RFC 8900

Internet Engineering Task Force (IETF)                         R. Bonica
Request for Comments: 8900                              Juniper Networks
BCP: 230                                                        F. Baker
Category: Best Current Practice                             Unaffiliated
ISSN: 2070-1721                                                G. Huston
                                                               R. Hinden
                                                    Check Point Software
                                                                O. Troan
                                                                 F. Gont
                                                            SI6 Networks
                                                          September 2020

                  IP Fragmentation Considered Fragile


   This document describes IP fragmentation and explains how it
   introduces fragility to Internet communication.

   This document also proposes alternatives to IP fragmentation and
   provides recommendations for developers and network operators.

Status of This Memo

   This memo documents an Internet Best Current Practice.

   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
   BCPs 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

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
     1.1.  Requirements Language
   2.  IP Fragmentation
     2.1.  Links, Paths, MTU, and PMTU
     2.2.  Fragmentation Procedures
     2.3.  Upper-Layer Reliance on IP Fragmentation
   3.  Increased Fragility
     3.1.  Virtual Reassembly
     3.2.  Policy-Based Routing
     3.3.  Network Address Translation (NAT)
     3.4.  Stateless Firewalls
     3.5.  Equal-Cost Multipath, Link Aggregate Groups, and Stateless
           Load Balancers
     3.6.  IPv4 Reassembly Errors at High Data Rates
     3.7.  Security Vulnerabilities
     3.8.  PMTU Black-Holing Due to ICMP Loss
       3.8.1.  Transient Loss
       3.8.2.  Incorrect Implementation of Security Policy
       3.8.3.  Persistent Loss Caused by Anycast
       3.8.4.  Persistent Loss Caused by Unidirectional Routing
     3.9.  Black-Holing Due to Filtering or Loss
   4.  Alternatives to IP Fragmentation
     4.1.  Transport-Layer Solutions
     4.2.  Application-Layer Solutions
   5.  Applications That Rely on IPv6 Fragmentation
     5.1.  Domain Name Service (DNS)
     5.2.  Open Shortest Path First (OSPF)
     5.3.  Packet-in-Packet Encapsulations
     5.4.  UDP Applications Enhancing Performance
   6.  Recommendations
     6.1.  For Application and Protocol Developers
     6.2.  For System Developers
     6.3.  For Middlebox Developers
     6.4.  For ECMP, LAG, and Load-Balancer Developers And Operators
     6.5.  For Network Operators
   7.  IANA Considerations
   8.  Security Considerations
   9.  References
     9.1.  Normative References
     9.2.  Informative References

   Authors' Addresses

1.  Introduction

   Operational experience [Kent] [Huston] [RFC7872] reveals that IP
   fragmentation introduces fragility to Internet communication.  This
   document describes IP fragmentation and explains the fragility it
   introduces.  It also proposes alternatives to IP fragmentation and
   provides recommendations for developers and network operators.

   While this document identifies issues associated with IP
   fragmentation, it does not recommend deprecation.  Legacy protocols
   that depend upon IP fragmentation would do well to be updated to
   remove that dependency.  However, some applications and environments
   (see Section 5) require IP fragmentation.  In these cases, the
   protocol will continue to rely on IP fragmentation, but the designer
   should be aware that fragmented packets may result in black holes.  A
   design should include appropriate safeguards.

   Rather than deprecating IP fragmentation, this document recommends
   that upper-layer protocols address the problem of fragmentation at
   their layer, reducing their reliance on IP fragmentation to the
   greatest degree possible.

1.1.  Requirements Language

   The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
   "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.  IP Fragmentation

2.1.  Links, Paths, MTU, and PMTU

   An Internet path connects a source node to a destination node.  A
   path may contain links and routers.  If a path contains more than one
   link, the links are connected in series, and a router connects each
   link to the next.

   Internet paths are dynamic.  Assume that the path from one node to
   another contains a set of links and routers.  If a link or a router
   fails, the path can also change so that it includes a different set
   of links and routers.

   Each link is constrained by the number of octets that it can convey
   in a single IP packet.  This constraint is called the link Maximum
   Transmission Unit (MTU).  IPv4 [RFC0791] requires every link to
   support an MTU of 68 octets or greater (see NOTE 1).  IPv6 [RFC8200]
   similarly requires every link to support an MTU of 1280 octets or
   greater.  These are called the IPv4 and IPv6 minimum link MTUs.

   Some links, and some ways of using links, result in additional
   variable overhead.  For the simple case of tunnels, this document
   defers to other documents.  For other cases, such as MPLS, this
   document considers the link MTU to include appropriate allowance for
   any such overhead.

   Likewise, each Internet path is constrained by the number of octets
   that it can convey in a single IP packet.  This constraint is called
   the Path MTU (PMTU).  For any given path, the PMTU is equal to the
   smallest of its link MTUs.  Because Internet paths are dynamic, PMTU
   is also dynamic.

   For reasons described below, source nodes estimate the PMTU between
   themselves and destination nodes.  A source node can produce
   extremely conservative PMTU estimates in which:

   *  The estimate for each IPv4 path is equal to the IPv4 minimum link

   *  The estimate for each IPv6 path is equal to the IPv6 minimum link

   While these conservative estimates are guaranteed to be less than or
   equal to the actual PMTU, they are likely to be much less than the
   actual PMTU.  This may adversely affect upper-layer protocol

   By executing Path MTU Discovery (PMTUD) procedures [RFC1191]
   [RFC8201], a source node can maintain a less conservative estimate of
   the PMTU between itself and a destination node.  In PMTUD, the source
   node produces an initial PMTU estimate.  This initial estimate is
   equal to the MTU of the first link along the path to the destination
   node.  It can be greater than the actual PMTU.

   Having produced an initial PMTU estimate, the source node sends non-
   fragmentable IP packets to the destination node (see NOTE 2).  If one
   of these packets is larger than the actual PMTU, a downstream router
   will not be able to forward the packet through the next link along
   the path.  Therefore, the downstream router drops the packet and
   sends an Internet Control Message Protocol (ICMP) [RFC0792] [RFC4443]
   Packet Too Big (PTB) message to the source node (see NOTE 3).  The
   ICMP PTB message indicates the MTU of the link through which the
   packet could not be forwarded.  The source node uses this information
   to refine its PMTU estimate.

   PMTUD produces a running estimate of the PMTU between a source node
   and a destination node.  Because PMTU is dynamic, the PMTU estimate
   can be larger than the actual PMTU.  In order to detect PMTU
   increases, PMTUD occasionally resets the PMTU estimate to its initial
   value and repeats the procedure described above.

   Ideally, PMTUD operates as described above.  However, in some
   scenarios, PMTUD fails.  For example:

   *  PMTUD relies on the network's ability to deliver ICMP PTB messages
      to the source node.  If the network cannot deliver ICMP PTB
      messages to the source node, PMTUD fails.

   *  PMTUD is susceptible to attack because ICMP messages are easily
      forged [RFC5927] and not authenticated by the receiver.  Such
      attacks can cause PMTUD to produce unnecessarily conservative PMTU

   NOTE 1:  In IPv4, every host must be able to reassemble a packet
      whose length is less than or equal to 576 octets.  However, the
      IPv4 minimum link MTU is not 576.  Section 3.2 of RFC 791
      [RFC0791] explicitly states that the IPv4 minimum link MTU is 68

   NOTE 2:  A non-fragmentable packet can be fragmented at its source.
      However, it cannot be fragmented by a downstream node.  An IPv4
      packet whose Don't Fragment (DF) bit is set to 0 is fragmentable.
      An IPv4 packet whose DF bit is set to 1 is non-fragmentable.  All
      IPv6 packets are also non-fragmentable.

   NOTE 3:  The ICMP PTB message has two instantiations.  In ICMPv4
      [RFC0792], the ICMP PTB message is a Destination Unreachable
      message with Code equal to 4 (fragmentation needed and DF set).
      This message was augmented by [RFC1191] to indicate the MTU of the
      link through which the packet could not be forwarded.  In ICMPv6
      [RFC4443], the ICMP PTB message is a Packet Too Big Message with
      Code equal to 0.  This message also indicates the MTU of the link
      through which the packet could not be forwarded.

2.2.  Fragmentation Procedures

   When an upper-layer protocol submits data to the underlying IP
   module, and the resulting IP packet's length is greater than the
   PMTU, the packet is divided into fragments.  Each fragment includes
   an IP header and a portion of the original packet.

   [RFC0791] describes IPv4 fragmentation procedures.  An IPv4 packet
   whose DF bit is set to 1 may be fragmented by the source node, but
   may not be fragmented by a downstream router.  An IPv4 packet whose
   DF bit is set to 0 may be fragmented by the source node or by a
   downstream router.  When an IPv4 packet is fragmented, all IP options
   (which are within the IPv4 header) appear in the first fragment, but
   only options whose "copy" bit is set to 1 appear in subsequent

   [RFC8200], notably in Section 4.5, describes IPv6 fragmentation
   procedures.  An IPv6 packet may be fragmented only at the source
   node.  When an IPv6 packet is fragmented, all extension headers
   appear in the first fragment, but only per-fragment headers appear in
   subsequent fragments.  Per-fragment headers include the following:

   *  The IPv6 header.

   *  The Hop-by-Hop Options header (if present).

   *  The Destination Options header (if present and if it precedes a
      Routing header).

   *  The Routing header (if present).

   *  The Fragment header.

   In IPv4, the upper-layer header usually appears in the first
   fragment, due to the sizes of the headers involved.  In IPv6, the
   upper-layer header must appear in the first fragment.

2.3.  Upper-Layer Reliance on IP Fragmentation

   Upper-layer protocols can operate in the following modes:

   *  Do not rely on IP fragmentation.

   *  Rely on IP fragmentation by the source node only.

   *  Rely on IP fragmentation by any node.

   Upper-layer protocols running over IPv4 can operate in all of the
   above-mentioned modes.  Upper-layer protocols running over IPv6 can
   operate in the first and second modes only.

   Upper-layer protocols that operate in the first two modes (above)
   require access to the PMTU estimate.  In order to fulfill this
   requirement, they can:

   *  Estimate the PMTU to be equal to the IPv4 or IPv6 minimum link

   *  Access the estimate that PMTUD produced.

   *  Execute PMTUD procedures themselves.

   *  Execute Packetization Layer PMTUD (PLPMTUD) procedures [RFC4821]

   According to PLPMTUD procedures, the upper-layer protocol maintains a
   running PMTU estimate.  It does so by sending probe packets of
   various sizes to its upper-layer peer and receiving acknowledgements.
   This strategy differs from PMTUD in that it relies on acknowledgement
   of received messages, as opposed to ICMP PTB messages concerning
   dropped messages.  Therefore, PLPMTUD does not rely on the network's
   ability to deliver ICMP PTB messages to the source.

3.  Increased Fragility

   This section explains how IP fragmentation introduces fragility to
   Internet communication.

3.1.  Virtual Reassembly

   Virtual reassembly is a procedure in which a device conceptually
   reassembles a packet, forwards its fragments, and discards the
   reassembled copy.  In Address plus Port (A+P) [RFC6346] and Carrier
   Grade NAT (CGN) [RFC6888], virtual reassembly is required in order to
   correctly translate fragment addresses.  It could be useful to
   address the problems in Sections 3.2, 3.3, 3.4, and 3.5.

   Virtual reassembly is computationally expensive and holds state for
   indeterminate periods of time.  Therefore, it is prone to errors and
   attacks (Section 3.7).

3.2.  Policy-Based Routing

   IP fragmentation causes problems for routers that implement policy-
   based routing.

   When a router receives a packet, it identifies the next hop on route
   to the packet's destination and forwards the packet to that next hop.
   In order to identify the next hop, the router interrogates a local
   data structure called the Forwarding Information Base (FIB).

   Normally, the FIB contains destination-based entries that map a
   destination prefix to a next hop.  Policy-based routing allows
   destination-based and policy-based entries to coexist in the same
   FIB.  A policy-based FIB entry maps multiple fields, drawn from
   either the IP or transport-layer header, to a next hop.

   |Entry| Type              | Dest. Prefix    | Next  | Next Hop      |
   |     |                   |                 | Hdr / |               |
   |     |                   |                 | Dest. |               |
   |     |                   |                 | Port  |               |
   |  1  | Destination-based | 2001:db8::1/128 | Any / | 2001:db8:2::2 |
   |     |                   |                 | Any   |               |
   |  2  | Policy-based      | 2001:db8::1/128 | TCP / | 2001:db8:3::3 |
   |     |                   |                 | 80    |               |

                     Table 1: Policy-Based Routing FIB

   Assume that a router maintains the FIB in Table 1.  The first FIB
   entry is destination-based.  It maps a destination prefix
   2001:db8::1/128 to a next hop 2001:db8:2::2.  The second FIB entry is
   policy-based.  It maps the same destination prefix 2001:db8::1/128
   and a destination port (TCP / 80) to a different next hop
   (2001:db8:3::3).  The second entry is more specific than the first.

   When the router receives the first fragment of a packet that is
   destined for TCP port 80 on 2001:db8::1, it interrogates the FIB.
   Both FIB entries satisfy the query.  The router selects the second
   FIB entry because it is more specific and forwards the packet to

   When the router receives the second fragment of the packet, it
   interrogates the FIB again.  This time, only the first FIB entry
   satisfies the query, because the second fragment contains no
   indication that the packet is destined for TCP port 80.  Therefore,
   the router selects the first FIB entry and forwards the packet to

   Policy-based routing is also known as filter-based forwarding.

3.3.  Network Address Translation (NAT)

   IP fragmentation causes problems for Network Address Translation
   (NAT) devices.  When a NAT device detects a new, outbound flow, it
   maps that flow's source port and IP address to another source port
   and IP address.  Having created that mapping, the NAT device

   *  The source IP address and source port on each outbound packet.

   *  The destination IP address and destination port on each inbound

   A+P [RFC6346] and Carrier Grade NAT (CGN) [RFC6888] are two common
   NAT strategies.  In both approaches, the NAT device must virtually
   reassemble fragmented packets in order to translate and forward each

3.4.  Stateless Firewalls

   As discussed in more detail in Section 3.7, IP fragmentation causes
   problems for stateless firewalls whose rules include TCP and UDP
   ports.  Because port information is only available in the first
   fragment and not available in the subsequent fragments, the firewall
   is limited to the following options:

   *  Accept all subsequent fragments, possibly admitting certain
      classes of attack.

   *  Block all subsequent fragments, possibly blocking legitimate

   Neither option is attractive.

3.5.  Equal-Cost Multipath, Link Aggregate Groups, and Stateless Load

   IP fragmentation causes problems for Equal-Cost Multipath (ECMP),
   Link Aggregate Groups (LAG), and other stateless load-distribution
   technologies.  In order to assign a packet or packet fragment to a
   link, an intermediate node executes a hash (i.e., load-distributing)
   algorithm.  The following paragraphs describe a commonly deployed
   hash algorithm.

   If the packet or packet fragment contains a transport-layer header,
   the algorithm accepts the following 5-tuple as input:

   *  IP Source Address.

   *  IP Destination Address.

   *  IPv4 Protocol or IPv6 Next Header.

   *  transport-layer source port.

   *  transport-layer destination port.

   If the packet or packet fragment does not contain a transport-layer
   header, the algorithm accepts only the following 3-tuple as input:

   *  IP Source Address.

   *  IP Destination Address.

   *  IPv4 Protocol or IPv6 Next Header.

   Therefore, non-fragmented packets belonging to a flow can be assigned
   to one link while fragmented packets belonging to the same flow can
   be divided between that link and another.  This can cause suboptimal
   load distribution.

   [RFC6438] offers a partial solution to this problem for IPv6 devices
   only.  According to [RFC6438]:

   |  At intermediate routers that perform load distribution, the hash
   |  algorithm used to determine the outgoing component-link in an ECMP
   |  and/or LAG toward the next hop MUST minimally include the 3-tuple
   |  {dest addr, source addr, flow label} and MAY also include the
   |  remaining components of the 5-tuple.

   If the algorithm includes only the 3-tuple {dest addr, source addr,
   flow label}, it will assign all fragments belonging to a packet to
   the same link.  (See [RFC6437] and [RFC7098]).

   In order to avoid the problem described above, implementations SHOULD
   implement the recommendations provided in Section 6.4 of this

3.6.  IPv4 Reassembly Errors at High Data Rates

   IPv4 fragmentation is not sufficiently robust for use under some
   conditions in today's Internet.  At high data rates, the 16-bit IP
   identification field is not large enough to prevent duplicate IDs,
   resulting in frequent incorrectly assembled IP fragments, and the TCP
   and UDP checksums are insufficient to prevent the resulting corrupted
   datagrams from being delivered to upper-layer protocols.  [RFC4963]
   describes some easily reproduced experiments demonstrating the
   problem and discusses some of the operational implications of these

   These reassembly issues do not occur as frequently in IPv6 because
   the IPv6 identification field is 32 bits long.

3.7.  Security Vulnerabilities

   Security researchers have documented several attacks that exploit IP
   fragmentation.  The following are examples:

   *  Overlapping fragment attacks [RFC1858] [RFC3128] [RFC5722].

   *  Resource exhaustion attacks.

   *  Attacks based on predictable fragment identification values

   *  Evasion of Network Intrusion Detection Systems (NIDS)

   In the overlapping fragment attack, an attacker constructs a series
   of packet fragments.  The first fragment contains an IP header, a
   transport-layer header, and some transport-layer payload.  This
   fragment complies with local security policy and is allowed to pass
   through a stateless firewall.  A second fragment, having a nonzero
   offset, overlaps with the first fragment.  The second fragment also
   passes through the stateless firewall.  When the packet is
   reassembled, the transport-layer header from the first fragment is
   overwritten by data from the second fragment.  The reassembled packet
   does not comply with local security policy.  Had it traversed the
   firewall in one piece, the firewall would have rejected it.

   A stateless firewall cannot protect against the overlapping fragment
   attack.  However, destination nodes can protect against the
   overlapping fragment attack by implementing the procedures described
   in RFC 1858, RFC 3128, and RFC 8200.  These reassembly procedures
   detect the overlap and discard the packet.

   The fragment reassembly algorithm is a stateful procedure in an
   otherwise stateless protocol.  Therefore, it can be exploited by
   resource exhaustion attacks.  An attacker can construct a series of
   fragmented packets with one fragment missing from each packet so that
   the reassembly is impossible.  Thus, this attack causes resource
   exhaustion on the destination node, possibly denying reassembly
   services to other flows.  This type of attack can be mitigated by
   flushing fragment reassembly buffers when necessary, at the expense
   of possibly dropping legitimate fragments.

   Each IP fragment contains an "Identification" field that destination
   nodes use to reassemble fragmented packets.  Some implementations set
   the Identification field to a predictable value, thus making it easy
   for an attacker to forge malicious IP fragments that would cause the
   reassembly procedure for legitimate packets to fail.

   NIDS aims at identifying malicious activity by analyzing network
   traffic.  Ambiguity in the possible result of the fragment reassembly
   process may allow an attacker to evade these systems.  Many of these
   systems try to mitigate some of these evasion techniques (e.g., by
   computing all possible outcomes of the fragment reassembly process,
   at the expense of increased processing requirements).

3.8.  PMTU Black-Holing Due to ICMP Loss

   As mentioned in Section 2.3, upper-layer protocols can be configured
   to rely on PMTUD.  Because PMTUD relies upon the network to deliver
   ICMP PTB messages, those protocols also rely on the networks to
   deliver ICMP PTB messages.

   According to [RFC4890], ICMPv6 PTB messages must not be filtered.
   However, ICMP PTB delivery is not reliable.  It is subject to both
   transient and persistent loss.

   Transient loss of ICMP PTB messages can cause transient PMTU black
   holes.  When the conditions contributing to transient loss abate, the
   network regains its ability to deliver ICMP PTB messages and
   connectivity between the source and destination nodes is restored.
   Section 3.8.1 of this document describes conditions that lead to
   transient loss of ICMP PTB messages.

   Persistent loss of ICMP PTB messages can cause persistent black
   holes.  Sections 3.8.2, 3.8.3, and 3.8.4 of this document describe
   conditions that lead to persistent loss of ICMP PTB messages.

   The problem described in this section is specific to PMTUD.  It does
   not occur when the upper-layer protocol obtains its PMTU estimate
   from PLPMTUD or from any other source.

3.8.1.  Transient Loss

   The following factors can contribute to transient loss of ICMP PTB

   *  Network congestion.

   *  Packet corruption.

   *  Transient routing loops.

   *  ICMP rate limiting.

   The effect of rate limiting may be severe, as RFC 4443 recommends
   strict rate limiting of ICMPv6 traffic.

3.8.2.  Incorrect Implementation of Security Policy

   Incorrect implementation of security policy can cause persistent loss
   of ICMP PTB messages.

   For example, assume that a Customer Premises Equipment (CPE) router
   implements the following zone-based security policy:

   *  Allow any traffic to flow from the inside zone to the outside

   *  Do not allow any traffic to flow from the outside zone to the
      inside zone unless it is part of an existing flow (i.e., it was
      elicited by an outbound packet).

   When a correct implementation of the above-mentioned security policy
   receives an ICMP PTB message, it examines the ICMP PTB payload in
   order to determine whether the original packet (i.e., the packet that
   elicited the ICMP PTB message) belonged to an existing flow.  If the
   original packet belonged to an existing flow, the implementation
   allows the ICMP PTB to flow from the outside zone to the inside zone.
   If not, the implementation discards the ICMP PTB message.

   When an incorrect implementation of the above-mentioned security
   policy receives an ICMP PTB message, it discards the packet because
   its source address is not associated with an existing flow.

   The security policy described above has been implemented incorrectly
   on many consumer CPE routers.

3.8.3.  Persistent Loss Caused by Anycast

   Anycast can cause persistent loss of ICMP PTB messages.  Consider the
   example below:

   A DNS client sends a request to an anycast address.  The network
   routes that DNS request to the nearest instance of that anycast
   address (i.e., a DNS server).  The DNS server generates a response
   and sends it back to the DNS client.  While the response does not
   exceed the DNS server's PMTU estimate, it does exceed the actual

   A downstream router drops the packet and sends an ICMP PTB message
   the packet's source (i.e., the anycast address).  The network routes
   the ICMP PTB message to the anycast instance closest to the
   downstream router.  That anycast instance may not be the DNS server
   that originated the DNS response.  It may be another DNS server with
   the same anycast address.  The DNS server that originated the
   response may never receive the ICMP PTB message and may never update
   its PMTU estimate.

3.8.4.  Persistent Loss Caused by Unidirectional Routing

   Unidirectional routing can cause persistent loss of ICMP PTB
   messages.  Consider the example below:

   A source node sends a packet to a destination node.  All intermediate
   nodes maintain a route to the destination node but do not maintain a
   route to the source node.  In this case, when an intermediate node
   encounters an MTU issue, it cannot send an ICMP PTB message to the
   source node.

3.9.  Black-Holing Due to Filtering or Loss

   In RFC 7872, researchers sampled Internet paths to determine whether
   they would convey packets that contain IPv6 extension headers.
   Sampled paths terminated at popular Internet sites (e.g., popular
   web, mail, and DNS servers).

   The study revealed that at least 28% of the sampled paths did not
   convey packets containing the IPv6 Fragment extension header.  In
   most cases, fragments were dropped in the destination autonomous
   system.  In other cases, the fragments were dropped in transit
   autonomous systems.

   Another study [Huston] confirmed this finding.  It reported that 37%
   of sampled endpoints used IPv6-capable DNS resolvers that were
   incapable of receiving a fragmented IPv6 response.

   It is difficult to determine why network operators drop fragments.
   Possible causes follow:

   *  Hardware inability to process fragmented packets.

   *  Failure to change vendor defaults.

   *  Unintentional misconfiguration.

   *  Intentional configuration (e.g., network operators consciously
      chooses to drop IPv6 fragments in order to address the issues
      raised in Sections 3.2 through 3.8, above.)

4.  Alternatives to IP Fragmentation

4.1.  Transport-Layer Solutions

   The Transport Control Protocol (TCP) [RFC0793]) can be operated in a
   mode that does not require IP fragmentation.

   Applications submit a stream of data to TCP.  TCP divides that stream
   of data into segments, with no segment exceeding the TCP Maximum
   Segment Size (MSS).  Each segment is encapsulated in a TCP header and
   submitted to the underlying IP module.  The underlying IP module
   prepends an IP header and forwards the resulting packet.

   If the TCP MSS is sufficiently small, then the underlying IP module
   never produces a packet whose length is greater than the actual PMTU.
   Therefore, IP fragmentation is not required.

   TCP offers the following mechanisms for MSS management:

   *  Manual configuration.

   *  PMTUD.

   *  PLPMTUD.

   Manual configuration is always applicable.  If the MSS is configured
   to a sufficiently low value, the IP layer will never produce a packet
   whose length is greater than the protocol minimum link MTU.  However,
   manual configuration prevents TCP from taking advantage of larger
   link MTUs.

   Upper-layer protocols can implement PMTUD in order to discover and
   take advantage of larger Path MTUs.  However, as mentioned in
   Section 2.1, PMTUD relies upon the network to deliver ICMP PTB
   messages.  Therefore, PMTUD can only provide an estimate of the PMTU
   in environments where the risk of ICMP PTB loss is acceptable (e.g.,
   known to not be filtered).

   By contrast, PLPMTUD does not rely upon the network's ability to
   deliver ICMP PTB messages.  It utilizes probe messages sent as TCP
   segments to determine whether the probed PMTU can be successfully
   used across the network path.  In PLPMTUD, probing is separated from
   congestion control, so that loss of a TCP probe segment does not
   cause a reduction of the congestion control window.  [RFC4821]
   defines PLPMTUD procedures for TCP.

   While TCP will never knowingly cause the underlying IP module to emit
   a packet that is larger than the PMTU estimate, it can cause the
   underlying IP module to emit a packet that is larger than the actual
   PMTU.  For example, if routing changes and as a result the PMTU
   becomes smaller, TCP will not know until the ICMP PTB message
   arrives.  If this occurs, the packet is dropped, the PMTU estimate is
   updated, the segment is divided into smaller segments, and each
   smaller segment is submitted to the underlying IP module.

   The Datagram Congestion Control Protocol (DCCP) [RFC4340] and the
   Stream Control Transmission Protocol (SCTP) [RFC4960] also can be
   operated in a mode that does not require IP fragmentation.  They both
   accept data from an application and divide that data into segments,
   with no segment exceeding a maximum size.

   DCCP offers manual configuration, PMTUD, and PLPMTUD as mechanisms
   for managing that maximum size.  Datagram protocols can also
   implement PLPMTUD to estimate the PMTU via [RFC8899].  This proposes
   procedures for performing PLPMTUD with UDP, UDP options, SCTP, QUIC,
   and other datagram protocols.

   Currently, User Datagram Protocol (UDP) [RFC0768] lacks a
   fragmentation mechanism of its own and relies on IP fragmentation.
   However, [UDP-OPTIONS] proposes a fragmentation mechanism for UDP.

4.2.  Application-Layer Solutions

   [RFC8085] recognizes that IP fragmentation reduces the reliability of
   Internet communication.  It also recognizes that UDP lacks a
   fragmentation mechanism of its own and relies on IP fragmentation.
   Therefore, [RFC8085] offers the following advice regarding
   applications the run over the UDP:

   |  An application SHOULD NOT send UDP datagrams that result in IP
   |  packets that exceed the Maximum Transmission Unit (MTU) along the
   |  path to the destination.  Consequently, an application SHOULD
   |  either use the path MTU information provided by the IP layer or
   |  implement Path MTU Discovery (PMTUD) itself [RFC1191] [RFC1981]
   |  [RFC4821] to determine whether the path to a destination will
   |  support its desired message size without fragmentation.

   RFC 8085 continues:

   |  Applications that do not follow the recommendation to do PMTU/
   |  PLPMTUD discovery SHOULD still avoid sending UDP datagrams that
   |  would result in IP packets that exceed the path MTU.  Because the
   |  actual path MTU is unknown, such applications SHOULD fall back to
   |  sending messages that are shorter than the default effective MTU
   |  for sending (EMTU_S in [RFC1122]).  For IPv4, EMTU_S is the
   |  smaller of 576 bytes and the first-hop MTU [RFC1122].  For IPv6,
   |  EMTU_S is 1280 bytes [RFC2460].  The effective PMTU for a directly
   |  connected destination (with no routers on the path) is the
   |  configured interface MTU, which could be less than the maximum
   |  link payload size.  Transmission of minimum-sized UDP datagrams is
   |  inefficient over paths that support a larger PMTU, which is a
   |  second reason to implement PMTU discovery.

   RFC 8085 assumes that for IPv4 an EMTU_S of 576 is sufficiently small
   to be supported by most current Internet paths, even though the IPv4
   minimum link MTU is 68 octets.

   This advice applies equally to any application that runs directly
   over IP.

5.  Applications That Rely on IPv6 Fragmentation

   The following applications rely on IPv6 fragmentation:

   *  DNS [RFC1035].

   *  OSPFv2 [RFC2328].

   *  OSPFv3 [RFC5340].

   *  Packet-in-packet encapsulations.

   Each of these applications relies on IPv6 fragmentation to a varying
   degree.  In some cases, that reliance is essential and cannot be
   broken without fundamentally changing the protocol.  In other cases,
   that reliance is incidental, and most implementations already take
   appropriate steps to avoid fragmentation.

   This list is not comprehensive, and other protocols that rely on IP
   fragmentation may exist.  They are not specifically considered in the
   context of this document.

5.1.  Domain Name Service (DNS)

   DNS relies on UDP for efficiency, and the consequence is the use of
   IP fragmentation for large responses, as permitted by the Extension
   Mechanisms for DNS (EDNS0) options in the query.  It is possible to
   mitigate the issue of fragmentation-based packet loss by having
   queries use smaller EDNS0 UDP buffer sizes or by having the DNS
   server limit the size of its UDP responses to some self-imposed
   maximum packet size that may be less than the preferred EDNS0 UDP
   buffer size.  In both cases, large responses are truncated in the
   DNS, signaling to the client to re-query using TCP to obtain the
   complete response.  However, the operational issue of the partial
   level of support for DNS over TCP, particularly in the case where
   IPv6 transport is being used, becomes a limiting factor of the
   efficacy of this approach [Damas].

   Larger DNS responses can normally be avoided by aggressively pruning
   the Additional section of DNS responses.  One scenario where such
   pruning is ineffective is in the use of DNSSEC, where large key sizes
   act to increase the response size to certain DNS queries.  There is
   no effective response to this situation within the DNS other than
   using smaller cryptographic keys and adopting of DNSSEC
   administrative practices that attempt to keep DNS response as short
   as possible.

5.2.  Open Shortest Path First (OSPF)

   OSPF implementations can emit messages large enough to cause
   fragmentation.  However, in order to optimize performance, most OSPF
   implementations restrict their maximum message size to a value that
   will not cause fragmentation.

5.3.  Packet-in-Packet Encapsulations

   This document acknowledges that in some cases, packets must be
   fragmented within IP-in-IP tunnels.  Therefore, this document makes
   no additional recommendations regarding IP-in-IP tunnels.

   In this document, packet-in-packet encapsulations include IP-in-IP
   [RFC2003], Generic Routing Encapsulation (GRE) [RFC2784], GRE-in-UDP
   [RFC8086], and Generic Packet Tunneling in IPv6 [RFC2473].  [RFC4459]
   describes fragmentation issues associated with all of the above-
   mentioned encapsulations.

   The fragmentation strategy described for GRE in [RFC7588] has been
   deployed for all of the above-mentioned encapsulations.  This
   strategy does not rely on IP fragmentation except in one corner case.
   (See Section of [RFC7588] and Section 7.1 of [RFC2473].)
   Section 3.3 of [RFC7676] further describes this corner case.

   See [TUNNELS] for further discussion.

5.4.  UDP Applications Enhancing Performance

   Some UDP applications rely on IP fragmentation to achieve acceptable
   levels of performance.  These applications use UDP datagram sizes
   that are larger than the Path MTU so that more data can be conveyed
   between the application and the kernel in a single system call.

   To pick one example, the Licklider Transmission Protocol (LTP)
   [RFC5326], which is in current use on the International Space Station
   (ISS), uses UDP datagram sizes larger than the Path MTU to achieve
   acceptable levels of performance even though this invokes IP
   fragmentation.  More generally, SNMP and video applications may
   transmit an application-layer quantum of data, depending on the
   network layer to fragment and reassemble as needed.

6.  Recommendations

6.1.  For Application and Protocol Developers

   Developers SHOULD NOT develop new protocols or applications that rely
   on IP fragmentation.  When a new protocol or application is deployed
   in an environment that does not fully support IP fragmentation, it
   SHOULD operate correctly, either in its default configuration or in a
   specified alternative configuration.

   While there may be controlled environments where IP fragmentation
   works reliably, this is a deployment issue and can not be known to
   someone developing a new protocol or application.  It is not
   recommended that new protocols or applications be developed that rely
   on IP fragmentation.  Protocols and applications that rely on IP
   fragmentation will work less reliably on the Internet.

   Legacy protocols that depend upon IP fragmentation SHOULD be updated
   to break that dependency.  However, in some cases, there may be no
   viable alternative to IP fragmentation (e.g., IPSEC tunnel mode, IP-
   in-IP encapsulation).  Applications and protocols cannot necessarily
   know or control whether they use lower layers or network paths that
   rely on such fragmentation.  In these cases, the protocol will
   continue to rely on IP fragmentation but should only be used in
   environments where IP fragmentation is known to be supported.

   Protocols may be able to avoid IP fragmentation by using a
   sufficiently small MTU (e.g., The protocol minimum link MTU),
   disabling IP fragmentation, and ensuring that the transport protocol
   in use adapts its segment size to the MTU.  Other protocols may
   deploy a sufficiently reliable PMTU discovery mechanism (e.g.,

   UDP applications SHOULD abide by the recommendations stated in
   Section 3.2 of [RFC8085].

6.2.  For System Developers

   Software libraries SHOULD include provision for PLPMTUD for each
   supported transport protocol.

6.3.  For Middlebox Developers

   Middleboxes, which are systems that "transparently" perform policy
   functions on passing traffic but do not participate in the routing
   system, should process IP fragments in a manner that is consistent
   with [RFC0791] and [RFC8200].  In many cases, middleboxes must
   maintain state in order to achieve this goal.

   Price and performance considerations frequently motivate network
   operators to deploy stateless middleboxes.  These stateless
   middleboxes may perform suboptimally, process IP fragments in a
   manner that is not compliant with RFC 791 or RFC 8200, or even
   discard IP fragments completely.  Such behaviors are NOT RECOMMENDED.
   If a middlebox implements nonstandard behavior with respect to IP
   fragmentation, then that behavior MUST be clearly documented.

6.4.  For ECMP, LAG, and Load-Balancer Developers And Operators

   In their default configuration, when the IPv6 Flow Label is not equal
   to zero, IPv6 devices that implement Equal-Cost Multipath (ECMP)
   Routing as described in OSPF [RFC2328] and other routing protocols,
   Link Aggregation Grouping (LAG) [RFC7424], or other load-distribution
   technologies SHOULD accept only the following fields as input to
   their hash algorithm:

   *  IP Source Address.

   *  IP Destination Address.

   *  Flow Label.

   Operators SHOULD deploy these devices in their default configuration.

   These recommendations are similar to those presented in [RFC6438] and
   [RFC7098].  They differ in that they specify a default configuration.

6.5.  For Network Operators

   Operators MUST ensure proper PMTUD operation in their network,
   including making sure the network generates PTB packets when dropping
   packets too large compared to outgoing interface MTU.  However,
   implementations MAY rate limit the generation of ICMP messages per
   [RFC1812] and [RFC4443].

   As per RFC 4890, network operators MUST NOT filter ICMPv6 PTB
   messages unless they are known to be forged or otherwise
   illegitimate.  As stated in Section 3.8, filtering ICMPv6 PTB packets
   causes PMTUD to fail.  Many upper-layer protocols rely on PMTUD.

   As per RFC 8200, network operators MUST NOT deploy IPv6 links whose
   MTU is less than 1280 octets.

   Network operators SHOULD NOT filter IP fragments if they are known to
   have originated at a domain name server or be destined for a domain
   name server.  This is because domain name services are critical to
   operation of the Internet.

7.  IANA Considerations

   This document has no IANA actions.

8.  Security Considerations

   This document mitigates some of the security considerations
   associated with IP fragmentation by discouraging its use.  It does
   not introduce any new security vulnerabilities, because it does not
   introduce any new alternatives to IP fragmentation.  Instead, it
   recommends well-understood alternatives.

9.  References

9.1.  Normative References

   [RFC0768]  Postel, J., "User Datagram Protocol", STD 6, RFC 768,
              DOI 10.17487/RFC0768, August 1980,

   [RFC0791]  Postel, J., "Internet Protocol", STD 5, RFC 791,
              DOI 10.17487/RFC0791, September 1981,

   [RFC0792]  Postel, J., "Internet Control Message Protocol", STD 5,
              RFC 792, DOI 10.17487/RFC0792, September 1981,

   [RFC0793]  Postel, J., "Transmission Control Protocol", STD 7,
              RFC 793, DOI 10.17487/RFC0793, September 1981,

   [RFC1035]  Mockapetris, P., "Domain names - implementation and
              specification", STD 13, RFC 1035, DOI 10.17487/RFC1035,
              November 1987, <https://www.rfc-editor.org/info/rfc1035>.

   [RFC1191]  Mogul, J. and S. Deering, "Path MTU discovery", RFC 1191,
              DOI 10.17487/RFC1191, November 1990,

   [RFC2119]  Bradner, S., "Key words for use in RFCs to Indicate
              Requirement Levels", BCP 14, RFC 2119,
              DOI 10.17487/RFC2119, March 1997,

   [RFC4443]  Conta, A., Deering, S., and M. Gupta, Ed., "Internet
              Control Message Protocol (ICMPv6) for the Internet
              Protocol Version 6 (IPv6) Specification", STD 89,
              RFC 4443, DOI 10.17487/RFC4443, March 2006,

   [RFC4821]  Mathis, M. and J. Heffner, "Packetization Layer Path MTU
              Discovery", RFC 4821, DOI 10.17487/RFC4821, March 2007,

   [RFC6437]  Amante, S., Carpenter, B., Jiang, S., and J. Rajahalme,
              "IPv6 Flow Label Specification", RFC 6437,
              DOI 10.17487/RFC6437, November 2011,

   [RFC6438]  Carpenter, B. and S. Amante, "Using the IPv6 Flow Label
              for Equal Cost Multipath Routing and Link Aggregation in
              Tunnels", RFC 6438, DOI 10.17487/RFC6438, November 2011,

   [RFC8085]  Eggert, L., Fairhurst, G., and G. Shepherd, "UDP Usage
              Guidelines", BCP 145, RFC 8085, DOI 10.17487/RFC8085,
              March 2017, <https://www.rfc-editor.org/info/rfc8085>.

   [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>.

   [RFC8200]  Deering, S. and R. Hinden, "Internet Protocol, Version 6
              (IPv6) Specification", STD 86, RFC 8200,
              DOI 10.17487/RFC8200, July 2017,

   [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,

   [RFC8899]  Fairhurst, G., Jones, T., Tüxen, M., Rüngeler, I., and T.
              Völker, "Packetization Layer Path MTU Discovery for
              Datagram Transports", RFC 8899, DOI 10.17487/RFC8899,
              September 2020, <https://www.rfc-editor.org/info/rfc8899>.

9.2.  Informative References

   [Damas]    Damas, J. and G. Huston, "Measuring ATR", April 2018,

   [Huston]   Huston, G., "IPv6, Large UDP Packets and the DNS", August

   [Kent]     Kent, C. and J. Mogul, "Fragmentation Considered Harmful",
              SIGCOMM '87: Proceedings of the ACM workshop on Frontiers
              in computer communications technology,
              DOI 10.1145/55482.55524, August 1987,

              Ptacek, T. H. and T. N. Newsham, "Insertion, Evasion and
              Denial of Service: Eluding Network Intrusion Detection",

   [RFC1122]  Braden, R., Ed., "Requirements for Internet Hosts -
              Communication Layers", STD 3, RFC 1122,
              DOI 10.17487/RFC1122, October 1989,

   [RFC1812]  Baker, F., Ed., "Requirements for IP Version 4 Routers",
              RFC 1812, DOI 10.17487/RFC1812, June 1995,

   [RFC1858]  Ziemba, G., Reed, D., and P. Traina, "Security
              Considerations for IP Fragment Filtering", RFC 1858,
              DOI 10.17487/RFC1858, October 1995,

   [RFC1981]  McCann, J., Deering, S., and J. Mogul, "Path MTU Discovery
              for IP version 6", RFC 1981, DOI 10.17487/RFC1981, August
              1996, <https://www.rfc-editor.org/info/rfc1981>.

   [RFC2003]  Perkins, C., "IP Encapsulation within IP", RFC 2003,
              DOI 10.17487/RFC2003, October 1996,

   [RFC2328]  Moy, J., "OSPF Version 2", STD 54, RFC 2328,
              DOI 10.17487/RFC2328, April 1998,

   [RFC2460]  Deering, S. and R. Hinden, "Internet Protocol, Version 6
              (IPv6) Specification", RFC 2460, DOI 10.17487/RFC2460,
              December 1998, <https://www.rfc-editor.org/info/rfc2460>.

   [RFC2473]  Conta, A. and S. Deering, "Generic Packet Tunneling in
              IPv6 Specification", RFC 2473, DOI 10.17487/RFC2473,
              December 1998, <https://www.rfc-editor.org/info/rfc2473>.

   [RFC2784]  Farinacci, D., Li, T., Hanks, S., Meyer, D., and P.
              Traina, "Generic Routing Encapsulation (GRE)", RFC 2784,
              DOI 10.17487/RFC2784, March 2000,

   [RFC3128]  Miller, I., "Protection Against a Variant of the Tiny
              Fragment Attack (RFC 1858)", RFC 3128,
              DOI 10.17487/RFC3128, June 2001,

   [RFC4340]  Kohler, E., Handley, M., and S. Floyd, "Datagram
              Congestion Control Protocol (DCCP)", RFC 4340,
              DOI 10.17487/RFC4340, March 2006,

   [RFC4459]  Savola, P., "MTU and Fragmentation Issues with In-the-
              Network Tunneling", RFC 4459, DOI 10.17487/RFC4459, April
              2006, <https://www.rfc-editor.org/info/rfc4459>.

   [RFC4890]  Davies, E. and J. Mohacsi, "Recommendations for Filtering
              ICMPv6 Messages in Firewalls", RFC 4890,
              DOI 10.17487/RFC4890, May 2007,

   [RFC4960]  Stewart, R., Ed., "Stream Control Transmission Protocol",
              RFC 4960, DOI 10.17487/RFC4960, September 2007,

   [RFC4963]  Heffner, J., Mathis, M., and B. Chandler, "IPv4 Reassembly
              Errors at High Data Rates", RFC 4963,
              DOI 10.17487/RFC4963, July 2007,

   [RFC5326]  Ramadas, M., Burleigh, S., and S. Farrell, "Licklider
              Transmission Protocol - Specification", RFC 5326,
              DOI 10.17487/RFC5326, September 2008,

   [RFC5340]  Coltun, R., Ferguson, D., Moy, J., and A. Lindem, "OSPF
              for IPv6", RFC 5340, DOI 10.17487/RFC5340, July 2008,

   [RFC5722]  Krishnan, S., "Handling of Overlapping IPv6 Fragments",
              RFC 5722, DOI 10.17487/RFC5722, December 2009,

   [RFC5927]  Gont, F., "ICMP Attacks against TCP", RFC 5927,
              DOI 10.17487/RFC5927, July 2010,

   [RFC6346]  Bush, R., Ed., "The Address plus Port (A+P) Approach to
              the IPv4 Address Shortage", RFC 6346,
              DOI 10.17487/RFC6346, August 2011,

   [RFC6888]  Perreault, S., Ed., Yamagata, I., Miyakawa, S., Nakagawa,
              A., and H. Ashida, "Common Requirements for Carrier-Grade
              NATs (CGNs)", BCP 127, RFC 6888, DOI 10.17487/RFC6888,
              April 2013, <https://www.rfc-editor.org/info/rfc6888>.

   [RFC7098]  Carpenter, B., Jiang, S., and W. Tarreau, "Using the IPv6
              Flow Label for Load Balancing in Server Farms", RFC 7098,
              DOI 10.17487/RFC7098, January 2014,

   [RFC7424]  Krishnan, R., Yong, L., Ghanwani, A., So, N., and B.
              Khasnabish, "Mechanisms for Optimizing Link Aggregation
              Group (LAG) and Equal-Cost Multipath (ECMP) Component Link
              Utilization in Networks", RFC 7424, DOI 10.17487/RFC7424,
              January 2015, <https://www.rfc-editor.org/info/rfc7424>.

   [RFC7588]  Bonica, R., Pignataro, C., and J. Touch, "A Widely
              Deployed Solution to the Generic Routing Encapsulation
              (GRE) Fragmentation Problem", RFC 7588,
              DOI 10.17487/RFC7588, July 2015,

   [RFC7676]  Pignataro, C., Bonica, R., and S. Krishnan, "IPv6 Support
              for Generic Routing Encapsulation (GRE)", RFC 7676,
              DOI 10.17487/RFC7676, October 2015,

   [RFC7739]  Gont, F., "Security Implications of Predictable Fragment
              Identification Values", RFC 7739, DOI 10.17487/RFC7739,
              February 2016, <https://www.rfc-editor.org/info/rfc7739>.

   [RFC7872]  Gont, F., Linkova, J., Chown, T., and W. Liu,
              "Observations on the Dropping of Packets with IPv6
              Extension Headers in the Real World", RFC 7872,
              DOI 10.17487/RFC7872, June 2016,

   [RFC8086]  Yong, L., Ed., Crabbe, E., Xu, X., and T. Herbert, "GRE-
              in-UDP Encapsulation", RFC 8086, DOI 10.17487/RFC8086,
              March 2017, <https://www.rfc-editor.org/info/rfc8086>.

   [TUNNELS]  Touch, J. and M. Townsley, "IP Tunnels in the Internet
              Architecture", Work in Progress, Internet-Draft, draft-
              ietf-intarea-tunnels-10, 12 September 2019,

              Touch, J., "Transport Options for UDP", Work in Progress,
              Internet-Draft, draft-ietf-tsvwg-udp-options-08, 12
              September 2019, <https://tools.ietf.org/html/draft-ietf-


   Thanks to Mikael Abrahamsson, Brian Carpenter, Silambu Chelvan,
   Lorenzo Colitti, Gorry Fairhurst, Joel Halpern, Mike Heard, Tom
   Herbert, Tatuya Jinmei, Suresh Krishnan, Jen Linkova, Paolo Lucente,
   Manoj Nayak, Eric Nygren, Fred Templin, and Joe Touch for their

Authors' Addresses

   Ron Bonica
   Juniper Networks
   2251 Corporate Park Drive
   Herndon, Virginia 20171
   United States of America

   Email: rbonica@juniper.net

   Fred Baker
   Santa Barbara, California 93117
   United States of America

   Email: FredBaker.IETF@gmail.com

   Geoff Huston
   6 Cordelia St
   Brisbane 4101 QLD

   Email: gih@apnic.net

   Robert M. Hinden
   Check Point Software
   959 Skyway Road
   San Carlos, California 94070
   United States of America

   Email: bob.hinden@gmail.com

   Ole Troan
   Philip Pedersens vei 1
   N-1366 Lysaker

   Email: ot@cisco.com

   Fernando Gont
   SI6 Networks
   Evaristo Carriego 2644
   Provincia de Buenos Aires

   Email: fgont@si6networks.com