live (TTL) and IPv6 Hop Limit field each serve this
   purpose [RFC791][RFC2460]. The IPv4 TTL field was originally intended
   to indicate packet expiration time, measured in seconds. A router is
   required to decrement the TTL by at least one or the number of
   seconds the packet is delayed, whichever is larger [RFC1812]. Packets
   are rarely held that long, and so the field has come to represent the

Touch, Townsley        Expires December 8, 2017               [Page 21]

Internet-Draft         Tunnels in the Internet                June 2017

   count of the number of routers traversed. IPv6 makes this meaning
   more explicit.

   These hop count fields represent the number of network forwarding
   elements (routers) traversed by an IP datagram. An IP datagram with a
   hop count of zero can traverse a link between two hosts because it
   never visits a router (where it would need to be decremented and
   would have been dropped).

   An IP datagram traversing a tunnel thus need not have its hop count
   modified, i.e., the tunnel transit header need not be affected. A
   zero hop count datagram should be able to traverse a tunnel as easily
   as it traverses a link. A router MAY be configured to decrement
   packets traversing a particular link (and thus a tunnel), which may
   be useful in emulating a tunnel path as if it were a network path
   that traversed one or more routers, but this is strictly optional.
   The ability of the outer network M and tunnel network N to avoid
   indefinitely looping packets does not rely on the hop counts of the
   transit packet and tunnel link packet being related.

   The hop count field is also used by several protocols to determine
   whether endpoints are 'local', i.e., connected to the same subnet
   (link-local discovery and related protocols [RFC4861]). A tunnel is a
   way to make a remote network address appear directly-connected, so it
   makes sense that the other ends of the tunnel appear local and that
   such link-local protocols operate over tunnels unless configured
   explicitly otherwise. When the interfaces of a tunnel are numbered,
   these can be interpreted the same way as if they were on the same
   link subnet.

4.1.4. IP Fragment Identification Fields

   Both IPv4 and IPv6 include an IP Identification (ID) field to support
   IP datagram fragmentation and reassembly [RFC791][RFC1122][RFC2460].
   When used, the ID field is intended to be unique for every packet for
   a given source address, destination address, and protocol, such that
   it does not repeat within the Maximum Segment Lifetime (MSL).

   For IPv4, this field is in the default header and is meaningful only
   when either source fragmented or DF=0 ("non-atomic packets")
   [RFC6864]. For IPv6, this field is contained in the optional Fragment
   Header [RFC2460]. Although IPv6 supports only source fragmentation,
   the field may occur in atomic fragments [RFC6946].

   Although the ID field was originally intended for fragmentation and
   reassembly, it can also be used to detect and discard duplicate
   packets, e.g., at congested routers (see Sec. of [RFC1122]).

Touch, Townsley        Expires December 8, 2017               [Page 22]

Internet-Draft         Tunnels in the Internet                June 2017

   For this reason, and because IPv4 packets can be fragmented anywhere
   along a path, all non-atomic IPv4 packets and all IPv6 packets
   between a source and destination of a given protocol must have unique
   ID values over the potential fragment reordering period

   The uniqueness of the IP ID is a known problem for high speed nodes,
   because it limits the speed of a single protocol between two
   endpoints [RFC4963]. Although this RFC suggests that the uniqueness
   of the IP ID is moot, tunnels exacerbate this condition. A tunnel
   often aggregates traffic from a number of different source and
   destination addresses, of different protocols, and encapsulates them
   in a header with the same ingress and egress addresses, all using a
   single encapsulation protocol. If the ingress enforces IP ID
   uniqueness, this can either severely limit tunnel throughput or can
   require substantial resources; the alternative is to ignore IP ID
   uniqueness and risk reassembly errors. Although fragmentation is
   somewhat rare in the current Internet at large, it can be common
   along a tunnel. Reassembly errors are not always detected by other
   protocol layers (see Sec. 4.3.3) , and even when detected they can
   result in excessive overall packet loss and can waste bandwidth
   between the egress and ultimate packet destination.

   The 32-bit IPv6 ID field in the Fragment Header is typically used
   only during source fragmentation. The size of the ID field is
   typically sufficient that a single counter can be used at the tunnel
   ingress, regardless of the endpoint addresses or next-header
   protocol, allowing efficient support for very high throughput

   The smaller 16-bit IPv4 ID is more difficult to correctly support. A
   recent update to IPv4 allows the ID to be repeated for atomic packets
   [RFC6864]. When either source fragmentation or on-path fragmentation
   is supported, the tunnel ingress may need to keep independent ID
   counters for each tunnel source/destination/protocol tuple.

4.1.5. Checksums

   IP traffic transiting a tunnel needs to expect a similar level of
   error detection and correction as it would expect from any other
   link. In the case of IPv4, there are no such expectations, which is
   partly why it includes a header checksum [RFC791].

   IPv6 omitted the header checksum because it already expects most link
   errors to be detected and dropped by the link layer and because it
   also assumes transport protection [RFC2460]. When transiting IPv6
   over IPv6, the tunnel fails to provide the expected error detection.

Touch, Townsley        Expires December 8, 2017               [Page 23]

Internet-Draft         Tunnels in the Internet                June 2017

   This is why IPv6 is often tunneled over layers that include separate
   protection, such as GRE [RFC2784].

   The fragmentation created by the tunnel ingress can increase the need
   for stronger error detection and correction, especially at the tunnel
   egress to avoid reassembly errors. The Internet checksum is known to
   be susceptible to reassembly errors that could be common [RFC4963],
   and should not be relied upon for this purpose. This is why some
   tunnel protocols, e.g., SEAL and AERO [RFC5320][Te17] and GRE
   [RFC2784] as well as legacy protocols swIPe and the Internet
   Encapsulation Protocol [RFC1853], include a separate checksum. This
   requirement can be undermined when using UDP as a tunnel with no UDP
   checksum (as per [RFC6935][RFC6936]) when fragmentation occurs
   because the egress has no checksum with which to validate reassembly.
   For this reason, it is safe to use UDP with a zero checksum for
   atomic tunnel link packets only; when used on fragments, whether
   generated at the ingress or en-route inside the tunnel, omission of
   such a checksum can result in reassembly errors that can cause
   additional work (capacity, forwarding processing, receiver
   processing) downstream of the egress.

4.2. MTU Issues

   Link MTUs, IP datagram limits, and transport protocol segment sizes
   are already related by several requirements
   [RFC768][RFC791][RFC1122][RFC1812][RFC2460] and by a variety of
   protocol mechanisms that attempt to establish relationships between
   them, including path MTU discovery (PMTUD) [RFC1191][RFC1981],
   packetization layer path MTU discovery (PLMTUD) [RFC4821], as well as
   mechanisms inside transport protocols [RFC793][RFC4340][RFC4960]. The
   following subsections summarize the interactions between tunnels and
   MTU issues, including minimum tunnel MTUs, tunnel fragmentation and
   reassembly, and MTU discovery.

4.2.1. Minimum MTU Considerations

   There are a variety of values of minimum MTU values to consider, both
   in a conventional network and in a tunnel as a link in that network.
   These are indicated in Figure 10, an annotated variant of Figure 4.
   Note that a (link) MTU (a) corresponds to a tunnel MTU (d) and that a
   path MTU (b) corresponds to a tunnel path MTU (e). The tunnel MTU is
   the EMTU_R of the egress interface, because that defines the largest
   transit packet message that can traverse the tunnel as a link in
   network M. The ability to traverse the hops of the tunnel - in
   network N - is not related, and only the ingress need be concerned
   with that value.

Touch, Townsley        Expires December 8, 2017               [Page 24]

Internet-Draft         Tunnels in the Internet                June 2017

                    --_                            --
        +------+   /  \                           /  \   +------+
        | Hsrc |--+ Ra +       --       --       + Rd +--| Hdst |
        +------+   \  //\     /  \     /  \     /\\  /   +------+
                    --/I \---+ Rb +---+ Rc +---/E \--
                      \  /    \  /     \  /    \  /
                       \/      --       --      \/
                        <----- Network N ------->
         <-------------------- Network M --------------------->

   Communication in network M viewed at that layer:
    (a)         <->          Link MTU
    (b)                <---- Tunnel MTU --------->
    (c)         <----------- Path MTU ----------------->
    (d) <------------------- EMTU_R --------------------------->

   Communication in network N viewed at that layer:
    (e)                   <--> Link MTU
    (f)                   <--- Path MTU ------>
    (g)                 <----- EMTU_R --------->

   Communication in network N viewed from network M:
    (h)                   <--> MFS
    (i)                   <--- Path MFS ------>
    (j)                 <----- EMFS_R --------->

                    Figure 10 The variety of MTU values

   Consider the following example values. For IPv6 transit packets, the
   minimum (link) MTU (a) is 1280 bytes, which similarly applies to
   tunnels as the tunnel MTU (b). The path MTU (c) is the minimum of the
   links (including tunnels as links) along a path, and indicates the
   smallest IP message (packet or fragment) that can traverse a path
   between a source and destination without on-path fragmentation (e.g.,
   supported in IPv4 with DF=0). Path MTU discovery, either at the
   network layer (PMTUD [RFC1191][RFC1981]) or packetization layer
   (PLPMTUD [RFC4821]) attempts to tune the source IP packets and
   fragments (i.e., EMTU_S) to fit within this path MTU size to avoid
   fragmentation and reassembly [Ke95]. The minimum EMTU_R (d) is 1500
   bytes, i.e., the minimum MTU for endpoint-to-endpoint communication.

   The tunnel is a source-destination communication in network N.
   Messages between the tunnel source (the ingress interface) and tunnel
   destination (egress interface) similarly experience a variety of
   network N MTU values, including a link MTU (e), a path MTU (f), and
   an EMTU_R (g). The network N message maximum is limited by the path
   MTU, and the source-destination message maximum (EMTU_S) is limited

Touch, Townsley        Expires December 8, 2017               [Page 25]

Internet-Draft         Tunnels in the Internet                June 2017

   by the path MTU when source fragmentation is disabled and by EMTU_R
   otherwise, just as it was in for those types of MTUs in network M.
   For an IPv6 network N, its link and path MTUs must be at least 1280
   and its EMTU_R must be at least 1500.

   However, viewed from the context of network M, these network N MTUs
   are link layer properties, i.e., maximum frame sizes (MFS (h)). The
   network N EMTU_R determines the largest message that can transit
   between the source (ingress) and destination (egress), but viewed
   from network M this is a link layer, i.e., EMFS_R (j). The tunnel
   EMTU_R is EMFS_R minus the link (encapsulation) headers and includes
   the encapsulation headers of the link layer. Just as the path MTU has
   no bearing on EMTU_R, the path MFS (i) in network N has no bearing on
   the MTU of the tunnel.

   For IPv6 networks M and N, these relationships are summarized as

   o  Network M MTU = 1280, the largest transit packet (i.e., payload)
      over a single IPv6 link in the base network without source

   o  Network M path MTU = 1280, the transit packet (i.e., payload) that
      can traverse a path of links in the base network without source

   o  Network M EMTU_R = 1500, the largest transit packet (i.e.,
      payload) that can traverse a path in the base network with source

   o  Network N MTU = 1280 (for the same reasons as for network M)

   o  Network N path MTU = 1280 (for the same reasons as for network M)

   o  Network N EMTU_R = 1500 (for the same reasons as for network M)

   o  Tunnel MTU = 1500-encapsulation (typically 1460), the network N
      EMTU_R payload

   o  Tunnel MAP (maximum atomic packet) = largest network M message
      that transits a tunnel as an atomic packet using network N as a
      link layer: 1280-encapsulation, i.e., the network N path MTU
      payload (which is itself limited by the tunnel path MFS)

   The difference between the network N MTU and its treatment as a link
   layer in network M is the reason why the tunnel ingress interfaces
   need to support fragmentation and tunnel egress interfaces need to

Touch, Townsley        Expires December 8, 2017               [Page 26]

Internet-Draft         Tunnels in the Internet                June 2017

   support reassembly in the encapsulation layer(s). The high cost of
   fragmentation and reassembly is why it is useful for applications to
   avoid sending messages too close to the size of the tunnel path MTU
   [Ke95], although there is no signaling mechanism that can achieve
   this (see Section 4.2.3).

4.2.2. Fragmentation

   A tunnel interacts with fragmentation in two different ways. As a
   link in network M, transit packets might be fragmented before they
   reach the tunnel - i.e., in network M either during source
   fragmentation (if generated at the same node as the ingress
   interface) or forwarding fragmentation (for IPv4 DF=0 datagrams). In
   addition, link packets traversing inside the tunnel may require
   fragmentation by the ingress interface - i.e., source fragmentation
   by the ingress as a host in network N. These two fragmentation
   operations are no more related than are conventional IP fragmentation
   and ATM segmentation and reassembly; one occurs at the (transit)
   network layer, the other at the (virtual) link layer.

   Although many of these issues with tunnel fragmentation and MTU
   handling were discussed in [RFC4459], that document described a
   variety of alternatives as if they were independent. This document
   explains the combined approach that is necessary.

   Like any other link, an IPv4 tunnel must transit 68 byte packets
   without requiring source fragmentation [RFC791][RFC1122] and an IPv6
   tunnel must transit 1280 byte packets without requiring source
   fragmentation [RFC2460]. The tunnel MTU interacts with routers or
   hosts it connects the same way as would any other link MTU. The
   pseudocode examples in this section use the following values:

   o  TP: transit packet

   o  TLP: tunnel link packet

   o  TPsize: size of the transit packet (including its headers)

   o  encaps: ingress encapsulation overhead (tunnel link headers)

   o  tunMTU: tunnel MTU, i.e., network N egress EMTU_R - encaps

   o  tunMAP: tunnel maximum atomic packet as limited by the tunnel path

Touch, Townsley        Expires December 8, 2017               [Page 27]

Internet-Draft         Tunnels in the Internet                June 2017

   These rules apply at the host/router where the tunnel is attached,
   i.e., at the network layer of the transit packet (we assume that all
   tunnels, including multipoint tunnels, have a single, uniform MTU).
   These are basic source fragmentation rules (or transit
   refragmentation for IPv4 DF=0 datagrams), and have no relation to the
   tunnel itself other than to consider the tunnel MTU as the effective
   link MTU of the next hop.

   Inside the source during transit packet generation or a router during
   transit packet forwarding, the tunnel is treated as if it were any
   other link (i.e., this is not tunnel processing, but rather typical
   source or router processing), as indicated in the pseudocode in
   Figure 11.

      if (TPsize > tunMTU) then
         if (TP can be on-path fragmented, e.g., IPv4 DF=0) then
            split TP into TP fragments of tunMTU size
            and send each TP fragment to the tunnel ingress interface
            drop the TP and send ICMP "too big" to the TP source
         send TP to the tunnel ingress (i.e., as an outbound interface)

         Figure 11 Router / host packet size processing algorithm

   The tunnel ingress acts as host on the tunnel path, i.e., as source
   fragmentation of tunnel link packets (we assume that all tunnels,
   even multipoint tunnels, have a single, uniform tunnel MTU), using
   the pseudocode shown in Figure 12. Note that ingress source
   fragmentation occurs in the encapsulation process, which may involve
   more than one protocol layer. In those cases, fragmentation can occur
   at any of the layers of encapsulation in which it is supported, based
   on the configuration of the ingress.

      if (TPsize <= tunMAP) then
         encapsulate the TP and emit
         if (tunMAP < TPsize) then
            encapsulate the TP, creating the TLP
            fragment the TLP into tunMAP chunks
            emit the TLP fragments

                  Figure 12 Ingress processing algorithm

Touch, Townsley        Expires December 8, 2017               [Page 28]

Internet-Draft         Tunnels in the Internet                June 2017

   Note that these Figure 11 and Figure 12 indicate that a node might
   both "fragment then encapsulate" and "encapsulate then fragment",
   i.e., the effect is "on-path fragment, then encapsulate, then source
   fragment". The first (on-path) fragmentation occurs only for IPv4
   DF=0 packets, based on the tunnel MTU. The second (source)
   fragmentation occurs for all packets, based on the tunnel maximum
   atomic packet (MAP) size. The first fragmentation is a convenience
   for a subset of IPv4 packets; it is the second (source) fragmentation
   that ensures that messages traverse the tunnel.

   Just as a network interface should never receive a message larger
   than its MTU, a tunnel should never receive a message larger than its
   tunnel MTU limit (see the host/router processing above). A router
   attempting to process such a message would already have generated an
   ICMP "packet too big" and the transit packet would have been dropped
   before entering into this algorithm. Similarly, a host would have
   generated an error internally and aborted the attempted transmission.

   As an example, consider IPv4 over IPv6 or IPv6 over IPv6 tunneling,
   where IPv6 encapsulation adds a 40 byte fixed header plus IPv6
   options (i.e., IPv6 header extensions) of total size 'EHsize'. The
   tunnel MTU will be at least 1500 - (40 + EHsize) bytes. The tunnel
   path MTU will be at least 1280 - (40 + EHsize) bytes, which then also
   represents the tunnel maximum atomic packet size (MAP). Transit
   packets larger than the tunnel MTU will be dropped by a node before
   ingress processing, and so do not need to be addressed as part of
   ingress processing. Considering these minimum values, the previous
   algorithm uses actual values shown in the pseudocode in Figure 13.

      if (TPsize <= (1240 - EHsize)) then
         encapsulate TP and emit
         if ((1240 - EHsize) < TPsize) then
            encapsulate the TP, creating the TLP
            fragment the TLP into (1240 - EHsize) chunks
            emit the TLP fragments

           Figure 13 Ingress processing for an tunnel over IPv6

   IPv6 cannot necessarily support all tunnel encapsulations. When the
   egress EMTU_R is the default of 1500 bytes, an IPv6 tunnel supports
   IPv6 transit only if EHsize is 180 bytes or less; otherwise the
   incoming transit packet would have been dropped as being too large by
   the host/router. Under the same EMTU_R assumption, an IPv6 tunnel
   supports IPv4 transit only if EHsize is 884 bytes or less. In this

Touch, Townsley        Expires December 8, 2017               [Page 29]

Internet-Draft         Tunnels in the Internet                June 2017

   example, transit packets of up to (1240 - Ehsize) can traverse the
   tunnel without ingress source fragmentation and egress reassembly.

   When using IP directly over IP, the minimum transit packet EMTU_R for
   IPv4 is 576 bytes and for IPv6 is 1500 bytes. This means that tunnels
   of IPv4-over-IPv4, IPv4-over-IPv6, and IPv6-over-IPv6 are possible
   without additional requirements, but this may involve ingress
   fragmentation and egress reassembly. IPv6 cannot be tunneled directly
   over IPv4 without additional requirements, notably that the egress
   EMTU_R is at least 1280 bytes.

   When ongoing ingress fragmentation and egress reassembly would be
   prohibitive or costly, larger MTUs can be supported by design and
   confirmed either out-of-band (by design) or in-band (e.g., using
   PLPMTUD [RFC4821], as done in SEAL [RFC5320] and AERO [Te17]). In
   particular, many tunnel specifications are often able to avoid
   persistent fragmentation because they operationally assume larger
   EMTU_R and tunnel MAP sizes than are guaranteed for IPv4 [RFC1122] or
   IPv6 [RFC2460].

4.2.3. Path MTU Discovery

   Path MTU discovery (PMTUD) enables a network path to support a larger
   PMTU than it can assume from the minimum requirements of protocol
   over which it operates. Note, however, that PMTUD never discovers
   EMTU_R that is larger than the required minimum; that information is
   available to some upper layer protocols, such as TCP [RFC1122], but
   cannot be determined at the IP layer.

   There is temptation to optimize tunnel traversal so that packets are
   not fragmented between ingress and egress, i.e., to attempt tune the
   network M PMTU to the tunnel MAP size rather than to the tunnel MTU,
   to avoid ingress fragmentation. This is often impossible because the
   ICMP "packet too big" message (IPv4 fragmentation needed [RFC792] or
   IPv6 packet too big [RFC4443]) indicates the complete failure of a
   link to transit a packet, not a preference for a size that matches
   that internal the mechanism of the link. ICMP messages are intended
   to indicate whether a tunnel MTU is insufficient; there is no ICMP
   message that can indicate when a transit packet is "too big for the
   tunnel path MTU, but not larger than the tunnel MTU". If there were,
   endpoints might receive that message for IP packets larger than 40
   bytes (the payload of a single ATM cell, allowing for the 8-byte AAL5
   trailer), but smaller than 9K (the ATM EMTU_R payload).

   In addition, attempting to try to tune the network transit size to
   natively match that of the link internal transit can be hazardous for
   many reasons:

Touch, Townsley        Expires December 8, 2017               [Page 30]

Internet-Draft         Tunnels in the Internet                June 2017

   o  The tunnel is capable of transiting packets as large as the
      network N EMTU_R - encapsulation, which is always at least as
      large as the tunnel MTU and typically is larger.

   o  ICMP has only one type of error message regarding large packets -
      "too big", i.e., too large to transit. There is no optimization
      message of "bigger than I'd like, but I can deal with if needed".

   o  IP tunnels often involve some level of recursion, i.e.,
      encapsulation over itself [RFC4459].

   Tunnels that use IPv4 as the encapsulation layer SHOULD set DF=0, but
   this requires generating unique fragmentation ID values, which may
   limit throughput [RFC6864]. These tunnels might have difficulty
   assuming ingress EMTU_S values over 64 bytes, so it may not be
   feasible to assume that larger packets with DF=1 are safe.

   Recursive tunneling occurs whenever a protocol ends up encapsulated
   in itself. This happens directly, as when IPv4 is encapsulated in
   IPv4, or indirectly, as when IP is encapsulated in UDP which then is
   a payload inside IP. It can involve many layers of encapsulation
   because a tunnel provider isn't always aware of whether the packets
   it transits are already tunneled.

   Recursion is impossible when the tunnel transit packets are limited
   to that of the native size of the ingress payload. Arriving tunnel
   transit packets have a minimum supported size (1280 for IPv6) and the
   tunnel PMFS has the same requirement; there would be no room for the
   tunnel's "link layer" headers, i.e., the encapsulation layer. The
   result would be an IPv6 tunnel that cannot satisfy IPv6 transit

   It is more appropriate to require the tunnel to satisfy IP transit
   requirements and enforce that requirement at design time or during
   operation (the latter using PLPMTUD [RFC4821]). Conventional path MTU
   discovery (PMTUD) relies on existing endpoint ICMP processing of
   explicit negative feedback from routers along the path via "packet to
   big" ICMP packets in the reverse direction of the tunnel
   [RFC1191][RFC1981]. This technique is susceptible to the "black hole"
   phenomenon, in which the ICMP messages never return to the source due
   to policy-based filtering [RFC2923]. PLPMTUD requires a separate,
   direct control channel from the egress to the ingress that provides
   positive feedback; the direct channel is not blocked by policy
   filters and the positive feedback ensures fail-safe operation if
   feedback messages are lost [RFC4821].

Touch, Townsley        Expires December 8, 2017               [Page 31]

Internet-Draft         Tunnels in the Internet                June 2017

   PLPMTUD might require that the ingress consider the potential impact
   of multipath forwarding (see Section 4.3.4). In such cases, probes
   generated by the ingress might need to track different flows, e.g.,
   that might traverse different tunnel paths. Additionally,
   encapsulation might need to consider mechanisms to ensure that probes
   traverse the same path as their corresponding traffic, even when
   labeled as the same flow (e.g., using the IPv6 flow ID). In such
   cases, the transit packet and probe may need to be encrypted or
   encapsulated in an additional flow-based transport header, to avoid
   differential path traversal based on deep-packet inspection within
   the tunnel.

4.3. Coordination Issues

   IP tunnels interact with link layer signals and capabilities in a
   variety of ways. The following subsections address some key issues of
   these interactions. In general, they are again informed by treating a
   tunnel as any other link layer and considering the interactions
   between the IP layer and link layers [RFC3819].

4.3.1. Signaling

   In the current Internet architecture, signaling goes upstream, either
   from routers along a path or from the destination, back toward the
   source. Such signals are typically contained in ICMP messages, but
   can involve other protocols such as RSVP, transport protocol signals
   (e.g., TCP RSTs), or multicast control or transport protocols.

   A tunnel behaves like a link and acts like a link interface at the
   nodes where it is attached. As such, it can provide information that
   enhances IP signaling (e.g., ICMP), but itself does not directly
   generate ICMP messages.

   For tunnels, this means that there are two separate signaling paths.
   The outer network M nodes can each signal the source of the tunnel
   transit packets, Hsrc (Figure 14). Inside the tunnel, the inner
   network N nodes can signal the source of the tunnel link packets, the
   ingress I (Figure 15).

Touch, Townsley        Expires December 8, 2017               [Page 32]

Internet-Draft         Tunnels in the Internet                June 2017

           |        |                           |        |
           v        --_                         --       v
        +------+   /  \                        /  \   +------+
        | Hsrc |--+ Ra +      --      --      + Rd +--| Hdst |
        +------+   \  //\    /  \    /  \    /\\  /   +------+
                    --/I \--+ Rb +--+ Rc +--/E \--
                      \  /   \  /    \  /   \  /
                       \/     --      --     \/
                        <---- Network N ----->
        <-------------------- Network M --------------------->

                   Figure 14 Signals outside the tunnel

                    --_ |     |       |      |  --
        +------+   /  \ v     |       |      | /  \   +------+
        | Hsrc |--+ Ra +      --      --      + Rd +--| Hdst |
        +------+   \  //\    /  \    /  \    /\\  /   +------+
                    --/I \--+ Rb +--+ Rc +--/E \--
                      \  /   \  /    \  /   \  /
                       \/     --      --     \/
                        <----- Network N ---->
        <--------------------- Network M -------------------->

                    Figure 15 Signals inside the tunnel

   These two signal paths are inherently distinct except where
   information is exchanged between the network interface of the tunnel
   (the ingress) and its attached node (Ra, in both figures).

   It is always possible for a network interface to provide hints to its
   attached node (host or router), which can be used for optimization.
   In this case, when signals inside the tunnel indicate a change to the
   tunnel, the ingress (i.e., the tunnel network interface) can provide
   information to the router (Ra, in both figures), so that Ra can
   generate the appropriate signal in return to Hsrc. This relaying may
   be difficult, because signals inside the tunnel may not return enough
   information to the ingress to support direct relaying to Hsrc.

   In all cases, the tunnel ingress needs to determine how to relay the
   signals from inside the tunnel into signals back to the source. For
   some protocols this is either simple or impossible (such as for
   ICMP), for others, it can even be undefined (e.g., multicast). In
   some cases, the individual signals relayed from inside the tunnel may
   result in corresponding signals in the outside network, and in other
   cases they may just change state of the tunnel interface. In the

Touch, Townsley        Expires December 8, 2017               [Page 33]

Internet-Draft         Tunnels in the Internet                June 2017

   latter case, the result may cause the router Ra to generate new ICMP
   errors when later messages arrive from Hsrc or other sources in the
   outer network.

   The meaning of the relayed information must be carefully translated.
   An ICMP error within a tunnel indicates a failure of the path inside
   the tunnel to support an egress atomic packet or packet fragment
   size. It can be very difficult to convert that ICMP error into a
   corresponding ICMP message from the ingress node back to the transit
   packet source. The ICMP message may not contain enough of a packet
   prefix to extract the transit packet header sufficient to generate
   the appropriate ICMP message. The relationship between the egress
   EMTU_R and the transit packet may be indirect, e.g., the ingress node
   may be performing source fragmentation that should be adjusted
   instead of propagating the ICMP upstream.

   Some messages have detailed specifications for relaying between the
   tunnel link packet and transit packet, including Explicit Congestion
   Notification (ECN [RFC6040]) and multicast (IGMP, e.g.).

4.3.2. Congestion

   Tunnels carrying IP traffic (i.e., the focus of this document) need
   not react directly to congestion any more than would any other link
   layer [RFC8085]. IP transit packet traffic is already expected to be
   congestion controlled.

   It is useful to relay network congestion notification between the
   tunnel link and the tunnel transit packets. Explicit congestion
   notification requires that ECN bits are copied from the tunnel
   transit packet to the tunnel link packet on encapsulation, as well as
   copied back at the egress based on a combination of the bits of the
   two headers [RFC6040]. This allows congestion notification within the
   tunnel to be interpreted as if it were on the direct path.

4.3.3. Multipoint Tunnels and Multicast

   Multipoint tunnels are tunnels with more than two ingress/egress
   endpoints [RFC2529][RFC5214][Te17]. Just as tunnels emulate links,
   multipoint tunnels emulate multipoint links, and can support
   multicast as a tunnel capability. Multipoint tunnels can be useful on
   their own, or may be used as part of more complex systems, e.g., LISP
   and TRILL configurations [RFC6830][RFC6325].

   Multipoint tunnels require a support for egress determination, just
   as multipoint links do. This function is typically supported by ARP
   [RFC826] or ARP emulation (e.g., LAN Emulation, known as LANE

Touch, Townsley        Expires December 8, 2017               [Page 34]

Internet-Draft         Tunnels in the Internet                June 2017

   [RFC2225]) for multipoint links. For multipoint tunnels, a similar
   mechanism is required for the same purpose - to determine the egress
   address for proper ingress encapsulation (e.g., LISP Map-Service

   All multipoint systems - tunnels and links - might support different
   MTUs between each ingress/egress (or link entrance/exit) pair. In
   most cases, it is simpler to assume a uniform MTU throughout the
   multipoint system, e.g., the minimum MTU supported across all
   ingress/egress pairs. This applies to both the ingress EMTU_S and
   egress EMTU_R (the latter determining the tunnel MTU). Values valid
   across all receivers need to be confirmed in advance (e.g., via IPv6
   ND announcements or out-of-band configuration information) before a
   multipoint tunnel or link can use values other than the default,
   otherwise packets may reach some receivers but be "black-holed" to
   others (e.g., if PMTUD fails [RFC2923]).

   A multipoint tunnel MUST have support for broadcast and multicast (or
   their equivalent), in exactly the same way as this is already
   required for multipoint links [RFC3819]. Both modes can be supported
   either by a native mechanism inside the tunnel or by emulation using
   serial replication at the tunnel ingress (e.g., AMT [RFC7450]), in
   the same way that links may provide the same support either natively
   (e.g., via promiscuous or automatic replication in the link itself)
   or network interface emulation (e.g., as for non-broadcast
   multiaccess networks, i.e., NBMAs).

   IGMP snooping enables IP multicast to be coupled with native link
   layer multicast support [RFC4541]. A similar technique may be
   relevant to couple transit packet multicast to tunnel link packet
   multicast, but the coupling of the protocols may be more complex
   because many tunnel link protocols rely on their own network N
   multicast control protocol, e.g., via PIM-SM [RFC6807][RFC7761].

4.3.4. Load Balancing

   Load balancing can impact the way in which a tunnel operates. In
   particular, multipath routing inside the tunnel can impact some of
   the tunnel parameters to vary, both over time and for different
   transit packets. The use of multiple paths can be the result of MPLS
   link aggregation groups (LAGs), equal-cost multipath routing (ECMP
   [RFC2991]), or other load balancing mechanisms. In some cases, the
   tunnel exists as the mechanism to support ECMP, as for GRE in UDP

   A tunnel may have multiple paths between the ingress and egress with
   different tunnel path MTU or tunnel MAP values, causing the ingress

Touch, Townsley        Expires December 8, 2017               [Page 35]

Internet-Draft         Tunnels in the Internet                June 2017

   EMTU_S to vary [RFC7690]. When individual values cannot be correlated
   to transit traffic, the EMTU_S can be set to the minimum of these
   different path MTU and MAP values.

   In some cases, these values can be correlated to paths, e.g., IPv6
   packets include a flow label to enable multipath routing to keep
   packets of a single flow following the same path, as well as to help
   differentiate path properties (e.g., for path MTU discovery
   [RFC4821]). It is important to preserve the semantics of that flow
   label as an aggregate identifier of the encapsulated link packets of
   a tunnel. This is achieved by hashing the transit IP addresses and
   flow label to generate a new flow label for use between the ingress
   and egress addresses [RFC6438]. It is not appropriate to simply copy
   the flow label from the transit packet into the link packet because
   of collisions that might arise if a label is used for flows between
   different transit packet addresses that traverse the same tunnel.

   When the transit packet is visible to forwarding nodes inside the
   tunnel (e.g., when it is not encrypted), those nodes use deep packet
   inspection (DPI) context to send a single flow over different paths.
   This sort of "DPI override" of the IP flow information can interfere
   with both PMTUD and PLPMTUD mechanisms. The only way to ensure that
   intermediate nodes do not interfere with PLPMTUD is to encrypt the
   transit packet when it is encapsulated for tunnel traversal, or to
   provide some other signals (e.g., an additional layer of
   encapsulation header including transport ports) that preserves the
   flow semantics.

4.3.5. Recursive Tunnels

   The rules described in this document already support tunnels over
   tunnels, sometimes known as "recursive" tunnels, in which IP is
   transited over IP either directly or via intermediate encapsulation
   (IP-UDP-IP, as in GUE [He16]).

   There are known hazards to recursive tunneling, notably that the
   independence of the tunnel transit header and tunnel link header hop
   counts can result in a tunneling loop. Such looping can be avoided
   when using direct encapsulation (IP in IP) by use of a header option
   to track the encapsulation count and to limit that count [RFC2473].
   This looping cannot be avoided when other protocols are used for
   tunneling, e.g., IP in UDP in IP, because the encapsulation count may
   not be visible where the recursion occurs.

Touch, Townsley        Expires December 8, 2017               [Page 36]

Internet-Draft         Tunnels in the Internet                June 2017

5. Observations

   The following subsections summarize the observations of this document
   and a summary of issues with existing tunnel protocol specifications.
   It also includes advice for tunnel protocol designers, implementers,
   and operators. It also includes

5.1. Summary of Recommendations

   o  Tunnel endpoints are network interfaces, tunnel are virtual links

       o ICMP messages MUST NOT be generated by the tunnel (as a link)

       o ICMP messages received by the ingress inside link change the
          link properties (they do not generate transit-layer ICMP

       o Link headers (hop, ID, options) are largely independent of
          arriving ID (with few exceptions based on translation, not
          direct copying, e.g., ECN and IPv6 flow IDs)

   o  MTU values should treat the tunnel as any other link

       o Require source ingress source fragmentation and egress
          reassembly at the tunnel link packet layer

       o The tunnel MTU is the tunnel egress EMTU_R less headers, and
          not related at all to the ingress-egress MFS

   o  Tunnels must obey core IP requirements

       o Obey IPv4 DF=1 on arrival at a node (nodes MUST NOT fragment
          IPv4 packets where DF=1 and routers MUST NOT clear the DF bit)

       o Shut down an IP tunnel if the tunnel MTU falls below the
          required minimum

5.2. Impact on Existing Encapsulation Protocols

   Many existing and proposed encapsulation protocols are inconsistent
   with the guidelines of this document. The following list summarizes
   only those inconsistencies, but omits places where a protocol is
   inconsistent solely by reference to another protocol.

   [should this be inverted as a table of issues and a list of which
   RFCs have problems?]

Touch, Townsley        Expires December 8, 2017               [Page 37]

Internet-Draft         Tunnels in the Internet                June 2017

   o  IP in IP / mobile IP [RFC2003][RFC4459] - IPv4 in IPv4

       o Sets link DF when transit DF=1 (fails without PLPMTUD)

       o Drops at egress if hopcount = 0 (host-host tunnels fail)

       o Drops based on transit source (same as router IP, matches
          egress), i.e., performs routing functions it should not

       o Ingress generates ICMP messages (based on relayed context),
          rather than using inner ICMP messages to set interface
          properties only

       o Treats tunnel MTU as tunnel path MTU, not tunnel egress MTU

   o  IPv6 tunnels [RFC2473] -- IPv6 or IPv4 in IPv6

       o Treats tunnel MTU as tunnel path MTU, not tunnel egress MTU

       o Decrements transiting packet hopcount (by 1)

       o Copies traffic class from tunnel link to tunnel transit header

       o Ignores IPv4 DF=0 and fragments at that layer upon arrival

       o Fails to retain soft ingress state based on inner ICMP messages
          affecting tunnel MTU

       o Tunnel ingress issues ICMPs

       o Fragments IPv4 over IPv6 fragments only if IPv4 DF=0
          (misinterpreting the "can fragment the IPv4 packet" as
          permission to fragment at the IPv6 link header)

   o  IPsec tunnel mode (IP in IPsec in IP) [RFC4301] -- IP in IPsec

       o Uses security policy to set, clear, or copy DF (rather than
          generating it independently, which would also be more secure)

       o Intertwines tunnel selection with security selection, rather
          than presenting tunnel as an interface and using existing
          forwarding (as with transport mode over IP-in-IP [RFC3884])

   o  GRE (IP in GRE in IP or IP in GRE in UDP in IP)

       o Treats tunnel MTU as tunnel path MTU, not tunnel egress MTU

Touch, Townsley        Expires December 8, 2017               [Page 38]

Internet-Draft         Tunnels in the Internet                June 2017

       o Requires ingress to generate ICMP errors

       o Copies IPv4 DF to outer IPv4 DF

       o Violates IPv6 MTU requirements when using IPv6 encapsulation

   o  LISP [RFC6830]

       o Treats tunnel MTU as tunnel path MTU, not tunnel egress MTU

       o Requires ingress to generate ICMP errors

       o Copies inner hop limit to outer

   o  L2TP [RFC3931]

       o Treats tunnel MTU as tunnel path MTU, not tunnel egress MTU

       o Requires ingress to generate ICMP errors

   o  PWE [RFC3985]

       o Treats tunnel MTU as tunnel path MTU, not tunnel egress MTU

       o Requires ingress to generate ICMP errors

   o  GUE (Generic UDP encapsulation) [He16] - IP (et. al) in UDP in IP

       o Allows inner encapsulation fragmentation

   o  Geneve [RFC7364][Gr17] - IP (et al.) in Geneve in UDP in IP

       o Treats tunnel MTU as tunnel path MTU, not tunnel egress MTU

   o  SEAL/AERO [RFC5320][Te17] - IP in SEAL/AERO in IP

       o Some issues with SEAL (MTU, ICMP), corrected in AERO

   o  RTG DT encapsulations [No16]

       o Assumes fragmentation can be avoided completely

       o Allows encapsulation protocols that lack fragmentation

       o Relies on ICMP PTB to correct for tunnel path MTU

   o  No known issues

Touch, Townsley        Expires December 8, 2017               [Page 39]

Internet-Draft         Tunnels in the Internet                June 2017

       o L2VPN (framework for L2 virtualization) [RFC4664]

       o L3VPN (framework for L3 virtualization) [RFC4176]

       o MPLS (IP in MPLS) [RFC3031]

       o TRILL (Ethernet in Ethernet) [RFC5556][RFC6325]

5.3. Tunnel Protocol Designers

   [To be completed]

   Recursive tunneling + minimum MTU = frag/reassembly is inevitable, at
   least to be able to split/join two fragments

   Account for egress MTU/path MTU differences.

   Include a stronger checksum.

   Ensure the egress MTU is always larger than the path MTU.

   Ensure that the egress reassembly can keep up with line rate OR
   design PLPMTUD into the tunneling protocol.

5.3.1. For Future Standards

   [To be completed]

   Larger IPv4 MTU (2K? or just 2x path MTU?) for reassembly

   Always include frag support for at least two frags; do NOT try to
   deprecate fragmentation.

   Limit encapsulation option use/space.

   Augment ICMP to have two separate messages: PTB vs P-bigger-than-

   Include MTU as part of BGP as a hint - SB

   Hazards of multi-MTU draft-van-beijnum-multi-mtu-04

5.3.2. Diagnostics

   [To be completed]

Touch, Townsley        Expires December 8, 2017               [Page 40]

Internet-Draft         Tunnels in the Internet                June 2017

   Some current implementations include diagnostics to support
   monitoring the impact of tunneling, especially the impact on
   fragmentation and reassembly resources, the status of path MTU
   discovery, etc.

   >> Because a tunnel ingress/egress is a network interface, it SHOULD
   have similar resources as any other network interface. This includes
   resources for packet processing as well as monitoring.

5.4. Tunnel Implementers

   [To be completed]

   Detect when the egress MTU is exceeded.

   Detect when the egress MTU drops below the required minimum and shut
   down the tunnel if that happens - configuring the tunnel down and
   issuing a hard error may be the only way to detect this anomaly, and
   it's sufficiently important that the tunnel SHOULD be disabled. This
   is always better than blindly assuming the tunnel has been deployed
   correctly, i.e., that the solution has been engineered.

   Do NOT decrement the TTL as part of being a tunnel. It's always
   already OK for a router to decrement the TTL based on different next-
   hop routers, but TTL is a property of a router not a link.

5.5. Tunnel Operators

   [To be completed]

   Keep the difference between "enforced by operators" vs. "enforced by
   active protocol mechanism" in mind. It's fine to assume something the
   tunnel cannot or does not test, as long as you KNOW you can assume
   it. When the assumption is wrong, it will NOT be signaled by the
   tunnel. Do NOT decrement the TTL as part of being a tunnel. It's
   always already OK for a router to decrement the TTL based on
   different next-hop routers, but TTL is a property of a router not a

   Consider the circuit breakers doc to provide diagnostics and last-
   resort control to avoid overload for non-reactive traffic (see
   Gorry's RFC-to-be)

   Do NOT decrement the TTL as part of being a tunnel. It's always
   already OK for a router to decrement the TTL based on different next-
   hop routers, but TTL is a property of a router not a link.

Touch, Townsley        Expires December 8, 2017               [Page 41]

Internet-Draft         Tunnels in the Internet                June 2017

   >>>> PLPMTUD can give multiple conflicting PMTU values during ECMP or
   LAG if PMTU is cached per endpoint pair rather than per flow -- but
   so can PMTUD! This is another reason why ICMP should never drive up
   the effective MTU (if aggregate, treat as the minimum of received
   messages over an interval).

6. Security Considerations

   Tunnels may introduce vulnerabilities or add to the potential for
   receiver overload and thus DOS attacks. These issues are primarily
   related to the fact that a tunnel is a link that traverses a network
   path and to fragmentation and reassembly. ICMP signal translation
   introduces a new security issue and must be done with care. ICMP
   generation at the router or host attached to a tunnel is already
   covered by existing requirements (e.g., should be throttled).

   Tunnels traverse multiple hops of a network path from ingress to
   egress. Traffic along such tunnels may be susceptible to on-path and
   off-path attacks, including fragment injection, reassembly buffer
   overload, and ICMP attacks. Some of these attacks may not be as
   visible to the endpoints of the architecture into which tunnels are
   deployed and these attacks may thus be more difficult to detect.

   Fragmentation at routers or hosts attached to tunnels may place an
   undue burden on receivers where traffic is not sufficiently diffuse,
   because tunnels may induce source fragmentation at hosts and path
   fragmentation (for IPv4 DF=0) more for tunnels than for other links.
   Care should be taken to avoid this situation, notably by ensuring
   that tunnel MTUs are not significantly different from other link

   Tunnel ingresses emitting IP datagrams MUST obey all existing IP
   requirements, such as the uniqueness of the IP ID field. Failure to
   either limit encapsulation traffic, or use additional ingress/egress
   IP addresses, can result in high speed traffic fragments being
   incorrectly reassembled.

   Tunnels are susceptible to attacks at both the inner and outer
   network layers. The tunnel ingress/egress endpoints appear as network
   interfaces in the outer network, and are as susceptible as any other
   network interface. This includes vulnerability to fragmentation
   reassembly overload, traffic overload, and spoofed ICMP messages that
   misreport the state of those interfaces. Similarly, the
   ingress/egress appear as hosts to the path traversed by the tunnel,
   and thus are as susceptible as any other host to attacks as well.


Touch, Townsley        Expires December 8, 2017               [Page 42]

Internet-Draft         Tunnels in the Internet                June 2017

   [Access control?]

   describe relationship to [RFC6169] - JT (as per INTAREA meeting
   notes, don't cover Teredo-specific issues in RFC6169, but include
   generic issues here)

7. IANA Considerations

   This document has no IANA considerations.

   The RFC Editor should remove this section prior to publication.

8. References

8.1. Normative References

   [RFC2119] Bradner, S., "Key words for use in RFCs to Indicate
             Requirement Levels", BCP 14, RFC 2119, March 1997.

   [are there others? 3819? ECN? Flow label issues?]

8.2. Informative References

   [Cl88]    Clark, D., "The design philosophy of the DARPA internet
             protocols," Proc. Sigcomm 1988, p.106-114, 1988.

   [Er94]    Eriksson, H., "MBone: The Multicast Backbone,"
             Communications of the ACM, Aug. 1994, pp.54-60.

   [Gr17]    Gross, J. (Ed.), I. Ganga (Ed.), T. Sridhar (Ed.), "Geneve:
             Generic Network Virtualization Encapsulation," draft-ietf-
             nvo3-geneve-04, Mar. 2017.

   [He16]    Herbert, T., L. Yong, O. Zia, "Generic UDP Encapsulation,"
             draft-ietf-nvo3-gue-05, Oct. 2016.

   [Ke95]    Kent, S., J. Mogul, "Fragmentation considered harmful," ACM
             Sigcomm Computer Communication Review (CCR), V25 N1, Jan.
             1995, pp. 75-87.

   [No16]    Nordmark, E. (Ed.), A. Tian, J. Gross, J. Hudson, L.
             Kreeger, P. Garg, P. Thaler, T. Herbert, "Encapsulation
             Considerations," draft-ietf-rtgwg-dt-encap-02, Oct. 2016.

   [RFC5]    Rulifson, J, "Decode Encode Language (DEL)," RFC 5, June

Touch, Townsley        Expires December 8, 2017               [Page 43]

Internet-Draft         Tunnels in the Internet                June 2017

   [RFC768]  Postel, J, "User Datagram Protocol," RFC 768, Aug. 1980

   [RFC791]  Postel, J., "Internet Protocol," RFC 791 / STD 5, September

   [RFC792]  Postel, J., "Internet Control Message Protocol," RFC 792,
             Sep. 981.

   [RFC793]  Postel, J, "Transmission Control Protocol," RFC 793, Sept.

   [RFC826]  Plummer, D., "An Ethernet Address Resolution Protocol -- or
             -- Converting Network Protocol Addresses to 48.bit Ethernet
             Address for Transmission on Ethernet Hardware," RFC 826,
             Nov. 1982.

   [RFC1075] Waitzman, D., C. Partridge, S. Deering, "Distance Vector
             Multicast Routing Protocol," RFC 1075, Nov. 1988.

   [RFC1122] Braden, R., Ed., "Requirements for Internet Hosts -
             Communication Layers," RFC 1122 / STD 3, October 1989.

   [RFC1191] Mogul, J., S. Deering, "Path MTU discovery," RFC 1191,
             November 1990.

   [RFC1812] Baker, F., "Requirements for IP Version 4 Routers," RFC
             1812, June 1995.

   [RFC1853] Simpson, W., "IP in IP Tunneling," RFC 1853, Oct. 1995.

   [RFC1981] McCann, J., S. Deering, J. Mogul, "Path MTU Discovery for
             IP version 6," RFC 1981, Aug. 1996.

   [RFC2003] Perkins, C., "IP Encapsulation within IP," RFC 2003, Oct.

   [RFC2225] Laubach, M., J. Halpern, "Classical IP and ARP over ATM,"
             RFC 2225, Apr. 1998.

   [RFC2460] Deering, S., R. Hinden, "Internet Protocol, Version 6
             (IPv6) Specification," RFC 2460, Dec. 1998.

   [RFC2473] Conta, A., "Generic Packet Tunneling in IPv6
             Specification," RFC 2473, Dec. 1998.

   [RFC2529] Carpenter, B., C. Jung, "Transmission of IPv6 over IPv4
             Domains without Explicit Tunnels," RFC 2529, Mar. 1999.

Touch, Townsley        Expires December 8, 2017               [Page 44]

Internet-Draft         Tunnels in the Internet                June 2017

   [RFC2784] Farinacci, D., T. Li, S. Hanks, D. Meyer, P. Traina,
             "Generic Routing Encapsulation (GRE)", RFC 2784, March

   [RFC2923] Lahey, K., "TCP Problems with Path MTU Discovery," RFC
             2923, September 2000.

   [RFC2983] Black, D., "Differentiated Services and Tunnels," RFC 2983,
             Oct. 2000.

   [RFC2991] Thaler, D., C. Hopps, "Multipath Issues in Unicast and
             Multicast Next-Hop Selection," RFC 2991, Nov. 2000.

   [RFC2473] Conta, A., S. Deering, "Generic Packet Tunneling in IPv6
             Specification," RFC 2473, Dec. 1998.

   [RFC2546] Durand, A., B. Buclin, "6bone Routing Practice," RFC 2540,
             Mar. 1999.

   [RFC3031] Rosen, E., A. Viswanathan, R. Callon, "Multiprotocol Label
             Switching Architecture", RFC 3031, January 2001.

   [RFC3819] Karn, P., Ed., C. Bormann, G. Fairhurst, D. Grossman, R.
             Ludwig, J. Mahdavi, G. Montenegro, J. Touch, L. Wood,
             "Advice for Internet Subnetwork Designers," RFC 3819 / BCP
             89, July 2004.

   [RFC3884] Touch, J., L. Eggert, Y. Wang, "Use of IPsec Transport Mode
             for Dynamic Routing," RFC 3884, September 2004.

   [RFC3931] Lau, J., Ed., M. Townsley, Ed., I. Goyret, Ed., "Layer Two
             Tunneling Protocol - Version 3 (L2TPv3)," RFC 3931, March

   [RFC3985] Bryant, S., P. Pate (Eds.), "Pseudo Wire Emulation Edge-to-
             Edge (PWE3) Architecture", RFC 3985, March 2005.

   [RFC4176] El Mghazli, Y., Ed., T. Nadeau, M. Boucadair, K. Chan, A.
             Gonguet, "Framework for Layer 3 Virtual Private Networks
             (L3VPN) Operations and Management," RFC 4176, October 2005.

   [RFC4301] Kent, S., and K. Seo, "Security Architecture for the
             Internet Protocol," RFC 4301, December 2005.

   [RFC4340] Kohler, E., M. Handley, S. Floyd, "Datagram Congestion
             Control Protocol (DCCP)," RFC 4340, Mar. 2006.

Touch, Townsley        Expires December 8, 2017               [Page 45]

Internet-Draft         Tunnels in the Internet                June 2017

   [RFC4443] Conta, A., S. Deering, M. Gupta (Ed.), "Internet Control
             Message Protocol (ICMPv6) for the Internet Protocol Version
             6 (IPv6) Specification," RFC 4443, Mar. 2006.

   [RFC4459] Savola, P., "MTU and Fragmentation Issues with In-the-
             Network Tunneling," RFC 4459, April 2006.

   [RFC4541] Christensen, M., K. Kimball, F. Solensky, "Considerations
             for Internet Group Management Protocol (IGMP) and Multicast
             Listener Discovery (MLD) Snooping Switches," RFC 4541, May

   [RFC4664] Andersson, L., Ed., E. Rosen, Ed., "Framework for Layer 2
             Virtual Private Networks (L2VPNs)," RFC 4664, September

   [RFC4821] Mathis, M., J. Heffner, "Packetization Layer Path MTU
             Discovery," RFC 4821, March 2007.

   [RFC4861] Narten, T., E. Nordmark, W. Simpson, H. Soliman, "Neighbor
             Discovery for IP version 6 (IPv6)," RFC 4861, Sept. 2007.

   [RFC4960] Stewart, R. (Ed.), "Stream Control Transmission Protocol,"
             RFC 4960, Sep. 2007.

   [RFC4963] Heffner, J., M. Mathis, B. Chandler, "IPv4 Reassembly
             Errors at High Data Rates," RFC 4963, July 2007.

   [RFC5214] Templin, F., T. Gleeson, D. Thaler, "Intra-Site Automatic
             Tunnel Addressing Protocol (ISATAP)," RFC 5214, Mar. 2008.

   [RFC5320] Templin, F., Ed., "The Subnetwork Encapsulation and
             Adaptation Layer (SEAL)," RFC 5320, Feb. 2010.

   [RFC5556] Touch, J., R. Perlman, "Transparently Interconnecting Lots
             of Links (TRILL): Problem and Applicability Statement," RFC
             5556, May 2009.

   [RFC5944] Perkins, C., Ed., "IP Mobility Support for IPv4, Revised"
             RFC 5944, Nov. 2010.

   [RFC6040] Briscoe, B., "Tunneling of Explicit Congestion
             Notification," RFC 6040, Nov. 2010.

   [RFC6169] Krishnan, S., D. Thaler, J. Hoagland, "Security Concerns
             With IP Tunneling," RFC 6169, Apr. 2011.

Touch, Townsley        Expires December 8, 2017               [Page 46]

Internet-Draft         Tunnels in the Internet                June 2017

   [RFC6325] Perlman, R., D. Eastlake, D. Dutt, S. Gai, A. Ghanwani,
             "Routing Bridges (RBridges): Base Protocol Specification,"
             RFC 6325, July 2011.

   [RFC6434] Jankiewicz, E., J. Loughney, T. Narten, "IPv6 Node
             Requirements," RFC 6434, Dec. 2011.

   [RFC6438] Carpenter, B., S. Amante, "Using the IPv6 Flow Label for
             Equal Cost Multipath Routing and Link Aggregation in
             Tunnels," RFC 6438, Nov. 2011.

   [RFC6807] Farinacci, D., G. Shepherd, S. Venaas, Y. Cai, "Population
             Count Extensions to Protocol Independent Multicast (PIM),"
             RFC 6807, Dec. 2012.

   [RFC6830] Farinacci, D., V. Fuller, D. Meyer, D. Lewis, "The
             Locator/ID Separation Protocol," RFC 6830, Jan. 2013.

   [RFC6833] Fuller, V., D. Farinacci, "Locator/ID Separation Protocol
             (LISP) Map-Server Interface," RFC 6833, Jan. 2013.

   [RFC6864] Touch, J., "Updated Specification of the IPv4 ID Field,"
             Proposed Standard, RFC 6864, Feb. 2013.

   [RFC6935] Eubanks, M., P. Chimento, M. Westerlund, "IPv6 and UDP
             Checksums for Tunneled Packets," RFC 6935, Apr. 2013.

   [RFC6936] Fairhurst, G., M. Westerlund, "Applicability Statement for
             the Use of IPv6 UDP Datagrams with Zero Checksums," RFC
             6936, Apr. 2013.

   [RFC6946] Gont, F., "Processing of IPv6 "Atomic" Fragments," RFC
             6946, May 2013.

   [RFC7364] Narten, T., Gray, E., Black, D., Fang, L., Kreeger, L., M.
             Napierala, "Problem Statement: Overlays for Network
             Virtualization", RFC 7364, Oct. 2014.

   [RFC7450] Bumgardner, G., "Automatic Multicast Tunneling," RFC 7450,
             Feb. 2015.

   [RFC7510] Xu, X., N. Sheth, L. Yong, R. Callon, D. Black,
             "Encapsulating MPLS in UDP," RFC 7510, April 2015.

   [RFC7588] Bonica, R., C. Pignataro, J. Touch, "A Widely-Deployed
             Solution to the Generic Routing Encapsulation Fragmentation
             Problem," RFC 7588, July 2015.

Touch, Townsley        Expires December 8, 2017               [Page 47]

Internet-Draft         Tunnels in the Internet                June 2017

   [RFC7676] Pignataro, C., R. Bonica, S. Krishnan, "IPv6 Support for
             Generic Routing Encapsulation (GRE)," RFC 7676, Oct 2015.

   [RFC7690] Byerly, M., M. Hite, J. Jaeggli, "Close Encounters of the
             ICMP Type 2 Kind (Near Misses with ICMPv6 Packet Too Big
             (PTB))," RFC 7690, Jan. 2016.

   [RFC7761] Fenner, B., M. Handley, H. Holbrook, I. Kouvelas, R.
             Parekh, Z. Zhang, L. Zheng, "Protocol Independent Multicast
             - Sparse Mode (PIM-SM): Protocol Specification (Revised),"
             RFC 7761, Mar. 2016.

   [RFC8085] Eggert, L., G. Fairhurst, G. Shepherd, "Unicast UDP Usage
             Guidelines," RFC 8085, Oct. 2015.

   [RFC8086] Yong, L. (Ed.), E. Crabbe, X. Xu, T. Herbert, "GRE-in-UDP
             Encapsulation," RFC 8086, Feb. 2017.

   [Sa84]    Saltzer, J., D. Reed, D. Clark, "End-to-end arguments in
             system design," ACM Trans. on Computing Systems, Nov. 1984.

   [Te17]    Templin, F., "Asymmetric Extended Route Optimization,"
             draft-templin-aerolink-75, May 2017.

   [To01]    Touch, J., "Dynamic Internet Overlay Deployment and
             Management Using the X-Bone," Computer Networks, July 2001,
             pp. 117-135.

   [To03]    Touch, J., Y. Wang, L. Eggert, G. Finn, "Virtual Internet
             Architecture," USC/ISI Tech. Report ISI-TR-570, Aug. 2003.

   [To16]    Touch, J., "Middleboxes Models Compatible with the
             Internet," USC/ISI Tech. Report ISI-TR-711, Oct. 2016.

   [To98]    Touch, J., S. Hotz, "The X-Bone," Proc. Globecom Third
             Global Internet Mini-Conference, Nov. 1998.

   [Zi80]    Zimmermann, H., "OSI Reference Model - The ISO Model of
             Architecture for Open Systems Interconnection," IEEE Trans.
             on Comm., Apr. 1980.

9. Acknowledgments

   This document originated as the result of numerous discussions among
   the authors, Jari Arkko, Stuart Bryant, Lars Eggert, Ted Faber, Gorry
   Fairhurst, Dino Farinacci, Matt Mathis, and Fred Templin. It
   benefitted substantially from detailed feedback from Toerless Eckert,

Touch, Townsley        Expires December 8, 2017               [Page 48]

Internet-Draft         Tunnels in the Internet                June 2017

   Vincent Roca, and Lucy Yong, as well as other members of the Internet
   Area Working Group.

   This work is partly supported by USC/ISI's Postel Center.

   This document was prepared using 2-Word-v2.0.template.dot.

Authors' Addresses

   Joe Touch
   4676 Admiralty Way
   Marina del Rey, CA 90292-6695

   Phone: +1 (310) 448-9151
   Email: touch@isi.edu

   W. Mark Townsley
   L'Atlantis, 11, Rue Camille Desmoulins
   Issy Les Moulineaux, ILE DE FRANCE 92782

   Email: townsley@cisco.com

Touch, Townsley        Expires December 8, 2017               [Page 49]

Internet-Draft         Tunnels in the Internet                June 2017

APPENDIX A: Fragmentation efficiency

A.1. Selecting fragment sizes

   There are different ways to fragment a packet. Consider a network
   with a PMTU as shown in Figure 16, where packets are encapsulated
   over the same network layer as they arrive on (e.g., IP in IP). If a
   packet as large as the PMTU arrives, it must be fragmented to
   accommodate the additional header.

         X===========================X (transit PMTU)
           |  X===========================X (tunnel 1 MTU)
           |  +---+----+------------------+
       (a) +->| H'| iH | DDDDDDDDDDDDDDDD |
           |  +---+----+------------------+
           |      |
           |      |  X===========================X (tunnel 2 MTU)
           |      |  +----+---+----+-------------+
           | (a1) +->| nH'| H | iH | DDDDDDDDDDD |
           |      |  +----+---+----+-------------+
           |      |
           |      |  +----+-------+
           | (a2) +->| nH"| DDDDD |
           |         +----+-------+
           |  +---+------+
       (b) +->| H"| DDDD |
                  |  +----+---+------+
             (b1) +->| nH'| H"| DDDD |

                   Figure 16 Fragmenting via maximum fit

   Figure 16 shows this process using "maximum fit", assuming outer
   fragmentation as an example (the situation is the same for inner
   fragmentation, but the headers that are affected differ). In maximum
   fit, the arriving packet is split into (a) and (b), where (a) is the
   size of the first tunnel, i.e., the tunnel 1 MTU (the maximum that
   fits over the first tunnel). However, this tunnel then traverses over
   another tunnel (number 2), whose impact the first tunnel ingress has
   not accommodated. The packet (a) arrives at the second tunnel

Touch, Townsley        Expires December 8, 2017               [Page 50]

Internet-Draft         Tunnels in the Internet                June 2017

   ingress, and needs to be encapsulated again, but it needs to be
   fragmented as well to fit into the tunnel 2 MTU, into (a1) and (a2).
   In this case, packet (b) arrives at the second tunnel ingress and is
   encapsulated into (b1) without fragmentation, because it is already
   below the tunnel 2 MTU size.

   In Figure 17, the fragmentation is done using "even split", i.e., by
   splitting the original packet into two roughly equal-sized
   components, (c) and (d). Note that (d) contains more packet data,
   because (c) includes the original packet header because this is an
   example of outer fragmentation. The packets (c) and (d) arrive at the
   second tunnel encapsulator, and are encapsulated again; this time,
   neither packet exceeds the tunnel 2 MTU, and neither requires further

         X===========================X (transit PMTU)
           |  X===========================X (tunnel 1 MTU)
           |  +---+----+----------+
       (c) +->| H'| iH | DDDDDDDD |
           |  +---+----+----------+
           |      |
           |      |  X===========================X (tunnel 2 MTU)
           |      |  +----+---+----+----------+
           | (c1) +->| nH | H'| iH | DDDDDDDD |
           |         +----+---+----+----------+
           |  +---+--------------+
       (d) +->| H"| DDDDDDDDDDDD |
                  |  +----+---+--------------+
             (d1) +->| nH | H"| DDDDDDDDDDDD |

                  Figure 17 Fragmenting via "even split"

A.2. Packing

   Encapsulating individual packets to traverse a tunnel can be
   inefficient, especially where headers are large relative to the
   packets being carried. In that case, it can be more efficient to
   encapsulate many small packets in a single, larger tunnel payload.

Touch, Townsley        Expires December 8, 2017               [Page 51]

Internet-Draft         Tunnels in the Internet                June 2017

   This technique, similar to the effect of packet bursting in Gigabit
   Ethernet (regardless of whether they're encoded using L2 symbols as
   delineators), reduces the overhead of the encapsulation headers
   (Figure 18). It reduces the work of header addition and removal at
   the tunnel endpoints, but increases other work involving the packing
   and unpacking of the component packets carried.

                     | iHa | iDa |
                           |     +-----+-----+
                           |     | iHb | iDb |
                           |     +-----+-----+
                           |           |
                           |           |     +-----+-----+
                           |           |     | iHc | iDc |
                           |           |     +-----+-----+
                           |           |           |
                           v           v           v
                | oH | iHa | iDa | iHb | iDb | iHc | iDc |

                  Figure 18 Packing packets into a tunnel

Touch, Townsley        Expires December 8, 2017               [Page 52]

Html markup produced by rfcmarkup 1.122, available from https://tools.ietf.org/tools/rfcmarkup/