One document matched: draft-mrw-nat66-02.xml
<?xml version="1.0" encoding="US-ASCII"?>
<!DOCTYPE rfc SYSTEM "rfc2629.dtd">
<?xml-stylesheet type='text/xsl' href='rfc2629.xslt' ?>
<?rfc strict="yes" ?>
<?rfc comments="no" ?>
<?rfc inline="no" ?>
<?rfc editing="no" ?>
<?rfc toc="yes"?>
<?rfc tocompact="yes"?>
<?rfc tocdepth="3"?>
<?rfc symrefs="yes"?>
<?rfc sortrefs="yes" ?>
<?rfc compact="yes" ?>
<?rfc subcompact="no" ?>
<rfc category="std" docName="draft-mrw-nat66-02" ipr="trust200902">
<front>
<title abbrev="NPTv6">IPv6-to-IPv6 Network Prefix Translation</title>
<author fullname="Margaret Wasserman" initials="M." surname="Wasserman">
<organization>Painless Security</organization>
<address>
<postal>
<street></street>
<city>North Andover</city>
<region>MA</region>
<code>01845</code>
<country>USA</country>
</postal>
<phone>+1 781 405 7464</phone>
<email>mrw@painless-security.com</email>
<uri>http://www.painless-secuirty.com</uri>
</address>
</author>
<author fullname="Fred Baker" initials="F.J." surname="Baker">
<organization>Cisco Systems</organization>
<address>
<postal>
<street></street>
<city>Santa Barbara</city>
<code>93117</code>
<region>California</region>
<country>USA</country>
</postal>
<phone>+1-408-526-4257</phone>
<email>fred@cisco.com</email>
</address>
</author>
<date year="2010" />
<area>Transport</area>
<workgroup></workgroup>
<abstract>
<t>This document describes a stateless, transport-agnostic IPv6-to-IPv6
Network Prefix Translation (NPTv6) function that provides the address
independence benefit associated with IPv4-to-IPv4 NAT (NAT44), and in
addition provides a 1:1 relationship between addresses in the "inside"
and "outside" prefixes, preserving end to end reachability at the
network layer.</t>
</abstract>
<note title="Requirements Terminology">
<t>The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
"SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this
document are to be interpreted as described in <xref
target="RFC2119">RFC 2119</xref>.</t>
</note>
</front>
<middle>
<section title="Introduction">
<t>This document describes a stateless IPv6-to-IPv6 Network Prefix
Translation (NPTv6) function, designed to provide address independence
to the edge network. It is transport-agnostic with respect to transports
that don't checksum the IP header, such as SCTP or DCCP, and to
transports that use the TCP/UDP pseudo-header and <xref
target="RFC1071">checksum</xref>.</t>
<t>This has several ramifications: <list style="symbols">
<t>Any security benefit that NAT44 might offer is not present in
NPTv6, necessitating the use of a firewall to obtain those benefits
if desired. An example of such a firewall is described in <xref
target="I-D.ietf-v6ops-cpe-simple-security"></xref>.</t>
<t>End to end reachability is preserved, although the address used
"inside" the edge network differs from the address used "outside"
the edge network. This has implications for application referrals
and other uses of Internet layer addresses.</t>
<t>If there are multiple identically-configured prefix translators
between two networks, there is no need for them to exchange dynamic
state, as there is no dynamic state - the algorithmic translation
will be identical across each of them. The network can therefore
asymmetrically route, load-share, and fail-over among them without
issue.</t>
<t>Since translation is 1:1 at the network layer, there is no need
to modify port numbers or other transport parameters.</t>
</list></t>
<section title="What is Address Independence?">
<t>For the purposes of this document, IPv6 Address Independence
consists of the following set of properties: <list style="hanging">
<t hangText="From the perspective of the edge network:"><list
style="symbols">
<t>The IPv6 addresses used inside the local network (for
interfaces, access lists, and logs) do not need to be
renumbered if the upstream network changes a site's external
prefix.</t>
<t>The IPv6 addresses used inside the edge network (for
interfaces, access lists, and logs) or within other upstream
networks (such as when multihoming) do not need to be
renumbered when a site adds, drops, or changes upstream
networks.</t>
<t>It is not necessary for an administration to convince an
upstream network to route its internal IPv6 prefixes, or for
it to advertise prefixes derived from other upstream networks
into it.</t>
<t>Unless it wants to optimize routing between multiple
upstream networks in the process of multihoming, there is
therefore no need for a BGP exchange with the upstream
network.</t>
</list></t>
<t hangText="From the perspective of the upstream network:"><list
style="symbols">
<t>IPv6 addresses used by the edge network are guaranteed to
have a provider-allocated prefix, eliminating the need and
concern for <xref target="RFC2827">BCP 38</xref> ingress
filtering and the advertisement of customer-specific
prefixes.</t>
</list></t>
</list></t>
<t>Thus, address independence has ramifications for the edge network,
networks it directly connects with (especially its upstream networks),
and for the Internet as a whole. The desire for address independence
has been a primary driver for IPv4 NAT deployment in medium to
large-sized enterprise networks, including NAT deployments in
enterprises that have plenty of IPv4 provider-independent address
space (from IPv4 "swamp space"). It has also been a driver for edge
networks to become members of RIR communities, seeking to obtain BGP
Autonomous System Numbers and provider-independent prefixes, and as a
result has been one of the drivers of the explosion of the IPv4 route
table. Service providers have stated that the lack of address
independence from their customers has been a negative incentive to
deployment, due to the impact of customer routing expected in their
networks.</t>
<t>The <xref target="RFC4864">Local Network Protection</xref> document
discusses a related concept called "Address Autonomy" as a benefit of
NAT44. <xref target="RFC4864"></xref> indicates that address autonomy
can be achieved by the simultaneous use of global addresses on all
nodes within a site that need external connectivity, and Unique Local
Addresses (ULAs) <xref target="RFC4193"></xref> for all internal
communication. However, this solution fails to meet the requirement
for address independence, because if an ISP renumbering event occurs,
all of the hosts, routers, DHCP servers, ACLs, firewalls and other
internal systems that are configured with global addresses from the
ISP will need to be renumbered before global connectivity is fully
restored.</t>
<t>The use of IPv6 Provider Independent (PI) addresses has also been
suggested as a means to fulfill the address independence requirement.
However, this solution requires that an enterprise qualify to receive
a PI assignment and persuade their ISP to install specific routes for
the enterprise's PI addresses. There are a number of practical issues
with this approach, especially if there is a desire to route to a
number of geographically and topologically diverse set of sites, which
can sometimes involve coordinating with several ISPs to route portions
of a single PI prefix. These problems have caused numerous enterprises
with plenty of IPv4 swamp space to choose to use IPv4 NAT for part, or
substantially all, of their internal network instead of using their
provider-independent address space.</t>
</section>
<section title="NPTv6 Applicability">
<t>NPTv6 provides a simple and compelling solution to meet the Address
Independence requirement in IPv6. The address independence benefit
stems directly from the translation function of the network prefix
translator. To avoid as many of the issues associated with NAT44 as
possible, NPTv6 is defined to include a two-way, checksum-neutral,
algorithmic translation function, and nothing else.</t>
<t>NPTv6 does not include a port mapping function, and the defined
address mapping mechanism is checksum-neutral. This avoids the need
for a NPTv6 Translator to re-write transport layer headers, making it
feasible to deploy new or improved transport layer protocols without
upgrading NPTv6 Translators. Because NPTv6 does not involve re-writing
transport-layer headers, NPTv6 will not interfere with encryption of
the full IP payload in many cases.</t>
<t>The default NPTv6 address mapping mechanism is purely algorithmic,
so NPTv6 translators do not need to maintain per-node or
per-connection state, allowing deployment of more robust and adaptive
networks than can be deployed using NAT44. Since the default NPTv6
mapping can be performed in either direction, it does not interfere
with inbound connection establishment, thus allowing internal nodes to
participate in direct Peer-to-Peer applications without the
application layer overhead one finds in many IPv4 Peer-to-Peer
applications.</t>
<t>Although NPTv6 compares favorably to NAT44 in several ways, it does
not eliminate all of the architectural problems associated with IPv4
NAT, as described in <xref target="RFC2993"></xref>. NPTv6 involves
modifying IP headers in transit, so it is not compatible with security
mechanisms, such as the IPsec Authentication Header, that provide
integrity protection for the IP header. NPTv6 may interfere with the
use of application protocols that transmit IP addresses in the
application-specific portion of the IP packet. These applications
currently require application layer gateways (ALGs) to work correctly
through NAT44 devices, and similar ALGs may be required for these
applications to work through NPTv6 Translators. The use of separate
internal and external prefixes creates complexity for DNS deployment,
due the desire for internal nodes to communicate with other internal
nodes using internal addresses, while external nodes need to obtain
external addresses to communicate with the same nodes. This frequently
results in the deployment of "split DNS", which may add complexity to
network configuration.</t>
<t>There are significant technical impacts associated with the
deployment of any prefix translation mechanism, including NPTv6, and
we strongly encourage anyone who is considering the implementation or
deployment of NPTv6 to read <xref target="RFC4864"></xref>, and to
carefully consider the alternatives described in that document, some
of which may cause fewer problems than NPTv6.</t>
</section>
</section>
<section title="NPTv6 Overview">
<t>NPTv6 may be implemented in an IPv6 router to map one IPv6 address
prefix to another IPv6 prefix as each IPv6 packet transits the router. A
router that implements a NPTv6 prefix translation function is referred
to as an NPTv6 Translator.</t>
<section title="NPTv6: the simplest case">
<t>In its simplest form, a NPTv6 Translator interconnects two network
links, one of which is an "internal" network link attached to a leaf
network within a single administrative domain, and the other of which
is an "external" network with connectivity to the global Internet. All
of the hosts on the internal network will use addresses from a single,
locally-routed prefix, and those addresses will be translated to/from
addresses in a globally-routable prefix as IP packets transit the
NPTv6 Translator. The lengths of these two prefixes will be
functionally the same; if they differ, the longer of the two will
limit the ability to use subnets in the shorter.</t>
<figure anchor="simpleNPTv6" title="A simple translator">
<artwork align="center"><![CDATA[
External Network: Prefix = 2001:0DB8:0001:/48
--------------------------------------
|
|
+-------------+
| NPTv6 |
| Translator |
+-------------+
|
|
--------------------------------------
Internal Network: Prefix = FD01:0203:0405:/48
]]></artwork>
</figure>
<t><xref target="simpleNPTv6"></xref> shows a NPTv6 Translator
attached to two networks. In this example, the internal network uses
<xref target="RFC4193">IPv6 Unique Local Addresses (ULAs)</xref> to
represent the internal IPv6 nodes, and the external network uses
globally routable IPv6 addresses to represent the same nodes.</t>
<t>When a NPTv6 Translator forwards packets in the "outbound"
direction, from the internal network to the external network, NPTv6
overwrites the IPv6 source prefix (in the IPv6 header) with a
corresponding external prefix. When packets are forwarded in the
"inbound" direction, from the external network to the internal
network, the IPv6 destination prefix is overwritten with a
corresponding prefix internal prefix. Using the prefixes shown in the
diagram above, as an IP packet passes through the NPTv6 Translator in
the outbound direction, the source prefix (FD01:0203:0405:/48) will be
overwritten with the external prefix (2001:0DB8:0001:/48). In an
inbound packet, the destination prefix (2001:0DB8:0001:/48) will be
overwritten with the internal prefix (FD01:0203:0405:/48). In both
cases, it is the local IPv6 prefix that is overwritten; the remote
IPv6 prefix remains unchanged. Nodes on the internal network are said
to be "behind" the NPTv6 Translator.</t>
</section>
<section title="NPTv6 between peer networks">
<t>NPTv6 can also be used between two private networks. In these
cases, both networks may use ULA prefixes, with each subnet in one
network mapped into a corresponding subnet in the other network, and
vice versa. Or, each network may use ULA prefixes for internal
addressing, and global unicast addresses on the other network.</t>
<figure anchor="flow" title="Flow of Information in Translation">
<artwork align="center"><![CDATA[
Internal Prefix = FD01:4444:5555:/48
--------------------------------------
V | External Prefix
V | 2001:0DB8:6666:/48
V +---------+ ^
V | NPTv6 | ^
V | Device | ^
V +---------+ ^
External Prefix | ^
2001:0DB8:0001:/48 | ^
--------------------------------------
Internal Prefix = FD01:0203:0405:/48
]]></artwork>
</figure>
</section>
<section title="NPTv6 redundnacy and load-sharing">
<t>In some cases, more than one NPTv6 Translator may be attached to a
network, as show in <xref target="parallel"></xref>. In such cases,
NPTv6 Translators are configured with the same internal and external
prefixes. Since there is only one translation, even though there are
multiple translators, they map only one external address (prefix and
IID) to the internal address.</t>
<figure anchor="parallel" title="Parallel Translators">
<artwork align="center"><![CDATA[
External Network: Prefix = 2001:0DB8:0001:/48
--------------------------------------
| |
| |
+-------------+ +-------------+
| NPTv6 | | NPTv6 |
| Translator | | Translator |
| #1 | | #2 |
+-------------+ +-------------+
| |
| |
--------------------------------------
Internal Network: Prefix = FD01:0203:0405:/48
]]></artwork>
</figure>
</section>
<section title="NPTv6 multihoming">
<figure anchor="distinct"
title="Parallel Translators with different upstream networks">
<artwork align="center"><![CDATA[
External Network #1: External Network #2:
Prefix = 2001:0DB8:0001:/48 Prefix = 2001:0DB8:5555:/48
--------------------------- --------------------------
| |
| |
+-------------+ +-------------+
| NPTv6 | | NPTv6 |
| Translator | | Translator |
| #1 | | #2 |
+-------------+ +-------------+
| |
| |
--------------------------------------
Internal Network: Prefix = FD01:0203:0405:/48
]]></artwork>
</figure>
<t>When multihoming, NPTv6 Translators are attached to an internal
network, as show in <xref target="distinct"></xref>, but connected to
different external networks. In such cases, NPTv6 Translators are
configured with the same internal prefix, but different external
prefixes. Since there are multiple translations, they map multiple
external addresses (prefix and IID) to the common internal address. A
system within the edge network is unable to determine which external
address it is using.</t>
<t>Multihoming in this sense has one negative feature as compared with
multihoming with a provider-independent address; when routes change
between NPTv6 Translators, since the upstream network changes, the
prefix used in shifting sessions changes. This obviously causes them
to fail. This is not expected to be a major real issue, however, in
networks where routing is generally stable.</t>
</section>
<section title="Mapping with No Per-Flow State">
<t>When NPTv6 is used as described in this document, no per-node or
per-flow state is maintained in the NPTv6 Translator. Both inbound and
outbound packets are translated algorithmically, using only
information found in the IPv6 header. Due to this property, NPTv6's
two-way, algorithmic address mapping can support both outbound and
inbound connection establishment without the need for state-priming or
rendezvous mechanisms, or the maintenance of mapping state. This is a
significant improvement over NAT44 devices, but it also has
significant security implications which are described in the Security
Considerations section.</t>
</section>
<section title="Checksum-Neutral Mapping">
<t>When a change is made to one of the IP header fields in the IPv6
pseudo-header checksum (such as one of the IP addresses), the checksum
field in the transport layer header may become invalid. Fortunately,
an incremental change in the area covered by the Internet standard
checksum <xref target="RFC1071"></xref> will result in a well-defined
change to the checksum value <xref target="RFC1624"></xref>. So, a
checksum change caused by modifying part of the area covered by the
checksum can be corrected by making a complementary change to a
different 16-bit field covered by the same checksum.</t>
<t>The NPTv6 mapping mechanisms described in this document are
checksum-neutral, which means that they result in IP headers that will
generate the same IPv6 pseudo-header checksum when the checksum is
calculated using the standard Internet checksum algorithm <xref
target="RFC1071"></xref>. Any changes that are made during translation
of the IPv6 prefix are offset by changes to other parts of the IPv6
address. This results in transport layers that use the Internet
checksum (such as TCP and UDP) calculating the same IPv6 pseudo header
checksum for both the internal and external forms of the same packet,
which avoids the need for the NPTv6 Translator to modify those
transport layer headers to correct the checksum value.</t>
</section>
</section>
<section title="NPTv6 Algorithmic Specification">
<t>The <xref target="RFC4291"></xref> IPv6 Address is reproduced for
clarity in <xref target="address"></xref>.</t>
<figure anchor="address"
title="Enumeration of the IPv6 Address [RFC4291]">
<artwork align="center"><![CDATA[
0 15 16 31 32 47 48 63 64 79 80 95 96 111 112 127
+-------+-------+-------+-------+-------+-------+-------+-------+
| Routing Prefix |Subnet| Interface Identifier (IID) |
+-------+-------+-------+-------+-------+-------+-------+-------+
]]></artwork>
</figure>
<section anchor="configuration" title="NPTv6 configuration calculations">
<t>When an NPTv6 Translation function is configured, it is configured
with <list style="symbols">
<t>one or more "internal" interfaces with their "internal" routing
domain prefixes, and</t>
<t>one or more "external" interfaces with their "external" routing
domain prefixes.</t>
</list></t>
<t>In the simple case, there is one of each. If a single router
provides NPTv6 translation services between a multiplicity of domains
(as might be true when multihoming), each internal/external pair must
be thought of as a separate NPTv6 Translator from the perspective of
this specification.</t>
<t>When an NPTv6 Translator is configured, the translation function
first ensures that the internal and external prefixes are the same
length, if necessary by extending the shorter of the two with zeroes.
These two prefixes will be used in the prefix translation function
described in <xref target="leaving"></xref> and <xref
target="returning"></xref>.</t>
<t>They are then zero-extended to /64, for the purposes of a
calculation. The translation function calculates the ones-complement
sum of the 16 bit words of the /64 external prefix and the /64
internal prefix. It then calculates the difference between these
values: external minus internal. This value, called the "adjustment",
is effectively constant for the lifetime of the NPTv6 Translator
configuration, and used in per-packet processing.</t>
</section>
<section anchor="leaving"
title="NPTv6 translation, internal network to external network">
<t>When a datagram passes through the NPTv6 Translator from an
internal to an external network, its IPv6 Source Address is changed in
two ways: <list style="symbols">
<t>The internal prefix is overwritten with the external prefix,
and</t>
<t>A 16-bit word of the address has the "adjustment" added to it
using one's complement arithmetic. If the result is 0xFFFF, it is
overwritten as zero. The choice of word is as specified in <xref
target="slash48"></xref> or <xref target="slash56"></xref> as
appropriate.</t>
</list></t>
</section>
<section anchor="returning"
title="NPTv6 translation, external network to internal network">
<t>When a datagram passes through the NPTv6 Translator from an
internal to an external network, its IPv6 Destination Address is
changed in two ways: <list style="symbols">
<t>The external prefix is overwritten with the internal prefix,
and</t>
<t>A 16-bit word of the address has the "adjustment" subtracted
from it (bitwise inverted and added to it) it using one's
complement arithmetic. If the result is 0xFFFF, it is overwritten
as zero. The choice of word is as specified in <xref
target="slash48"></xref> or <xref target="slash56"></xref> as
appropriate.</t>
</list></t>
</section>
<section anchor="slash48" title="NPTv6 with a /48 or shorter prefix">
<t>When a NPTv6 Translator is configured with internal and external
prefixes that are 48 bits in length (a /48) or shorter, the adjustment
MUST be added to or subtracted from bits 48..63 of the address.</t>
<t>This mapping results in no modification of the Interface Identifier
(IID), which is held in the lower half of the IPv6 address, so it will
not interfere with future protocols that may use unique IIDs for node
identification.</t>
<t>NPTv6 Translator implementations MUST implement the /48
mapping.</t>
</section>
<section anchor="slash56" title="NPTv6 with a /49 or longer prefix">
<t>When a NPTv6 Translator is configured with internal and external
prefixes that are longer than 48 bits in length (such as a /52, /56,
or /60), the adjustment must be added to or subtracted from one of the
words in bits 64..79, 80..95, 96..111, or 112..127 of the address.
While the choice of word is immaterial as long as it is consistent,
for consistency's sake, these words MUST be inspected in that
sequence, and the first that is not initially 0xFFFF chosen.</t>
<t>NPTv6 Translator implementations SHOULD implement the mapping for
longer prefixes.</t>
</section>
<section anchor="example48" title="/48 Prefix Mapping Example">
<t>For the network shown in <xref target="simpleNPTv6"></xref>, the
Internal Prefix is FD01:0203:0405:/48, and the External Prefix is
2001:0DB8:0001:/48</t>
<t>If a node with internal address FD01:0203:0405:0001::1234 sends an
outbound packet through the NPTv6 Translator, the resulting external
address will be 2001:0DB8:0001:D550::1234. The resulting address is
obtained by calculating the checksum of both the internal and external
48-bit prefixes, subtracting the internal prefix from the external
prefix using one's complement arithmetic to calculate the
"adjustment", and adding the adjustment to the 16-bit subnet field (in
this case 0x0001).</t>
<t>To show the work:</t>
<t>The one's complement checksum of FD01:0203:0405 is 0xFCF5. The
one's complement checksum of 2001:0DB8:0001 is 0xD245. Using one's
complement math, 0xD245 - 0xFCF5 = 0xD54F. The subnet in the original
packet is 0x0001. Using one's complement math, 0x0001 + 0xD54F =
0xD550. Since 0xD550 != 0xFFFF, it is not changed to 0x0000.</t>
<t>So, the value 0xD550 is written in the 16-bit subnet area,
resulting in a mapped external address of
2001:0DB8:0001:D550::1234.</t>
<t>When a response packet is received, it will contain the destination
address 2001:0DB8:0001:D550::0001, which will be mapped using the
inverse mapping algorithm, back to FD01:0203:0405:0001::1234.</t>
<t>In this case, the difference between the two prefixes will be
calculated as follows:</t>
<t>Using one's complement math, 0xFCF5 - 0xD245 = 0x2AB0. The subnet
in the original packet = 0xD550. Using one's complement math, 0xD550 +
0x2AB0 = 0x0001. Since 0x0001 != 0xFFFF, it is not changed to
0x0000.</t>
<t>So the value 0x0001 is written into the subnet field, and the
internal value of the subnet field is properly restored.</t>
</section>
<section title="Address Mapping for Longer Prefixes">
<t>If the prefix being mapped is longer than 48 bits, the algorithm is
slightly more complex. A common case will be that the internal and
external prefixes are of different length. In such a case, the shorter
prefix is zero-extended to the length of the longer as described in
<xref target="configuration"></xref> for the purposes of overwriting
the prefix. Then, they are both zero-extended to 64 bits to facilitate
one's complement arithmetic. The "adjustment" is calculated using
those 64 bit prefixes.</t>
<t>For example if the internal prefix is a /48 ULA and the external
prefix is a /56 provider-allocated prefix, the ULA becomes a /56 with
zeros in bits 48..55. For purposes of one's complement arithmetic,
they are then both zero-extended to 64 bits. A side-effect of this is
that a subset of the subnets possible in the shorter prefix are
untranslatable. While the security value of this is debatable, the
administration may choose to use them for subnets that it knows need
no external accessibility.</t>
<t>We then find the first word in the IID that does not have the value
0xFFFF, trying bits 64..79, and then 80..95, 96..111, and finally
112..127. We perform the same calculation (with the same proof of
correctness) as in <xref target="example48"></xref>, but applying it
to that word.</t>
<t>Although any 16-bit portion of an IPv6 IID could contain 0xFFFF, an
IID of all-ones is a reserved anycast identifier that should not be
used on the network <xref target="RFC2526"></xref>. If a NPTv6
Translator discovers a packet with an IID of all-zeros while
performing address mapping, that packet MUST be dropped, and an ICMPv6
Parameter Problem error SHOULD be generated <xref
target="RFC4443"></xref>.</t>
<t>Note: this mechanism does involve modification of the IID; it may
not be compatible with future mechanisms that use unique IIDs for node
identification.</t>
</section>
</section>
<section title="Implications of Network Address Translator Behavioral Requirements">
<section title="Prefix Configuration and generation">
<t>NPTv6 Translators MUST support manual configuration of internal and
external prefixes, and MUST NOT place any restrictions on those
prefixes except that they be valid IPv6 unicast prefixes as described
in <xref target="RFC4291"></xref>.</t>
<t>NPTv6 Translators that do not have a manually configured internal
prefix SHOULD randomly generate a ULA prefix for the internal network
and advertise that prefix in router advertisements. NPTv6 Translators
with more than one internal interface SHOULD assign a (non-0xFFFF)
subnet number to each link, and include the subnet number in router
advertisements on the corresponding link. NPTv6 Translators that
generate a ULA prefix MUST generate the prefix using a random number
as described in RFC4291 <xref target="RFC4193"></xref>, and SHOULD
store the randomly generated prefix in non-volatile storage for
continued use.</t>
</section>
<section title="NAT Behavioral Requirements">
<t>NPTv6 Translators MUST support hairpinning behavior, as defined in
the NAT Behavioral Requirements for UDP document <xref
target="RFC4787"></xref>. This means that when a NPTv6 Translator
receives a packet on the internal interface that has a destination
address that matches the site's external prefix, it will translate the
packet and forward it internally. This allows internal nodes to reach
other internal nodes using their external, global addresses when
necessary.</t>
<t>Conceptually, the datagram leaves the domain (is translated as
described in <xref target="leaving"></xref>), and returns (is again
translated as described in <xref target="returning"></xref>). As a
result, the datagram exchange will be through the NPTv6 Translator in
both directions for the lifetime of the session. The alternative would
be to require the NPTv6 Translator to drop the datagram, forcing the
sender to use the correct internal prefix for its peer. Performing
only the external-to-internal translation results in the datagram
being sent from the untranslated internal address of the source to the
translated and therefore internal address of its peer, which would
enable the session to bypass the NPTv6 Translator for future
datagrams. It would also mean that the original sender would be
unlikely to recognize the response when it arrived.</t>
<t>Because NPTv6 does not perform port mapping and uses a one-to-one,
reversible mapping algorithm, none of the other NAT behavioral
requirements apply to NPTv6.</t>
</section>
</section>
<section title="Implications for Applications">
<t>The use of NPTv6 Transition technology makes a capability available
to Applications (and the networks that contain them) that is not readily
possible in a NAT44 network. This is the ability to position an
externally-accessible application within the "internal" network. In an
IPv4 network using NAT44, externally-accessible application must be
positioned on systems with global addresses, forcing the edge network to
obtain global address allocation; if the application can be in the
translated routing domain, it automatically has an address in each of
its upstream prefixes without the edge network obtaining such. However,
there must be a means for the application to know what addresses are
usable. Reasons include at least advertisement in DNS (which might be
done statically if DNS is directly maintained by the administration, or
from the end system if Dynamic DNS is in use). If referrals and other
uses of network layer addressing do not use names, then the application
needs a means to determine what addresses are relevant, whether from DNS
or another means.</t>
<t>The means of address discovery is not within the scope of this
specification.</t>
</section>
<section title="A Note on Port Mapping">
<t>In addition to overwriting IP addresses when packets are forwarded,
NAPT44 devices overwrite the source port number in outbound traffic, and
the destination port number in inbound traffic. This mechanism is called
"port mapping".</t>
<t>The major benefit of port mapping is that it allows multiple
computers to share a single IPv4 address. A large number of internal
IPv4 addresses (typically from the 10.0.0.0/8 prefix) can be mapped into
a single external, globally routable IPv4 address, with the local port
number used to identify which internal node should receive each inbound
packet. This address amplification feature should not be needed in
IPv6.</t>
<t>Since port mapping requires re-writing a portion of the transport
layer header, it requires NAPT44 devices to be aware of all of the
transport protocols that they forward, thus stifling the development of
new and improved transport protocols and preventing the use of IPsec
encryption. Modifying the transport layer header is incompatible with
security mechanisms that encrypt the full IP payload, and restricts the
NAPT44 to forwarding transport layers that use weak checksum algorithms
that are easily recalculated in routers.</t>
<t>Since there is significant detriment caused by modifying transport
layer headers and very little, if any, benefit to the use of port
mapping in IPv6, NPTv6 Translators that comply with this specification
MUST NOT perform port mapping.</t>
</section>
<section title="Security Considerations">
<t>When NPTv6 is deployed using either of the two-way, algorithmic
mappings defined in the document, it allows direct inbound connections
to internal nodes. While this can be viewed as a benefit of NPTv6 vs.
NAT44, it does open internal nodes to attacks that would be more
difficult in a NAT44 network. Although this situation is not
substantially worse, from a security standpoint, than running IPv6 with
no NAT, some enterprises may assume that a NPTv6 Translator will offer
similar protection to a NAT44 device. For this reason, it is RECOMMENDED
that NPTv6 Translators also implement firewall functionality such as
described in <xref target="I-D.ietf-v6ops-cpe-simple-security"></xref>,
with appropriate configuration options including turning it on or
off.</t>
</section>
<section title="IANA Considerations">
<t>This document has no IANA considerations.</t>
</section>
<section title="Acknowledgements">
<t>The checksum-neutral algorithmic address mapping described in this
document is based on e-mail written by Iljtsch Van Beijnum.</t>
<t>The following people provided advice or review comments that
substantially improved this document: Jari Arrko, Iljtsch Van Beijnum,
Remi Depres, Tony Hain, Ed Jankiewicz, Dave Thaler, Mark Townsley, and
Steve Blake.</t>
<t>This document was written using the xml2rfc tool described in RFC
2629 <xref target="RFC2629"></xref>.</t>
</section>
<section title="Change Log">
<section title="Changes Between draft-mrw-behave-nat66-00 and -01">
<t>There were several minor changes made between the *behave-nat66-00
and -01 versions of this draft: <list style="symbols">
<t>Added Fred Baker as a co-author.</t>
<t>Minor mathematical corrections.</t>
<t>Added AH to paragraph on NAT security issues.</t>
<t>Added additional NAT topologies to overview (diagrams TBD).</t>
</list></t>
</section>
<section title="Changes between *behave-nat66-01 and -02">
<t>There were further changes made between *behave-nat66-01 and -02:
<list style="symbols">
<t>Removed topology hiding mechanism.</t>
<t>Added diagrams.</t>
<t>Made minor updates based on mailing list feedback.</t>
<t>Added discussion of IPv6 SAF document.</t>
<t>Added applicability section.</t>
<t>Added discussion of Address Independence requirement.</t>
<t>Added hairpinning requirement and discussion of applicability
of other NAT behavioral requirements.</t>
</list></t>
</section>
<section title="Changes between *behave-nat66-02 and *nat66-00">
<t>There were further changes made between behave-nat66-02 and
nat66-02: <list style="symbols">
<t>Added mapping for prefixes longer than /48.</t>
<t>Change draft name to remove reference to the behave WG.</t>
<t>Resolved various open issues and fixed typos.</t>
</list></t>
</section>
<section title="Changes between *nat66-00 and *nat66-01">
<t><list style="symbols">
<t>Change the acronym "NAT66" to "NPTv6", so people don't read
"NAT" and MEGO.</t>
<t>Change the term used to refer to the function from "NAT66
device" to "NPTv6 Translator". It's not a "device" function, it's
a function that is applied between two interfaces. Consider a
router with two upstreams and two legs in the local network; it
will not translate between the local legs, but will translate to
and from each upstream, and be configured differently for each of
the two ISPs.</t>
<t>Comment specifically on the security aspects.</t>
<t>Comment specifically on the application issues raised on this
list.</t>
<t>Comment specifically on multihoming, load-sharing, and
asymmetric routing.</t>
<t>Spell out the hairpinning requirement and its implications.</t>
<t>Spell out the service provider side of Address
Independence.</t>
<t>00 focuses on the edge's view</t>
<t>Detail the algorithm in a manner clearer to the implementor (I
think)</t>
<t>Spell out the case for GSE-style DMZs between the edge and the
transit network, which is about the implications for the global
routing table.</t>
<t>Refer to xref
target="I-D.ietf-v6ops-cpe-simple-security"/>.</t>
</list></t>
</section>
</section>
</middle>
<back>
<references title="Normative References">
<?rfc include="reference.RFC.2119" ?>
<?rfc include="reference.RFC.2526" ?>
<?rfc include="reference.RFC.4193" ?>
<?rfc include="reference.RFC.4291" ?>
<?rfc include="reference.RFC.4443" ?>
<?rfc include="reference.RFC.4787" ?>
</references>
<references title="Informative References">
<?rfc include="reference.RFC.1071" ?>
<?rfc include="reference.RFC.1624" ?>
<?rfc include="reference.RFC.2629" ?>
<?rfc include="reference.RFC.2827" ?>
<?rfc include="reference.RFC.2993" ?>
<?rfc include="reference.RFC.4864" ?>
<?rfc include="reference.I-D.ietf-v6ops-cpe-simple-security" ?>
<reference anchor="GSE"
target="http://tools.ietf.org/id/draft-ietf-ipngwg-gseaddr">
<front>
<title>GSE - An Alternate Addressing Architecture for IPv6</title>
<author fullname=" Mike O'Dell" initials="M." surname="O'Dell">
<organization>UUNET Technologies</organization>
</author>
<date month="February" year="1997" />
</front>
<format target="http://tools.ietf.org/id/draft-ietf-ipngwg-gseaddr"
type="TXT" />
</reference>
</references>
<section anchor="appendix1" title="Why GSE?">
<t>For the purpose of this discussion, let us over-simplify the
Internet's structure by distinguishing between two broad classes of
networks: transit and edge. A "transit network", in this context, is a
network that provides connectivity services to other networks. Its AS
number may show up in a non-final position in BGP AS paths, or in the
case of mobile and residential broadband networks, it may offer network
services to smaller networks that can't justify RIR membership. An "edge
network", in contrast, is any network that is not a transit network; it
is the ultimate customer, and while it provides internal connectivity
for its own use, it is in other respects is a consumer of transit
services. In terms of routing, a network in the transit domain generally
needs some way to make choices about how it routes to other networks; an
edge network is generally quite satisfied with a simple default
route.</t>
<t>The <xref target="GSE"></xref> proposal, and as a result this
proposal (which is similar to GSE in most respects and inspired by it),
responds directly to current concerns in the RIR communities. Edge
networks are used to an environment in IPv4 in which their addressing is
disjoint from that of their upstream transit networks; it is either
provider independent, or a network prefix translator makes their
external address distinct from their internal address, and they like the
distinction. In IPv6, there is a mantra that edge network addresses
should be derived from their upstream, and if they have multiple
upstreams, edge networks are expected to design their networks to use
all of those prefixes equivalently. They see this as unnecessary and
unwanted operational complexity, and are as a result pushing very hard
in the RIR communities for provider independent addressing.</t>
<t>Widespread use of provider independent addressing has a natural and
perhaps unavoidable side-effect that is likely to be very expensive in
the long term. It means that the routing table will enumerate the
networks at the edge of the transit domain, the edge networks, rather
than enumerating the transit domain. Per the CIDR Report, there are
currently more than 36,000 Autonomous Systems being advertised in BGP,
of which over 15,000 advertise only one prefix. There are in the
neighborhood of 5000 AS's that show up in a non-final position in AS
paths, and perhaps another 5000 networks whose AS numbers are terminal
in more than one AS path. In other words, we have prefixes for some
36,000 transit and edge networks in the route table now, many of which
arguably need an Autonomous System number only for multihoming. Current
estimates suggest that we could easily see that be on the order of
10,000,000 within fifteen years. Tens of thousands of entries in the
route table is very survivable; while our protocols and computers will
likely do quite well with tens of millions of routes, the heat produced
and power consumed by those routers, and the inevitable impact on the
cost of those routers, is not a good outcome. To avoid having a massive
and unscalable route table, we need to find a way that is politically
acceptable and returns us to enumerating the transit domain, not the
edge.</t>
<t>There have been a number of proposals. As described, shim6 moves the
complexity to the edge, and the edge is rebelling. Geographic addressing
in essence forces ISPs to "own" geographic territory from a routing
perspective, as otherwise there is no clue in the address as to what
network a datagram should be delivered to in order to reach it.
Metropolitan Addressing can imply regulatory authority, and even if it
is implemented using internet exchange consortia, visits a great deal of
complexity on the transit networks that directly serve the edge. The one
that is likely to be most acceptable is any proposal that enables an
edge network to be operationally independent of its upstreams, with no
obligation to renumber when it adds, drops, or changes ISPs, and with no
additional burden placed either on the ISP or the edge network as a
result. From an application perspective, an additional operational
requirement in the words of US NIST's Roadmap for the Smart Grid, is
that <list style="empty">
<t>"...the Network should enable an application in a particular
domain to communicate with an application in any other domain in the
information network, with proper management control over who and
where applications can be interconnected."</t>
</list></t>
<t>In other words, the structure of the network should allow for and
enable appropriate access control, but the structure of the network
should not inherently limit access.</t>
<t>The GSE model, by statelessly translating the prefix between an edge
network and its upstream transit network, accomplishes that with a
minimum of fuss and bother. Stated in the simplest terms, it enables the
edge network to behave as if it has a provider-independent prefix from a
multihoming and renumbering perspective without the overhead of RIR
membership or maintaining BGP connectivity, and it enables the transit
networks to aggressively aggregate what are from their perspective
provider-allocated customer prefixes, to maintain a rational-sized
routing table.</t>
</section>
</back>
</rfc>
| PAFTECH AB 2003-2026 | 2026-04-24 04:23:44 |