Extension Mechanisms for DNS (EDNS0)Internet Systems Consortium950 Charter StreetRedwood CityCalifornia94063US+1 650.423.1312joao@isc.orgInternet Systems Consortium950 Charter StreetRedwood CityCalifornia94063US+1 650.423.1304mgraff@isc.orgInternet Systems Consortium950 Charter StreetRedwood CityCalifornia94063US+1 650.423.1301vixie@isc.org
General
DNSEXT Working GroupRFCRequest for CommentsI-DInternet-Draftedns0dns
The Domain Name System's wire protocol includes a number of fixed
fields whose range has been or soon will be exhausted and does not
allow requestors to advertise their capabilities to responders. This
document describes backward compatible mechanisms for allowing the
protocol to grow.
This document updates the EDNS0 specification (RFC2671) based on 10
years of deployment experience.
DNS specifies a Message Format and within such
messages there are standard formats for encoding options, errors, and
name compression. The maximum allowable size of a DNS Message is fixed.
Many of DNS's protocol limits are too small for uses which are or which
are desired to become common. There is no way for implementations to
advertise their capabilities.
Unextended agents will not know how to interpret the protocol
extensions detailed here. In practice, these clients will be upgraded
when they have need of a new feature, and only new features will make
use of the extensions. Extended agents must be prepared for behavior
of unextended clients in the face of new protocol elements, and fall
back gracefully to unextended DNS.
proposed extensions to the basic DNS protocol to overcome these
deficiencies.
This memo refines that specification and obsoletes
.
specified extended label types. The only one
ever proposed was in RFC2673 for a label type called "Bitstring Labels."
For various reasons introducing a new label type was found to be
extremely difficult, and RFC2673 was moved to Experimental.
This document Obsoletes Extended Labels.
"Requestor" is the side which sends a request. "Responder" is
an authoritative, recursive resolver, or other DNS component which
responds to questions.
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
RFC 2119.
EDNS support is practically mandatory in a modern world. DNSSEC requires EDNS
support, and many other Features are made possible only by EDNS support
to request or advertise them. Many organizations are beginning to
require DNSSEC. Without common interoperability, DNSSEC cannot be as
easily deployed.
DNS publishers are wanting to put more data in answers. DNSSEC
DNSKEY records, negative answers, and many other DNSSEC queries
cause larger answers to be returned. In order to support this, DNS
servers, middleware, and stub resolvers MUST support larger packet
sizes advertised via EDNS0.
The DNS Message Header's second full 16-bit word is divided into a
4-bit OPCODE, a 4-bit RCODE, and a number of 1-bit flags
(see , section
4.1.1). Some of these were marked for
future use, and most these have since been allocated. Also, most of
the RCODE values are now in use. The OPT pseudo-RR specified below
contains extensions to the RCODE bit field as well as additional flag bits.
The first two bits of a wire format domain label are used to denote the
type of the label.
allocates two of the four
possible types and reserves the other two.
More label types were defined in . This document
obsoletes the use of the 2-bit combination defined by
to identify extended label types.
Traditional DNS Messages are limited to 512 octets in size when sent
over UDP (). Today, many organizations wish
to return many records in a single reply, and special tricks are
needed to make the responses fit in this 512-byte limit.
Additionally, DNSSEC signatures can easily generate a much larger
response than a 512 byte message can hold.
EDNS0 is intended to address these larger packet sizes and continue to
use UDP. It specifies a way to advertise additional features such
as larger response size capability, which is intended to help avoid
truncated UDP responses which then cause retry over TCP.
The first octet in the on-the-wire representation of a DNS label
specifies the label type; the basic DNS specification
dedicates the two most significant bits of that octet for this purpose.
defined DNS label type 0b01 for use as an
indication for Extended Label Types. A specific Extended Label Type
is selected by the 6 least significant bits of the first octet. Thus,
Extended Label Types are indicated by the values 64-127 (0b01xxxxxx)
in the first octet of the label.
This document does not describe any specific Extended Label Type.
In practice, Extended Label Types are difficult to use due to support in
clients and intermediate gateways. Therefore, the registry of Extended
Label Types is requested to be closed. They cause interoperability
problems and at present no defined label types are in use.
Bitstring labels were originally created to solve problems with IPv6
reverse zones. Due to the problems of introducing a new label type
they were moved to experimental. This document moves them from
experimental to historical, making them obsoleted.
An OPT pseudo-RR (sometimes called a meta-RR) MAY be added to the
additional data section of a request.
The OPT RR has been assigned RR type 41.
If present in requests, compliant responders MUST include an OPT record
in responses.
An OPT record does not carry any DNS data. It is used only to contain
control information pertaining to the question and answer sequence of
a specific transaction. OPT RRs MUST NOT be cached, forwarded, or
stored in or loaded from master files.
The OPT RR MAY be placed anywhere within the additional data section.
Only one OPT RR MAY be included within any DNS message. If a message
with more than one OPT RR is received, a FORMERR MUST be returned.
An OPT RR has a fixed part and a variable set of options expressed as
{attribute, value} pairs. The fixed part holds some DNS meta data and
also a small collection of basic extension elements which we expect
to be so popular that it would be a waste of wire space to encode
them as {attribute, value} pairs.
The fixed part of an OPT RR is structured as follows:
Field NameField TypeDescriptionNAMEdomain nameempty (root domain)TYPEu_int16_tOPTCLASSu_int16_trequestor's UDP payload sizeTTLu_int32_textended RCODE and flagsRDLENu_int16_tdescribes RDATARDATAoctet stream{attribute,value} pairs
The variable part of an OPT RR is encoded in its RDATA and is structured
as zero or more of the following:
Assigned by Expert Review.
Size (in octets) of OPTION-DATA.
Varies per OPTION-CODE.
The order of appearance of option tuples is not guaranteed. If one
option modifies the behavior of another or multiple options are
related to one another in some way, they have the same effect
regardless of ordering in the RDATA wire encoding.
Any OPTION-CODE values not understood by a responder or requestor
MUST be ignored. Specifications of such options might wish to
include some kind of signaled acknowledgement. For example, an option
specification might say that if a responder sees option XYZ, it MUST
include option XYZ in its response.
The OPT record must not be cached.
If a requestor detects that
the remote end does not support EDNS0, it MAY issue queries
without an OPT record. It MAY cache this knowledge
for a brief time in order to avoid fallback delays in the
future. However, if DNSSEC or any future option using EDNS
is required, no fallback should be performed as they are only
signaled through EDNS0.
The requestor's UDP payload size (which OPT stores in the RR CLASS
field) is the number of octets of the largest UDP payload that can be
reassembled and delivered in the requestor's network stack.
Note that path MTU, with or without fragmentation, may be smaller
than this. Values lower than 512 MUST be treated as equal to 512.
Requestors SHOULD place a value in this field that it can actually
receive. For example, if a requestor sits behind a firewall which
will block fragmented IP packets, a requestor SHOULD not choose a
value which will cause fragmentation. Doing so will prevent large
responses from being received, and can cause fallback to occur.
Note that a 512-octet UDP payload requires a 576-octet IP reassembly
buffer. Choosing between 1280 and 1410 bytes for IP (v4 or v6) over Ethernet
would be reasonable.
Choosing a very large value will guarantee fragmentation at the IP
layer, and may prevent answers from being received due to a single
fragment loss or misconfigured firewalls.
The requestor's maximum payload size can change over time. It
MUST not be cached for use beyond the transaction in which it
is advertised.
The responder's maximum payload size can change over time, but can
be reasonably expected to remain constant between two closely spaced sequential
transactions; for example, a meaningless QUERY to discover a
responder's maximum UDP payload size, followed immediately by an
UPDATE which takes advantage of this size. This is considered
preferable to the outright use of TCP for oversized requests, if
there is any reason to suspect that the responder implements
EDNS, and if a request will not fit in the default 512 payload size
limit.
Due to transaction overhead, it is unwise to advertise an
architectural limit as a maximum UDP payload size. Just because
your stack can reassemble 64KB datagrams, don't assume that you want
to spend more than about 4KB of state memory per ongoing transaction.
A requestor MAY choose to implement a fallback to smaller advertised
sizes to work around firewall or other network limitations.
A requestor SHOULD choose to use a fallback mechanism which begins
with a large size, such as 4096. If that fails, a fallback around
the 1280 byte range SHOULD be tried, as it has a reasonable chance
to fit within a single Ethernet frame. Failing that, a requestor MAY
choose a 512 byte packet, which with large answers may cause a TCP
retry.
Middleware boxes (e.g. firewalls, SOHO routers, load balancers, etc)
MUST NOT limit DNS messages over UDP to 512 bytes.
Middleware boxes which simply forward requests to a recursive
resolver MUST NOT modify and MUST NOT delete the OPT record
contents in either direction.
Middleware boxes which have additional functionality, such as
answering certain queries or acting like an intelligent forwarder,
MUST understand the OPT record. These boxes MUST consider the
incoming request and any outgoing requests as separate transactions
if the characteristics of the messages are different.
The extended RCODE and flags (which OPT stores in the RR TTL field)
are structured as follows:
Forms upper 8 bits of extended 12-bit RCODE (together with the 4 bits
defined in . Note that
EXTENDED-RCODE value 0 indicates that an unextended RCODE
is in use (values 0 through 15).
Indicates the implementation level of whoever sets it. Full
conformance with this specification is indicated by version ``0.''
Requestors are encouraged to set this to the lowest implemented
level capable of expressing a transaction, to minimize the responder
and network load of discovering the greatest common implementation
level between requestor and responder. A requestor's version
numbering strategy MAY ideally be a run time configuration option.
If a responder does not implement the VERSION level of the
request, then it answers with RCODE=BADVERS. All responses MUST
be limited in format to the VERSION level of the request, but the
VERSION of each response SHOULD be the highest implementation
level of the responder. In this way a requestor will learn the
implementation level of a responder as a side effect of every
response, including error responses and including RCODE=BADVERS.
DNSSEC OK bit as defined by .
Set to zero by senders and ignored by receivers,
unless modified in a subsequent specification.
Allocations assigned by expert review.
Assignment of Option Codes should be liberal, but duplicate
functionality is to be avoided.
The presence of an OPT pseudo-RR in a request should be taken as an
indication that the requestor fully implements the given version of EDNS,
and can correctly understand any response that conforms to that
feature's specification.
Lack of presence of an OPT record in a request MUST be taken as an
indication that the requestor does not implement any part of this
specification and that the responder MUST NOT include an OPT record in
its response.
Responders who do not implement these protocol extensions MUST respond
with FORMERR messages without any OPT record.
If there is a problem with processing the OPT record itself, such as an
option value that is badly formatted or includes out of range values, a
FORMERR MUST be returned. If this occurs the response MUST include an
OPT record. This is intended to allow the requestor to to distinguish
between servers which do not implement EDNS and format errors within
EDNS.
The minimal response must be the DNS header, question
section, and an OPT record. This must also occur when an truncated
response (using the DNS header's TC bit) is returned.
Requestor-side specification of the maximum buffer size may open a
DNS denial of service attack if responders can be made to send
messages which are too large for intermediate gateways to forward,
thus leading to potential ICMP storms between gateways and responders.
Announcing very large UDP buffer sizes may result in dropping by
middleboxes (see ). This could cause retransmissions
with no hope of success.
Some devices have been found to reject fragmented UDP packets.
Announcing too small UDP buffer sizes may result in fallback to TCP with
a corresponding load impact on DNS servers.
This is especially important with DNSSEC, where answers are much larger.
The IANA has assigned RR type code 41 for OPT.
specified a number of IANA sub-registries
within "DOMAIN NAME SYSTEM PARAMETERS:"
EDNS Extended Label TypeEDNS Option CodesEDNS Version NumbersDomain System Response Code
IANA is advised to re-parent these sub-registries to this document.
created the "EDNS Extended Label Type Registry".
We request that this registry be closed.
This document assigns option code 65535 in the "EDNS Option Codes"
registry to "Reserved for future expansion."
expands the RCODE space from 4 bits to 12 bits.
This allows more than the 16 distinct RCODE values
allowed in . IETF Standards
Action is required to add a new RCODE. Adding new RCODEs should
be avoided due to the difficulty in upgrading the installed base.
This document assigns EDNS Extended RCODE 16 to "BADVERS".
IETF Standards Action is required for assignments of new EDNS0 flags.
Flags SHOULD be used only when necessary for DNS resolution to
function. For many uses, a EDNS Option Code may be preferred.
IETF Standards Action is required to create new entries in the
EDNS Version Number registry. Expert Review is required for
allocation of an EDNS Option Code.
Following is a list of high-level changes made to the original
RFC2671.
Support for the OPT record is now mandatory.
Extended label types obsoleted and the registry is closed.
The bitstring label type, which was already moved from draft
to experimental, is requested to be moved to historical.
Changes in how EDNS buffer sizes are selected, with recommendations
on how to select them.
Front material (IPR notice and such) was updated to current
requirements.
Specified the method for allocation of constants.
Cleaned up a lot of wording, along with quite a bit of document
structure changes.