[Search] [txt|pdfized|bibtex] [Tracker] [Email] [Diff1] [Diff2] [Nits]
Versions: 00 01                                                         
   Network Working Group                                     Eliot Lear
   INTERNET-DRAFT                                         Cisco Systems
   Category: Informational

                            January 31, 2001

                  A Middle Box Architectural Framework

1  Status of this Memo

   This document is an Internet-Draft and is in full conformance with
   all provisions of Section 10 of RFC2026.

   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

   2. Copyright Notice

   Copyright (C) The Internet Society (2000).  All Rights Reserved.

2  Abstract

   It used to be reasonable to expect that any two points connected to
   the Internet to have the ability to hold any communication.  Such an
   expectation has not be reasonable for quite some time, thanks to
   firewalls, NATs, and other intermediate devices.  Today, we
   acknowledge a new architecture and we name the functional blocks of
   that architecture as well as several ways to get ends to communicate,
   and how two devices could expect to communicate with each other.
   This document does not define the protocols involved.

3  Introduction

   The IPv4 Internet consists of a network of interconnected networks
   that may use public or private address space.  The use of private
   address space [BCP5] has broken the classic connection model that
   applications use to speak to other devices.  Similarly, many networks
   are separated by firewalls.

   Now we consider the components necessary for end to end
   communications in this environment, as well as the forms of signaling
   necessary for those communications to occur in a reliable way.

   The reader should be familiar with RFCs 1918, 2663, and 2775.  We do
   not intend to fix all problems related to NATs or firewalls with this
   architecture.  For instance, one will not find below a way to save a
   TCP connection in the face of a NAT failure.

3.1  Motivation

   The currently envisioned sets of applications to make use of this
   architecture are voice and video conferencing and telephony.  Others
   may follow.  Today voice and video capture a small percentage of
   total network usage.  As demand for these uses grows it will become
   more important that we have in place mechanisms that have proper
   scaling properties.

3.2  Goals, Terms, and Limits

   Here are our design goals:

   1.   Enable hosts within different administrative domains and address
     realms to have end to end sessions.
   2.   The architecture must allow for multiple middle boxes that connect
     one realm to multiple realms.
   3.   Enable this with a minimal amount of signaling from the end host,
     and no new signaling beyond administratively defined boundaries.
   4.   The recommended methods must not radically alter the end host stack
     below the application.
   5.   Work within the existing interior and exterior routing framework.
   6.   The mechanisms used must easily integrate with the existing
     Internet mechanisms and services.

   The three middle box processes we consider in this document are
   discovery, diagnostics, and signaling.  Discovery means determining
   that one or more middle box sits between a host end the remote end of
   a session.  Once that box is discovered, either an end host or its
   proxy will exchange information with it in order for the
   communication to proceed.  Diagnostics is the untimely discovery of a
   problem involving a middle box.  Discovery and diagnostics differ in
   that an end host may receive a diagnostic message from one middle box
   while it might need to discover a separate middle box in order for
   the communication to proceed.  The architecture we propose elides the
   two functions.

   No changes are made to the layer 3 routing mechanism, other than the
   possible addition of a new ICMP message to be multiplexed back to the
   responsible application.  To do otherwise would dramatically increase
   implementation cost and complexity, as well as the cost of
   troubleshooting problems.  While the host may signal for additional
   information to and from the application layer, it cannot rely on new
   layer 3 routing facilities within the network.

   Nothing this architecture proposes may stop two oHOSTs from
   communicating with each other without the help of a middle box.
   Similarly, this architecture must not prevent two iHOSTS from
   communicating, just as they would have previously.  However, for all
   the reasons listed in RFC 2775 it is not possible for a middle box to
   enable all forms of communication between iHOSTs and oHOSTs.  The
   recommended method to clear at least some of these roadblocks is the
   wide deployment of IP version 6.

4  Architectural Components and Terms

   We now define components of the architecture based on the following

   |                             |
   |   Private Realm             |
   |                             |
   |                     [AC]    |
   |                             |           -------        [oHOST]
   |                             |          /        \
   |  [iHOST]               [middle box]    |Internet|
   |                             |          \        /
   |              [iHOST]        |           \------/
   |                             |

   A private realm consists either of hosts that are numbered within the
   space defined by BCP 5 or hosts that sit within a single security
   domain.  An iHOST is a host that sits within that realm.  An oHOST
   sits outside the private address realm.  It may be assigned public
   address space or private address space.  In the latter case it would
   sit within its own private address realm.   An AC is an application
   controller.  An application controller may be an application layer
   gateway, or it may merely arrange for communication between two
   hosts, be they iHOSTs or oHOSTs.  An AC may itself be a middle box.

   A middle box is a device that sits on the edge of a private realm.
   It is the responsibility of the middle box to police or transform
   sessions from its private realm to the outside world.  Middle boxes
   consist of firewalls, NATs and other devices that sit within the flow
   of packets and may impact a session from an iHOST to an oHOST.  At
   the border of a private realm there may be multiple middle boxes that
   connect the realm to other realms.

   The remainder of this document refers to communications between an
   iHOST and an oHOST.

4.2 Communication Models

   No matter the model we assume that iHOSTS may only signal to middle
   boxes that exist within the same administrative domain.

   There are two models of communication we will consider.  The first
   model may not require end host application changes, but rather
   requires changes to servers that assist end hosts in communicating
   with each other.  The second model allows end hosts themselves to
   exchange information with the intermediate devices.  These two models
   differ significantly.

4.2.1  Proxy Signaling

   In the first case, iHOSTs use something akin to an application layer
   gateway to first arrange a session.  We will generically refer to
   such a box as an application controller (AC).  An example would be
   H.225 signaling between an H.323 end point and a gate keeper.  As the
   gatekeeper processes call setup information it would communicate with
   any middle box necessary for the end points to establish end to end
   connectivity.  This model is best employed when there already exists
   some sort of application controller, since no additional signaling
   may be necessary between the iHOST and the AC.  Either the end
   application is required to know of the AC or the AC must reside
   within the data path between the iHOST and oHOST.

   |                             |
   |   Private Realm             |
   |                             |
   |          --- 1 --> [AC]->v  |
   |         /                v  |
   |        /                 2  |
   |       |                  v  |           -------        [oHOST]
   |       |                  v  |          /        \         ^
   |  [iHOST]-----3-------->[middle box]--->|Internet|---3---->^
   |                             |          \        /
   |              [iHOST]        |           \------/
   |                             |

   In order for the end hosts to continue sessions with some level of
   reliability, one of the following cases must be true:

   1.   The AC must maintain topological awareness, so that it can
     determine which middle boxes are involved.  It might do this by
     monitoring and examining link state information from the IGP.
   2.   The AC is configured with the address of a middle box and the
     middle box must then refer the request to other appropriate middle
     boxes, as necessary.  In this case either the middle box has sufficient
     topological knowledge or the request must be flooded to all middle

   This model may require no new signaling between iHOSTs and middle
   boxes, so long as the AC provides any required server functionality.

4.2.2 Direct Signaling

   In the second case, no AC is required.  Instead, an iHOST must
   determine that a middle box exists, signal to it for end to end
   configuration information, and then proceed.  Furthermore, the iHOST
   must determine when the path to the oHOST has changed during a
   session.  The application should recover as circumstances dictate.

   In this model it is important for the end host to determine that not
   only has a failure occurred, but that the failure occurred due to
   something the middle box could see or control.

   |                             |
   |   Private Realm             |
   |                             |
   |                             |
   |                             |           -------        [oHOST]
   |        >------2--------v    |          /        \
   |  [iHOST]<-----1--------[middle box]    |Internet|
   |       ^<------3--------v    |          \        /
   |                             |           \------/
   |                             |

   1.   Diagnostic/discovery, a message or messages that indicate that the
     middle box requires attention.  This message is sent in response to an
     attempt by iHOST to start a session with oHOST that the middle box knows
     will not succeed.
   2.   Signaling request from iHOST to middle box.
   3.   Response from middle box.

   Note that in this model there is no AC.

   However, the middle box must be prepared to signal to an iHOST that
   it is being contacted.  And indeed this mechanism will only work
   reasonably if it is integrated with DNS to allow for temporary
   allocation of addresses at the time of a query.  It is a policy
   decision for the middlebox to determine when such addresses should be

4.2.3 Current use of Application Layer Gateways

   An alternative model would be for end hosts to use application layer
   gateways to access external resources.  This model requires no new
   generic signaling, but a method for each iHOST to determine when it
   should use an application layer gateway, and when it should
   communicate directly with another iHOST.

   In this case, because the ALG is a middle box, it follows that this
   method requires the ALG to reside on the perimeter of the
   administrative domain.

   We mention this model not merely for completeness, but because it is
   the operative model for many applications that would eventually use
   one of the other two models.  Its clear benefit is that it exists
   today.  Furthermore, there may be benefits to having ALGs reside on
   the perimeter.  For instance, these devices will be able to
   statefully inspect each and every packet to and from an internal

4.3  Which model should we build?

   The astute reader will notice that the direct model is very close to
   a superset of the proxy model.  The proxy model needs a signaling
   mechanism between the AC and the middle box.  There is no reason that
   signaling mechanism couldn't be the same one used by the direct

   Indeed as diagnostics are introduced they can enhance both the proxy
   model and the ALG model by returning decent diagnostics to the end
   user when iHOST not properly configured or the AC or ALG unexpectedly
   falls outside the data stream.

5  Diagnostics and Discovery

   As mentioned in the introduction, we elide the diagnostics and
   discovery functions in this architecture.  No matter which of the
   above models is implemented diagnostics are required to indicate to
   the users and application when a failure has occurred.  These
   failures can take numerous forms, but here are some examples:

     o  A middle box might reboot and lose state of which holes are
   meant to have been open.
     o  The traffic may be rerouted to an unsuspecting middle box.
     o  The application on iHOST may be unaware of a middle box for
   which an information exchange is required.
     o  An AC may be misconfigured to prevent the iHOST from
   establishing a session with an oHOST.

   The diagnostic mechanism used on the Internet is ICMP.  When a middle
   box detects an attempt by an iHOST to start a session to an oHOST
   that will not succeed, it should send a message back to the iHOST
   indicating that a failure is likely.  The content of that ICMP
   message is discussed separately in [BLOCKER].

   Of the two models mentioned in Section 4, discovery may be different.
   With the proxy model it is fairly easy to configure by hand a
   relatively small number of ACs.  Indeed either ACs must know about
   the middle boxes or the middle boxes must know about them, since ACs
   do not sit within the packet flow.

   With the direct signaling model it is possible to piggyback discovery
   on top of the diagnostic message discussed earlier.

6  Signaling

   Once the application has determined that it must communicate with a
   middle box in order for a communication to properly proceed, either
   the AC or the iHOST initiates an exchange with the middle box.  The
   nature of this exchange depends on the function of the middle box.
   For instance, if the middle box is a firewall, the application is in
   essence asking permission from the firewall for the communication to
   proceed.  If, on the other hand, the box is a NAT or NAPT, the middle
   box may merely need to know the mapping between the two addressing
   realms for the communication.  These two functions can be combined,
   and it is reasonable to do so in order to reduce signaling overhead.

   Any signaling requests to reserve address space or open pinholes must
   be matched with similar requests to undo what was done.  However,
   firewalls will as a matter of policy not trust that all went well.
   Indeed they should be fairly conservative to reduce the risk that
   pinholes have not been left open beyond their legitimate purposes.

6.1 Firewalls

   Firewalls require sufficient information about the communication to
   determine whether or not it is authentic and whether or not it is
   authorized.  The firewall may query the application for specific
   information necessary for authorization, but we can assume that
   during the initial contact the firewall will need at least the
   following information:

   1.   Protocol to be used during a session
   2.   IP address and source port of the iHOST
   3.   IP address and source port of the oHOST
   4.   One or more methods to indicate when the session has ended.

   The firewall may also wish to know what person is initiating the
   request, the application that is being used, or even perhaps a token
   of some form.  While it might wish to have arbitrary amounts of
   information in order to make its decision, applications need to be
   aware of the sorts of information the firewall will demand.  Thus,
   the names and formats of the values of the requested information must
   be standardized, and should be managed by IANA.

   If a firewall is going to respond to a request from an iHOST or AC it
   SHOULD do so with a request for all the information it needs in
   response to an initial request from either the iHOST or AC.

   The format of the query the firewall makes must be standardized, as
   should the names and formats of the individual attributes.

   A firewall may accept, reject, or ignore such signaling requests.  If
   the firewall accepts a request it should respond with the protocol,
   source and destination IP addresses and ports, confirming the way the
   communication shall be terminated.  In addition, it may return
   additional information, as requested.  This brings us to NATs.

   Note that even if a communication is initially authorized, nothing we
   state here should prevent or even discourage further stateful
   inspection of any communication.

6.2  NATs

   An AC or iHOST may request the port and IP address mapping between
   two realms as part of the query discussed above.  The middle box may
   return the appropriate mappings.  If it does so, it MUST also return
   connection termination conditions.

   Just as NATs work today mappings may be created prior to any
   signaling exchange between the middle box and the AC or iHOST.

6.3  Termination Conditions

   It is important for the middle box and the application (AC or iHOST)
   to agree on when a session has ended.  In the case of a firewall it
   is critical that it properly close the pinhole it opened.  In the
   case of a NAT, once a session has terminated the NAT may reallocate
   addressees and ports to another iHOST.

   Termination conditions can be one of several methods:

   1.   A period of time, similar to a TTL used by DHCP.
   2.   A requirement that the application tell the middle box when the
     communication has ended.
   3.   An easily discernable in stream method to determine that the
     session is over.  For instance, a TCP session ends with the exchange of
     FINs, followed by their acknowledgement.  Such methods should be
     described in an RFC.

   While an application might request one or more method it is up to the
   middle box to decide which method to employ.  If more than one method
   is contained within a request or a response, the termination
   condition that occurs soonest will be used.

   For example, an H.323 gateway might request a UDP connection from the
   iHOST on port 7499 to oHOST on port 8233.  The AC might also request
   that it tell the middle box when the session is terminated.  A
   firewall might respond that it has opened a pinhole as described, and
   the termination conditions are that the AC will indicate the
   completion of the session AND the session will be considered closed
   after five minutes.

   Prior to the expiration of the TTL, if the call is still active, the
   AC might further request additional time from the middle box.

6.4  Communication between middle boxes

   Although it is theoretically possible for middle boxes to exchange
   connection state amongst each other, the overhead for doing so may
   well prove quite high, and the value is dubious.  If a middle box
   fails it is possible that a hot spare would be able to take over its
   responsibilities.  There exists at least one document that considers
   this possibility [YAKOV et.al].  However, we choose not to
   standardize this function at this time.

   More likely the case will be that any existing connections will fail
   due to a topological change, either the middle box failing or a route
   to the middle box failing.  Therefore, it is important that end hosts
   be able to re-establish communications and retain state above the
   transport layer, as is necessary and appropriate.

6.5  Signaling Protocol Choice

   Returning to the discussion of the two different models discussed in
   Section 4, we note that there are subtle differences in expected
   protocol characteristics between the proxy and direct models.  In the
   case of the direct model an iHOST could expect to issue a single
   request per session.  However, in the case of a proxy, it is likely
   to make many requests to the middle box on behalf many client iHOSTs.

   The signaling protocol should allow for ease of failover.  In
   addition, the protocol should also take minimal resources on both
   client and middle box.  The client itself may be a middle box. In any
   event the signaling end points - both middle box and client - MUST
   MUST MUST implement appropriate congestion control mechanisms.

6.6  Direction of Initiation

   In the most straight forward case, either the iHOST or the AC would
   initiate signaling to a middle box or group of middle boxes.  This is
   the simple case, and as we have seen, it's not all that simple.

   The harder case occurs when an oHOST needs to communicate with an
   iHOST, and wishes to initiate communication. Now, all the same
   roadblocks may need to be cleared, only neither the iHOST  nor the AC
   is aware of need to do so.  Possibly the middle box may transmit a
   diagnostic to the iHOST or AC, indicating the initiation attempt.
   Sufficient information must be sent so that the signaling request can
   be initiated.  Just as in normal connection attempts the iHOST MUST
   properly handle the case where it does not wish to allow the
   connection.  It would indicate this by not signaling back to the
   middle box or AC.

7  Multiple Middle Boxes

   When used with the direct model, a diagnostic message such as ICMP
   allows the application on an iHOST to determine not only that a
   middle box is in its path, but also which middle box is in its path.
   Once the iHOST identifies the middle box it can signal to it.  Should
   the data path change so that another middle box is chosen, the iHOST
   will once again receive a notification.

   Depending on the application or environment it may be possible for an
   iHOST to fail over between two middle boxes that are sharing state.
   Such failures must be transparent to the iHOST at all layers.

   As mentioned earlier, the matter of multiple middle boxes is somewhat
   more complex with the proxy model.  Because the AC is not in the data
   flow it must go to some additional measures to determine that a
   middle box has failed.  Indeed once the AC has determined that a
   particular middle box has failed, or that a path has changed, it must
   communicate appropriate information back to the iHOST.  Thus, unless
   the application already anticipates appropriate failure and restart
   conditions, modifications may be required, defeating the usefulness
   of the proxy model.

   There is one additional case to be considered. When an iHOST attempts
   to communicate across several concentric boundaries it might require
   several rounds of signaling before a session could proceed.  The
   fastest way for signaling to proceed within the direct model is for
   the middle box to forward a packet that has generated a diagnostic,
   so that the very same packet could cause the next middle box to
   generate a diagnostic as well, etc.

   Note that there are a number of deployment scenarios one could create
   in which the signaling itself could create a diagnostic from a middle
   box.  We declare such scenarios a matter of bad implementation and

8  The Stack Simplification Act of 2001

   Some mechanisms such as [RSIP] significantly complicate the host
   stack by not giving sufficient guidance as to when the mechanism
   should be used.

   We propose that there are only three alternatives:

   1.   the application relies on the host's network layer to get the
     packets directly to the other end;
   2.   the application communicates with a service that identifies
     failures and optionally enables a flow;
   3.   the application is explicitly configured to use an application
     layer gateway.

   Option 2 is of great concern, in as much as the host must properly
   multiplex any messages to an application, and those messages may need
   to be received out of band of the normal communication.

9  Future Work

   The first effort necessary is to conclude in fact whether or not
   additional diagnostics are necessary.  If so, we must next determine
   the exact mechanisms to deliver those diagnostics.  We must further
   agree on whether or not additional discovery mechanisms should be

   This document focuses on unicast based applications.  We believe it
   provides sufficient flexibility to allow for design of multicast
   applications that take advantage of those building blocks, but more
   study is clearly needed.

10  Security Considerations

   There are numerous security considerations that middle boxes will
   encounter, and the ones we list below should be viewed as far from

   Any time a request is made for information or for a configuration
   change it should be viewed with great suspicion.  It is as of yet
   unclear all the attacks that can be made using either the signaling
   mechanism proposed in this document or the diagnostic messages
   proposed in [BLOCKER].

   To minimize the risk of attacks, this mechanism should only be used
   in conjunction with strong authentication and a conservative
   authorization model.  The lower bound of risk is likely that of
   today's model, where sites generally allow outbound TCP connections.

   The signaling, diagnostics, and discovery discussed in this draft are
   useful only within the boundaries of a single administrative domain.
   The middle boxes on the borders of that domain should prevent
   external devices from participating by not transmitting diagnostic
   messages outside, and by not listening for signaling requests on
   interfaces external to that domain.

   Furthermore, intrusion detection systems would be well advised to
   look for such requests as an indication of either configuration error
   or a possible attack.

   The period of time between the termination of a communication and the
   termination of pinholes in firewalls may allow for mischief.  End
   hosts must be prepared to ignore unsolicited traffic to the ports

11 IANA Considerations

   The names of attributes and the format of their contents that a
   middle box can either furnish or request will need to be held in a

12  References

   [1] Rekhter et. al., "Address Allocation for Private Internets", RFC
   1918, February 1996.

   [4] H.323

   [6] Carpenter, B., "Internet Transparency",  RFC 2775, February 2000.

   [7] Postel, J., "Internet Control Message Protocol", RFC 792,
   USC/Information Sciences Institute, September 1981.

   [9] Postel, J., ed., "Transmission Control Protocol - DARPA Internet
   Program Protocol Specification", RFC 793, USC/Information Sciences
   Institute, NTIS AD Number A111091, September 1981.

   [11] Postel, J., "User Datagram Protocol", RFC 768, USC/Information
   Sciences Institute, August 1980.

   [12] Srisuresh, P., Holdrege, M., "IP Network Address Translator
   (NAT) Terminology and Considerations", RFC 2663, August, 1999.

   [13] Rekhter, Y., Bates, T., "Scalable Support for Multi-homed Multi-
   provider Connectivity", RFC 2260, January 1998.

   [14] Droms, R., "Dynamic Host Configuration Protocol", RFC 2131,
   March 1997.


13 Author's Address

   Eliot Lear
   Cisco Systems, Inc.
   170 W. Tasman Dr.
   San Jose, CA 95134-1706
   Email: lear@cisco.com
   Phone: +1 (408) 527 4020

14 Intellectual Property Statement

   The IETF takes no position regarding the validity or scope of any
   intellectual property or other rights that might be claimed to
   pertain to the implementation or use of the technology described in
   this document or the extent to which any license under such rights
   might or might not be available; neither does it represent that it
   has made any effort to identify any such rights.  Information on the
   IETF's procedures with respect to rights in standards-track and
   standards-related documentation can be found in BCP-11.  Copies of
   claims of rights made available for publication and any assurances
   of licenses to be made available, or the result of an attempt made
   to obtain a general license or permission for the use of such
   proprietary rights by implementors or users of this specification
   can be obtained from the IETF Secretariat.

   The IETF invites any interested party to bring to its attention any
   copyrights, patents or patent applications, or other proprietary
   rights which may cover technology that may be required to practice
   this standard.  Please address the information to the IETF Executive

16  Full Copyright Statement

   Copyright (C) The Internet Society (2000).  All Rights Reserved.

   This document and translations of it may be copied and furnished to
   others, and derivative works that comment on or otherwise explain it
   or assist in its implementation may be prepared, copied, published
   and distributed, in whole or in part, without restriction of any
   kind, provided that the above copyright notice and this paragraph
   are included on all such copies and derivative works.  However, this
   document itself may not be modified in any way, such as by removing
   the copyright notice or references to the Internet Society or other
   Internet organizations, except as needed for the purpose of
   developing Internet standards in which case the procedures for
   copyrights defined in the Internet Standards process must be
   followed, or as required to translate it into languages other than
   English.  The limited permissions granted above are perpetual and
   will not be revoked by the Internet Society or its successors or
   assigns.  This document and the information contained
   herein is provided on an "AS IS" basis and THE INTERNET SOCIETY AND

17 Expiration Date

   This memo is filed as <draft-lear-middlebox-arch-00.txt>, and expires
   July 31, 2001.