ARMWARE RFC Archive <- RFC Index (8901..9000)

RFC 8930




Internet Engineering Task Force (IETF)                  T. Watteyne, Ed.
Request for Comments: 8930                                Analog Devices
Category: Standards Track                                P. Thubert, Ed.
ISSN: 2070-1721                                            Cisco Systems
                                                              C. Bormann
                                                  Universität Bremen TZI
                                                           November 2020

     On Forwarding 6LoWPAN Fragments over a Multi-Hop IPv6 Network

Abstract

   This document provides generic rules to enable the forwarding of an
   IPv6 over Low-Power Wireless Personal Area Network (6LoWPAN) fragment
   over a route-over network.  Forwarding fragments can improve both
   end-to-end latency and reliability as well as reduce the buffer
   requirements in intermediate nodes; it may be implemented using RFC
   4944 and Virtual Reassembly Buffers (VRBs).

Status of This Memo

   This is an Internet Standards Track document.

   This document is a product of the Internet Engineering Task Force
   (IETF).  It represents the consensus of the IETF community.  It has
   received public review and has been approved for publication by the
   Internet Engineering Steering Group (IESG).  Further information on
   Internet Standards is available in Section 2 of RFC 7841.

   Information about the current status of this document, any errata,
   and how to provide feedback on it may be obtained at
   https://www.rfc-editor.org/info/rfc8930.

Copyright Notice

   Copyright (c) 2020 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
   (https://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.

Table of Contents

   1.  Introduction
   2.  Terminology
     2.1.  Requirements Language
     2.2.  Background
     2.3.  New Terms
   3.  Overview of 6LoWPAN Fragmentation
   4.  Limitations of Per-Hop Fragmentation and Reassembly
     4.1.  Latency
     4.2.  Memory Management and Reliability
   5.  Forwarding Fragments
   6.  Virtual Reassembly Buffer (VRB) Implementation
   7.  Security Considerations
   8.  IANA Considerations
   9.  References
     9.1.  Normative References
     9.2.  Informative References
   Acknowledgments
   Authors' Addresses

1.  Introduction

   The original 6LoWPAN fragmentation is defined in [RFC4944] for use
   over a single Layer 3 hop, though multiple Layer 2 hops in a mesh-
   under network is also possible, and was not modified by the update in
   [RFC6282]. 6LoWPAN operations including fragmentation depend on a
   link-layer security that prevents any rogue access to the network.

   In a route-over 6LoWPAN network, an IP packet is expected to be
   reassembled at each intermediate hop, uncompressed, pushed to Layer 3
   to be routed, and then compressed and fragmented again.  This
   document introduces an alternate approach called 6LoWPAN Fragment
   Forwarding (6LFF) whereby an intermediate node forwards a fragment
   (or the bulk thereof, MTU permitting) without reassembling if the
   next hop is a similar 6LoWPAN link.  The routing decision is made on
   the first fragment of the datagram, which has the IPv6 routing
   information.  The first fragment is forwarded immediately, and a
   state is stored to enable forwarding the next fragments along the
   same path.

   Done right, 6LoWPAN Fragment Forwarding techniques lead to more
   streamlined operations, less buffer bloat, and lower latency.  But it
   may be wasteful when fragments are missing, leading to locked
   resources and low throughput, and it may be misused to the point that
   the end-to-end latency of one packet falls behind that of per-hop
   reassembly.

   This specification provides a generic overview of 6LFF, discusses
   advantages and caveats, and introduces a particular 6LFF technique
   called "Virtual Reassembly Buffer" (VRB) that can be used while
   retaining the message formats defined in [RFC4944].  Basic
   recommendations such as the insertion of an inter-frame gap between
   fragments are provided to avoid the most typical caveats.

2.  Terminology

2.1.  Requirements Language

   The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
   "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and
   "OPTIONAL" in this document are to be interpreted as described in
   BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all
   capitals, as shown here.

2.2.  Background

   Past experience with fragmentation, e.g., as described in "IPv4
   Reassembly Errors at High Data Rates" [RFC4963] and references
   therein, has shown that misassociated or lost fragments can lead to
   poor network behavior and, occasionally, trouble at the application
   layer.  That experience led to the definition of the "Path MTU
   Discovery for IP version 6" [RFC8201] protocol that limits
   fragmentation over the Internet.

   "IP Fragmentation Considered Fragile" [RFC8900] discusses security
   threats that are linked to using IP fragmentation.  The 6LoWPAN
   fragmentation takes place underneath the IP Layer, but some issues
   described there may still apply to 6LoWPAN fragments (as discussed in
   further details in Section 7).

   Readers are expected to be familiar with all the terms and concepts
   that are discussed in "IPv6 over Low-Power Wireless Personal Area
   Networks (6LoWPANs): Overview, Assumptions, Problem Statement, and
   Goals" [RFC4919] and "Transmission of IPv6 Packets over IEEE 802.15.4
   Networks" [RFC4944].

   "Multiprotocol Label Switching Architecture" [RFC3031] states that
   with MPLS,

   |  packets are "labeled" before they are forwarded.  At subsequent
   |  hops, there is no further analysis of the packet's network layer
   |  header.  Rather, the label is used as an index into a table which
   |  specifies the next hop, and a new label.

   The MPLS technique is leveraged in the present specification to
   forward fragments that actually do not have a network-layer header,
   since the fragmentation occurs below IP.

2.3.  New Terms

   This specification uses the following terms:

   6LoWPAN Fragment Forwarding Endpoints:  The 6LFF endpoints are the
      first and last nodes in an unbroken string of 6LFF nodes.  They
      are also the only points where the fragmentation and reassembly
      operations take place.

   Compressed Form:  This specification uses the generic term
      "compressed form" to refer to the format of a datagram after the
      action of [RFC6282] and possibly [RFC8138] for Routing Protocol
      for Low-Power and Lossy Network (RPL) [RFC6550] artifacts.

   Datagram_Size:  The size of the datagram in its compressed form
      before it is fragmented.

   Datagram_Tag:  An identifier of a datagram that is locally unique to
      the Layer 2 sender.  Associated with the link-layer address of the
      sender, this becomes a globally unique identifier for the datagram
      within the duration of its transmission.

   Fragment_Offset:  The offset of a fragment of a datagram in its
      compressed form.

3.  Overview of 6LoWPAN Fragmentation

   Figure 1 illustrates 6LoWPAN fragmentation.  We assume node A
   forwards a packet to node B, possibly as part of a multi-hop route
   between 6LoWPAN Fragment Forwarding endpoints, which may be neither A
   nor B, though 6LoWPAN may compress the IP header better when they are
   both the 6LFF and the 6LoWPAN compression endpoints.

                  +---+                     +---+
           ... ---| A |-------------------->| B |--- ...
                  +---+                     +---+
                                 # (frag. 5)

                123456789                 123456789
               +---------+               +---------+
               |   #  ###|               |###  #   |
               +---------+               +---------+
                  outgoing                incoming
             fragmentation                reassembly
                    buffer                buffer

        Figure 1: Fragmentation at Node A, and Reassembly at Node B

   Typically, node A starts with an uncompressed packet and compacts the
   IPv6 packet using the header compression mechanism defined in
   [RFC6282].  If the resulting 6LoWPAN packet does not fit into a
   single link-layer frame, node A's 6LoWPAN sub-layer cuts it into
   multiple 6LoWPAN fragments, which it transmits as separate link-layer
   frames to node B.  Node B's 6LoWPAN sub-layer reassembles these
   fragments, inflates the compressed header fields back to the original
   IPv6 header, and hands over the full IPv6 packet to its IPv6 layer.

   In Figure 1, a packet forwarded by node A to node B is cut into nine
   fragments, numbered 1 to 9 as follows:

   *  Each fragment is represented by the '#' symbol.

   *  Node A has sent fragments 1, 2, 3, 5, and 6 to node B.

   *  Node B has received fragments 1, 2, 3, and 6 from node A.

   *  Fragment 5 is still being transmitted at the link layer from node
      A to node B.

   The reassembly buffer for 6LoWPAN is indexed in node B by:

   *  a unique identifier of node A (e.g., node A's link-layer address).

   *  the Datagram_Tag chosen by node A for this fragmented datagram.

   Because it may be hard for node B to correlate all possible link-
   layer addresses that node A may use (e.g., short versus long
   addresses), node A must use the same link-layer address to send all
   the fragments of the same datagram to node B.

   Conceptually, the reassembly buffer in node B contains:

   *  a Datagram_Tag as received in the incoming fragments, associated
      with the interface and the link-layer address of node A for which
      the received Datagram_Tag is unique,

   *  the actual packet data from the fragments received so far, in a
      form that makes it possible to detect when the whole packet has
      been received and can be processed or forwarded,

   *  a state indicating the fragments already received,

   *  a Datagram_Size, and

   *  a timer that allows discarding a partially reassembled packet
      after some timeout.

   A fragmentation header is added to each fragment; it indicates what
   portion of the packet that fragment corresponds to.  Section 5.3 of
   [RFC4944] defines the format of the header for the first and
   subsequent fragments.  All fragments are tagged with a 16-bit
   "Datagram_Tag", used to identify which packet each fragment belongs
   to.  Each datagram can be uniquely identified by the sender link-
   layer addresses of the frame that carries it and the Datagram_Tag
   that the sender allocated for this datagram.  [RFC4944] also mandates
   that the first fragment is sent first and with a particular format
   that is different than that of the next fragments.  Each fragment
   except for the first one can be identified within its datagram by the
   datagram-offset.

   Node B's typical behavior, per [RFC4944], is as follows.  Upon
   receiving a fragment from node A with a Datagram_Tag previously
   unseen from node A, node B allocates a buffer large enough to hold
   the entire packet.  The length of the packet is indicated in each
   fragment (the Datagram_Size field), so node B can allocate the buffer
   even if the fragment it receives first is not the first fragment.  As
   fragments come in, node B fills the buffer.  When all fragments have
   been received, node B inflates the compressed header fields into an
   IPv6 header and hands the resulting IPv6 packet to the IPv6 layer,
   which performs the route lookup.  This behavior typically results in
   per-hop fragmentation and reassembly.  That is, the packet is fully
   reassembled, then (re-)fragmented, at every hop.

4.  Limitations of Per-Hop Fragmentation and Reassembly

   There are at least two limitations to doing per-hop fragmentation and
   reassembly.  See [ARTICLE] for detailed simulation results on both
   limitations.

4.1.  Latency

   When reassembling, a node needs to wait for all the fragments to be
   received before being able to re-form the IPv6 packet and possibly
   forwarding it to the next hop.  This repeats at every hop.

   This may result in increased end-to-end latency compared to a case
   where each fragment is forwarded without per-hop reassembly.

4.2.  Memory Management and Reliability

   Constrained nodes have limited memory.  Assuming a reassembly buffer
   for a 6LoWPAN MTU of 1280 bytes as defined in Section 4 of [RFC4944],
   typical nodes only have enough memory for 1-3 reassembly buffers.

   To illustrate this, we use the topology from Figure 2, where nodes A,
   B, C, and D all send packets through node E.  We further assume that
   node E's memory can only hold 3 reassembly buffers.

                  +---+       +---+
          ... --->| A |------>| B |
                  +---+       +---+\
                                    \
                                    +---+    +---+
                                    | E |--->| F | ...
                                    +---+    +---+
                                    /
                                   /
                  +---+       +---+
          ... --->| C |------>| D |
                  +---+       +---+

             Figure 2: Illustrating the Memory Management Issue

   When nodes A, B, and C concurrently send fragmented packets, all
   three reassembly buffers in node E are occupied.  If, at that moment,
   node D also sends a fragmented packet, node E has no option but to
   drop one of the packets, lowering end-to-end reliability.

5.  Forwarding Fragments

   A 6LoWPAN Fragment Forwarding technique makes the routing decision on
   the first fragment, which is always the one with the IPv6 address of
   the destination.  Upon receiving a first fragment, a forwarding node
   (e.g., node B in an A->B->C sequence) that does fragment forwarding
   MUST attempt to create a state and forward the fragment.  This is an
   atomic operation, and if the first fragment cannot be forwarded, then
   the state MUST be removed.

   Since the Datagram_Tag is uniquely associated with the source link-
   layer address of the fragment, the forwarding node MUST assign a new
   Datagram_Tag from its own namespace for the next hop and rewrite the
   fragment header of each fragment with that Datagram_Tag.

   When a forwarding node receives a fragment other than a first
   fragment, it MUST look up state based on the source link-layer
   address and the Datagram_Tag in the received fragment.  If no such
   state is found, the fragment MUST be dropped; otherwise, the fragment
   MUST be forwarded using the information in the state found.

   Compared to Section 3, the conceptual reassembly buffer in node B now
   contains the following, assuming that node B is neither the source
   nor the final destination:

   *  a Datagram_Tag as received in the incoming fragments, associated
      with the interface and the link-layer address of node A for which
      the received Datagram_Tag is unique.

   *  the link-layer address that node B uses as the source to forward
      the fragments.

   *  the interface and the link-layer address of the next-hop C that is
      resolved on the first fragment.

   *  a Datagram_Tag that node B uniquely allocated for this datagram
      and that is used when forwarding the fragments of the datagram.

   *  a buffer for the remainder of a previous fragment left to be sent.

   *  a timer that allows discarding the stale 6LFF state after some
      timeout.  The duration of the timer should be longer than that
      which covers the reassembly at the receiving endpoint.

   A node that has not received the first fragment cannot forward the
   next fragments.  This means that if node B receives a fragment, node
   A was in possession of the first fragment at some point.  To keep the
   operation simple and consistent with [RFC4944], the first fragment
   MUST always be sent first.  When that is done, if node B receives a
   fragment that is not the first and for which it has no state, then
   node B treats it as an error and refrains from creating a state or
   attempting to forward.  This also means that node A should perform
   all its possible retries on the first fragment before it attempts to
   send the next fragments, and that it should abort the datagram and
   release its state if it fails to send the first fragment.

   Fragment forwarding obviates some of the benefits of the 6LoWPAN
   header compression [RFC6282] in intermediate hops.  In return, the
   memory used to store the packet is distributed along the path, which
   limits the buffer-bloat effect.  Multiple fragments may progress
   simultaneously along the network as long as they do not interfere.
   An associated caveat is that on a half-duplex radio, if node A sends
   the next fragment at the same time as node B forwards the previous
   fragment to node C down the path, then node B will miss it.  If node
   C forwards the previous fragment to node D at the same time and on
   the same frequency as node A sends the next fragment to node B, this
   may result in a hidden terminal problem.  In that case, the
   transmission from node C interferes at node B with that from node A,
   unbeknownst to node A.  Consecutive fragments of a same datagram MUST
   be separated with an inter-frame gap that allows one fragment to
   progress beyond the next hop and beyond the interference domain
   before the next shows up.  This can be achieved by interleaving
   packets or fragments sent via different next-hop routers.

6.  Virtual Reassembly Buffer (VRB) Implementation

   The VRB [LWIG-VRB] is a particular incarnation of a 6LFF that can be
   implemented without a change to [RFC4944].

   VRB overcomes the limitations listed in Section 4.  Nodes do not wait
   for the last fragment before forwarding, reducing end-to-end latency.
   Similarly, the memory footprint of VRB is just the VRB table,
   reducing the packet drop probability significantly.

   However, there are other caveats:

   Non-zero Packet Drop Probability:  The abstract data in a VRB table
      entry contains at a minimum the link-layer address of the
      predecessor and the successor, the Datagram_Tag used by the
      predecessor, and the local Datagram_Tag that this node will swap
      with it.  The VRB may need to store a few octets from the last
      fragment that may not have fit within MTU and that will be
      prepended to the next fragment.  This yields a small footprint
      that is 2 orders of magnitude smaller, compared to needing a
      1280-byte reassembly buffer for each packet.  Yet, the size of the
      VRB table necessarily remains finite.  In the extreme case where a
      node is required to concurrently forward more packets than it has
      entries in its VRB table, packets are dropped.

   No Fragment Recovery:  There is no mechanism in VRB for the node that
      reassembles a packet to request a single missing fragment.
      Dropping a fragment requires the whole packet to be resent.  This
      causes unnecessary traffic, as fragments are forwarded even when
      the destination node can never construct the original IPv6 packet.

   No Per-Fragment Routing:  All subsequent fragments follow the same
      sequence of hops from the source to the destination node as the
      first fragment, because the IP header is required in order to
      route the fragment and is only present in the first fragment.  A
      side effect is that the first fragment must always be forwarded
      first.

   The severity and occurrence of these caveats depend on the link layer
   used.  Whether they are acceptable depends entirely on the
   requirements the application places on the network.

   If the caveats are present and not acceptable for the application,
   alternative specifications may define new protocols to overcome them.
   One example is [RFC8931], which specifies a 6LFF technique that
   allows the end-to-end fragment recovery between the 6LFF endpoints.

7.  Security Considerations

   An attacker can perform a Denial-of-Service (DoS) attack on a node
   implementing VRB by generating a large number of bogus "fragment 1"
   fragments without sending subsequent fragments.  This causes the VRB
   table to fill up.  Note that the VRB does not need to remember the
   full datagram as received so far but only possibly a few octets from
   the last fragment that could not fit in it.  It is expected that an
   implementation protects itself to keep the number of VRBs within
   capacity, and that old VRBs are protected by a timer of a reasonable
   duration for the technology and destroyed upon timeout.

   Secure joining and the link-layer security that it sets up protects
   against those attacks from network outsiders.

   "IP Fragmentation Considered Fragile" [RFC8900] discusses security
   threats and other caveats that are linked to using IP fragmentation.
   The 6LoWPAN fragmentation takes place underneath the IP Layer, but
   some issues described there may still apply to 6LoWPAN fragments.

   *  Overlapping fragment attacks are possible with 6LoWPAN fragments,
      but there is no known firewall operation that would work on
      6LoWPAN fragments at the time of this writing, so the exposure is
      limited.  An implementation of a firewall SHOULD NOT forward
      fragments but instead should recompose the IP packet, check it in
      the uncompressed form, and then forward it again as fragments if
      necessary.  Overlapping fragments are acceptable as long as they
      contain the same payload.  The firewall MUST drop the whole packet
      if overlapping fragments are encountered that result in different
      data at the same offset.

   *  Resource-exhaustion attacks are certainly possible and a sensitive
      issue in a constrained network.  An attacker can perform a DoS
      attack on a node implementing VRB by generating a large number of
      bogus first fragments without sending subsequent fragments.  This
      causes the VRB table to fill up.  When hop-by-hop reassembly is
      used, the same attack can be more damaging if the node allocates a
      full Datagram_Size for each bogus first fragment.  With the VRB,
      the attack can be performed remotely on all nodes along a path,
      but each node suffers a lesser hit.  This is because the VRB does
      not need to remember the full datagram as received so far but only
      possibly a few octets from the last fragment that could not fit in
      it.  An implementation MUST protect itself to keep the number of
      VRBs within capacity and to ensure that old VRBs are protected by
      a timer of a reasonable duration for the technology and destroyed
      upon timeout.

   *  Attacks based on predictable fragment identification values are
      also possible but can be avoided.  The Datagram_Tag SHOULD be
      assigned pseudorandomly in order to reduce the risk of such
      attacks.  A larger size of the Datagram_Tag makes the guessing
      more difficult and reduces the chances of an accidental reuse
      while the original packet is still in flight, at the expense of
      more space in each frame.  Nonetheless, some level of risk remains
      because an attacker that is able to authenticate to and send
      traffic on the network can guess a valid Datagram_Tag value, since
      there are only a limited number of possible values.

   *  Evasion of Network Intrusion Detection Systems (NIDSs) leverages
      ambiguity in the reassembly of the fragment.  This attack makes
      little sense in the context of this specification since the
      fragmentation happens within the Low-Power and Lossy Network
      (LLN), meaning that the intruder should already be inside to
      perform the attack.  NIDS systems would probably not be installed
      within the LLN either but rather at a bottleneck at the exterior
      edge of the network.

8.  IANA Considerations

   This document has no IANA actions.

9.  References

9.1.  Normative References

   [RFC2119]  Bradner, S., "Key words for use in RFCs to Indicate
              Requirement Levels", BCP 14, RFC 2119,
              DOI 10.17487/RFC2119, March 1997,
              <https://www.rfc-editor.org/info/rfc2119>.

   [RFC4919]  Kushalnagar, N., Montenegro, G., and C. Schumacher, "IPv6
              over Low-Power Wireless Personal Area Networks (6LoWPANs):
              Overview, Assumptions, Problem Statement, and Goals",
              RFC 4919, DOI 10.17487/RFC4919, August 2007,
              <https://www.rfc-editor.org/info/rfc4919>.

   [RFC4944]  Montenegro, G., Kushalnagar, N., Hui, J., and D. Culler,
              "Transmission of IPv6 Packets over IEEE 802.15.4
              Networks", RFC 4944, DOI 10.17487/RFC4944, September 2007,
              <https://www.rfc-editor.org/info/rfc4944>.

   [RFC8174]  Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC
              2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174,
              May 2017, <https://www.rfc-editor.org/info/rfc8174>.

9.2.  Informative References

   [ARTICLE]  Tanaka, Y., Minet, P., and T. Watteyne, "6LoWPAN Fragment
              Forwarding", IEEE Communications Standards Magazine, Vol.
              3, Issue 1, pp. 35-39, DOI 10.1109/MCOMSTD.2019.1800029,
              March 2019,
              <https://ieeexplore.ieee.org/abstract/document/8771317>.

   [LWIG-VRB] Bormann, C. and T. Watteyne, "Virtual reassembly buffers
              in 6LoWPAN", Work in Progress, Internet-Draft, draft-ietf-
              lwig-6lowpan-virtual-reassembly-02, 9 March 2020,
              <https://tools.ietf.org/html/draft-ietf-lwig-6lowpan-
              virtual-reassembly-02>.

   [RFC3031]  Rosen, E., Viswanathan, A., and R. Callon, "Multiprotocol
              Label Switching Architecture", RFC 3031,
              DOI 10.17487/RFC3031, January 2001,
              <https://www.rfc-editor.org/info/rfc3031>.

   [RFC4963]  Heffner, J., Mathis, M., and B. Chandler, "IPv4 Reassembly
              Errors at High Data Rates", RFC 4963,
              DOI 10.17487/RFC4963, July 2007,
              <https://www.rfc-editor.org/info/rfc4963>.

   [RFC6282]  Hui, J., Ed. and P. Thubert, "Compression Format for IPv6
              Datagrams over IEEE 802.15.4-Based Networks", RFC 6282,
              DOI 10.17487/RFC6282, September 2011,
              <https://www.rfc-editor.org/info/rfc6282>.

   [RFC6550]  Winter, T., Ed., Thubert, P., Ed., Brandt, A., Hui, J.,
              Kelsey, R., Levis, P., Pister, K., Struik, R., Vasseur,
              JP., and R. Alexander, "RPL: IPv6 Routing Protocol for
              Low-Power and Lossy Networks", RFC 6550,
              DOI 10.17487/RFC6550, March 2012,
              <https://www.rfc-editor.org/info/rfc6550>.

   [RFC8138]  Thubert, P., Ed., Bormann, C., Toutain, L., and R. Cragie,
              "IPv6 over Low-Power Wireless Personal Area Network
              (6LoWPAN) Routing Header", RFC 8138, DOI 10.17487/RFC8138,
              April 2017, <https://www.rfc-editor.org/info/rfc8138>.

   [RFC8201]  McCann, J., Deering, S., Mogul, J., and R. Hinden, Ed.,
              "Path MTU Discovery for IP version 6", STD 87, RFC 8201,
              DOI 10.17487/RFC8201, July 2017,
              <https://www.rfc-editor.org/info/rfc8201>.

   [RFC8900]  Bonica, R., Baker, F., Huston, G., Hinden, R., Troan, O.,
              and F. Gont, "IP Fragmentation Considered Fragile",
              BCP 230, RFC 8900, DOI 10.17487/RFC8900, September 2020,
              <https://www.rfc-editor.org/info/rfc8900>.

   [RFC8931]  Thubert, P., Ed., "IPv6 over Low-Power Wireless Personal
              Area Network (6LoWPAN) Selective Fragment Recovery",
              RFC 8931, DOI 10.17487/RFC8931, November 2020,
              <https://www.rfc-editor.org/info/rfc8931>.

Acknowledgments

   The authors would like to thank Carles Gomez Montenegro, Yasuyuki
   Tanaka, Ines Robles, and Dave Thaler for their in-depth review of
   this document and suggestions for improvement.  Many thanks to
   Georgios Papadopoulos and Dominique Barthel for their contributions
   during the WG activities.  And many thanks as well to Roman Danyliw,
   Barry Leiba, Murray Kucherawy, Derrell Piper, Sarah Banks, Joerg Ott,
   Francesca Palombini, Mirja Kühlewind, Éric Vyncke, and especially
   Benjamin Kaduk for their constructive reviews through the IETF last
   call and IESG process.

Authors' Addresses

   Thomas Watteyne (editor)
   Analog Devices
   32990 Alvarado-Niles Road, Suite 910
   Union City, CA 94587
   United States of America

   Email: thomas.watteyne@analog.com

   Pascal Thubert (editor)
   Cisco Systems, Inc
   Building D
   45 Allee des Ormes - BP1200
   06254 Mougins - Sophia Antipolis
   France

   Phone: +33 497 23 26 34
   Email: pthubert@cisco.com

   Carsten Bormann
   Universität Bremen TZI
   Postfach 330440
   D-28359 Bremen
   Germany

   Phone: +49-421-218-63921
   Email: cabo@tzi.org