DNS Extensions Working Group                               Edward Lewis
INTERNET-DRAFT                                            NeuStar, Inc.
Expires: Octopber 1, 2009                                  April 2009
Updates: 1034, 1035 (if approved)
Intended status: Standards Track

                    DNS Zone Transfer Protocol (AXFR)
                  draft-ietf-dnsext-axfr-clarify-11.txt

Status of this Memo

  This Internet-Draft is submitted to IETF in full conformance with the
  provisions of BCP 78 and BCP 79.

  Internet-Drafts are working documents of the Internet Engineering
  Task Force (IETF), its areas, and its working groups.  Note that
  other groups may also distribute working documents as Internet-
  Drafts.

  Internet-Drafts are draft documents valid for a maximum of six months
  and may be updated, replaced, or obsoleted by other documents at any
  time.  It is inappropriate to use Internet-Drafts as reference
  material or to cite them other than as "work in progress."

  The list of current Internet-Drafts can be accessed at
  http://www.ietf.org/1id-abstracts.html

  The list of Internet-Draft Shadow Directories can be accessed at
  http://www.ietf.org/shadow.html

  This Internet-Draft will expire on October 1, 2009.

Copyright Notice

   Copyright (c) 2009 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
   (http://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.

Abstract

The Domain Name System standard mechanisms for maintaining coherent
servers for a zone consist of three elements.  One mechanism is the
Authoritative Transfer (AXFR) is defined in RFC 1034 and RFC 1035.
The definition of AXFR, has proven insufficient in detail, forcing
implementations intended to be compliant to make assumptions, impeding
interoperability. Yet today we have a satisfactory set of
implementations that do interoperate. This document is a new
definition of the AXFR, new in the sense that is it recording an
accurate definition of an interoperable AXFR mechanism.

 Introduction

The Domain Name System standard facilities for maintaining coherent
servers for a zone consist of three elements.  Authoritative
Transfer (AXFR) is defined in "Domain Names - Concepts and Facilities"
[RFC1034] (referred to in this document as RFC 1034) and "Domain Names
- Implementation and Specification" [RFC1035] (aka  RFC 1035).
Incremental Zone Transfer (IXFR) is defined in "Incremental Zone
Transfer in DNS" [RFC1995].  A mechanism for prompt notification of
zone changes (NOTIFY) is defined in "A Mechanism for Prompt
Notification of Zone Changes (DNS NOTIFY)" [RFC1996].  The goal of
these mechanisms is to enable a set of DNS name servers to remain
coherently authoritative for a given zone.

Comments on this draft ought to be addressed to the editor or to
namedroppers@ops.ietf.org.

1.1 Definition of Terms

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 "Key words for use in
RFCs to Indicate Requirement Levels" [BCP14].

"Newer"/"New" DNS and "older"/"old" DNS refers to implementations
written after and prior to the publication of this document.

1.2 Scope

In the greater context there are many ways to achieve coherency among
a set of name servers.  The AXFR, IXFR and NOTIFY mechanisms form
just one, the one defined in the RFCs cited.  For example, there are
DNS implementations that assemble answers from data stored in
relational databases (as opposed to master files) relying on the
database's non-DNS means to synchronize the database instances.  Some
of these non-DNS solutions interoperate in some fashion.  As far as
it is known, AXFR, IXFR and NOTIFY are the only in-band mechanisms
that provide an interoperable solution to the desire for coherency
within the definition of DNS, they certainly are the only mechanisms
documented by the IETF.

This document does not cover incoherent DNS situations.  There are
applications of the DNS in which servers for a zone are designed to be
incoherent.  For these configurations, a coherency mechanism as
described here would be unsuitable.

"General purpose DNS implementation" refers to DNS software developed
for wide-spread use.  This includes resolvers and servers freely
accessible as libraries and standalone processes.  This also includes
proprietary implementations used only in support of DNS service
offerings.

"Turnkey DNS implementation" refers to custom made, single use
implementations of DNS.  Such implementations consist of software
that employs the DNS protocol message format yet do not conform to
the entire range of DNS functionality.

A DNS implementation is not required to support AXFR, IXFR and NOTIFY.
A DNS implementation SHOULD have some means for maintaining name server
coherency.  A general purpose DNS implementation SHOULD include AXFR
(and in the same vein IXFR and NOTIFY), but turnkey DNS implementations
MAY exist without AXFR.  (An editorial note to readers of this section.
The mention of IXFR and NOTIFY is for context and illustration, this
document does not make any normative comments on those mechanisms.)

1.3 Context

Besides describing the mechanisms themselves, there is the context in
which they operate to consider.  When AXFR, IXFR and NOTIFY were
defined, there was little consideration given to security and privacy
issues.  Since the original definition of AXFR, new opinions have
appeared on the access to an entire zone's contents.  In this document,
the basic mechanisms will be discussed separately from the permission
to use these mechanisms.

1.4 Coverage and Relationship to Original AXFR Specification

This document concentrates on just the definition of AXFR.  Any effort
to update the IXFR or NOTIFY mechanisms would be done in different
documents.

The original "specification" of the AXFR sub-protocol is scattered
through RFC 1034 and RFC 1035.  Section 2.2 of RFC 1035 on page 5
depicts the scenario for which AXFR has been designed.  Section 4.3.5
of RFC 1034 describes the zone synchronization strategies in general
and rules for the invocation of a full zone transfer via AXFR; the
fifth paragraph of that section contains a very short sketch of the
AXFR protocol; Section 5.5 of RFC 2181 has corrected a significant
flaw in that specification.  Section 3.2.3 of RFC 1035 has assigned
the code point for the AXFR QTYPE (see section 2.1.2 below for more
details).  Section 4.2 of RFC 1035 discusses the transport layer use
of DNS and shortly explains why UDP transport is deemed inappropriate
for AXFR; the last paragraph of Section 4.2.2 gives details for the
TCP connection management with AXFR.  Finally, the second paragraph
of Section 6.3 in RFC 1035 mandates server behavior when zone data
changes occur during an ongoing zone transfer using AXFR.

This document will update the specification of AXFR in fully
specifying the record formats and processing rules for AXFR, largely
expanding on paragraph 5 of Section 4.3.5 of RFC 1034, and detailing
the transport considerations for AXFR, thus amending Section 4.2.2 of
RFC 1035.  Furthermore, it discusses backward compatibility issues
and provides policy/management considerations as well as specific
Security Considerations for AXFR.  The goal of this document is to
define AXFR as it exists, or is supposed to exist, currently.

 AXFR Messages

An AXFR session consists of an AXFR query message and the sequence of
AXFR response messages returned for it.  In this document, the AXFR
client is the sender of the AXFR query and the AXFR server is the
responder.  (Use of terms such as master, slave, primary, secondary
are not important to defining AXFR.)  The use of the word "session"
without qualification refers to an AXFR session.

An important aspect to keep in mind is that the definition of AXFR is
restricted to TCP [RFC0793].  The design of the AXFR process has
certain inherent features that are not easily ported to UDP [RFC0768].

The basic format of an AXFR message is the DNS message as defined in
RFC 1035, Section 4 ("MESSAGES") [RFC1035], updated by the following:
- "A Mechanism for Prompt Notification of Zone Changes (...)" [RFC1996]
- "Domain Name System (DNS) IANA Considerations" [RFC5395]
- "Dynamic Updates in the Domain Name System (DNS UPDATE)" [RFC2136]
- "Clarifications to the DNS Specification" [RFC2181]
- "Extension Mechanisms for DNS (EDNS0)" [RFC2671]
- "Secret Key Transaction Authentication for DNS (TSIG)" [RFC2845]
- "Secret Key Establishment for DNS (TKEY RR)" [RFC2930]
- "Obsoleting IQUERY" [RFC3425]
- "Handling of Unknown DNS Resource Record (RR) Types" [RFC3597]
- "Resource Records for the DNS Security Extensions" [RFC4034]
- "Protocol Modifications for the DNS Security Extensions" [RFC4035]
- "Use of SHA-256 in DNSSEC ... (DS) ... (RRs)" [RFC4509]
- "HMAC SHA TSIG Algorithm Identifiers" [RFC4635]
- "... (DNSSEC) Hashed Authenticated Denial of Existence" [RFC5155]

For completeness, the following, in process, documents contain
information about the DNS message.  These documents ought not interfere
with AXFR but these documents are helpful in understanding what will
be carried via AXFR.

- "Use of SHA-2 algorithms with RSA in DNSKEY and RRSIG Resource
  Records for DNSSEC" [DRAFT1]
- "Clarifications and Implementation Notes for DNSSECbis" [DRAFT2]

The upper limit on the permissible size of a DNS message over TCP is
only restricted by the TCP framing defined in RFC 1035, section 4.2.2
which specifies a two-octet message length field, understood to be
unsigned, and thus causing a limit of 65535 octets.  Unlike DNS
messages over UDP, this limit is not changed by EDNS0.

Note that the TC (truncation) bit is never set by an AXFR server nor
considered/read by an AXFR client.

Field names used in this document will correspond to the names as they
appear in the IANA registry for DNS Header Flags [DNSFLGS].

2.1 AXFR query

An AXFR query is sent by a client whenever there is a reason to ask.
This might be because of zone maintenance activities or as a result of
a command line request, say for debugging.

An AXFR query is sent by a client whenever there is a reason to ask.
This might be because of scheduled or triggered zone maintenance
activities (see section 4.3.5 of RFC 1034 and DNS NOTIFY [RFC1996],
respectively) or as a result of a command line request, say for
debugging.

2.1.1 Header Values

These are the DNS message header values for an AXFR query.

ID          See note 2.1.1.a
QR          MUST be 0 (Query)
OPCODE      MUST be 0 (Standard Query)
AA          See note 2.1.1.b
TC          See note 2.1.1.b
RD          See note 2.1.1.b
RA          See note 2.1.1.b
Z           See note 2.1.1.c
AD          See note 2.1.1.b
CD          See note 2.1.1.b
RCODE       MUST be 0 (No error)
QDCOUNT     MUST be 1
ANCOUNT     MUST be 0
NSCOUNT     MUST be 0
ARCOUNT     See note 2.1.1.d

Note 2.1.1.a Set to any value that the client is not already using
with the same server.  There is no specific means for selecting the
value in this field.  (Recall that AXFR is done only via TCP
connections.)

A server MUST reply using messages that use the same message ID to
allow a client to meaningfully have multiple AXFR queries outstanding.

Note 2.1.1.b The value in this field has no meaning in the context of
AXFR query messages.  For the client, it is RECOMMENDED that the
value be zero.  The server MUST ignore this value.

Note 2.1.1.c The client MUST set this bit to 0, the server MUST ignore
it.

Note 2.1.1.d The client MUST set this field to be the number of
resource records appearing in the additional section.  See Section
2.1.5 "Additional Section" for details.

The value MAY be 0, 1 or 2.  If it is 2, the additional
section MUST contain both an EDNS0 [RFC2671] OPT resource record and
a record carrying transaction integrity and authentication data,
currently a choice of TSIG [RFC2845] and SIG(0) [RFC2931].  If the
value is 1, then the additional section MUST contain either only an
EDNS0 OPT resource record or a record carrying transaction integrity
and authentication data.  If the value is 0, the additional section
MUST be empty.

2.1.2 Query Section

The Query section of the AXFR query MUST conform to section 4.1.2 of
RFC 1035, and contain the following values:

QNAME       the name of the zone requested
QTYPE       AXFR(= 252), the pseudo-RR type for zone transfer [DNSVALS]
QCLASS      the class of the zone requested

2.1.3 Answer Section

MUST be empty.

2.1.4 Authority Section

MUST be empty.

2.1.5 Additional Section

The client MAY include an EDNS0 OPT [RFC2671] resource record.  If the
server has indicated that it does not support EDNS0, the client MUST
send this section without an EDNS0 OPT resource record if there is a
retry.  Indication that a server does not support EDNS0 is not an
explicit element in the protocol, it is up to the client to interpret.
Most likely, the server will return a FORMERR which might be related to
the OPT resource record.

The client MAY include a transaction integrity and authentication
resource record, currently a choice of TSIG [RFC2845] or SIG(0)
[RFC2931].  If the server has indicated that it does not recognize the
resource record, and that the error is indeed caused by the resource
record, the client probably ought not try again.  Removing the security
data in the face of an obstacle ought to only be done with full
awareness of the implication of doing so.

In general, if an AXFR client is aware that an AXFR server does not
support a particular mechanism, the client SHOULD NOT attempt to engage
the server using the mechanism (or at all).  A client could become
aware of a server's abilities via a configuration setting or via some
other (as yet) undefined means.

The range of permissible resource records that MAY appear in the
additional section might change over time.  If either a change to an
existing resource record (like the OPT RR for EDNS0) is made or
a new additional section record is created, the new definitions ought
to include a discussion on the impact upon AXFR.  Although this is not
predictable, future additional section residing records may have an
effect that is orthogonal to AXFR, so can ride through the session as
opaque data.  In this case, a "wise" implementation ought to be able
to pass these records through without disruption.

2.2 AXFR response

The AXFR response will consist of 0 or more messages.  A "0 message"
response is covered in section 2.2.1.

An AXFR response that is transferring the zone's contents will consist
of a series (which could be a series of length 1) of DNS messages.
In such a series, the first message MUST begin with the SOA
resource record of the zone, the last message MUST conclude with the
same SOA resource record.  Intermediate messages MUST NOT contain the
SOA resource record.  The first message MUST copy the Query Section
from the corresponding AXFR query message in to the first response
message's query section.  Subsequent messages MAY do the same.

An AXFR response that is indicating an error MUST consist of a single
DNS message with the return code set to the appropriate value for the
condition encountered - once the error condition is detected. Such
a message MUST terminate the AXFR session; it MUST copy the Query
Section from the AXFR query into its Query Section, but the inclusion
of the terminating SOA resource record is not necessary.

An AXFR client might receive a number of AXFR response messages
free of an error condition before the message indicating an error
is received.

2.2.1 "0 Message" Response

A legitimate "0 message" response, i.e., the client sees no response
whatsoever, is very exceptional and controversial.  Unquestionably it
is unhealthy for there to be 0 responses in a protocol that is designed
around a query - response paradigm over an unreliable transport.  The
lack of a response could be a sign of underlying network problems and
cause the protocol state machine to react accordingly.  However, AXFR
uses TCP and not UDP, eliminating undetectable network errors.

A "0 message response" is reserved for situations in which the server
has a reason to suspect that the query is sent for the purpose of
abuse.  Due to the use of this being so controversial, a "0 message
response" is not being defined as a legitimate part of the protocol
but the use of it is being acknowledged as a warning to AXFR client
implementations.  Any earnest query has the expectation of some
response but may not get one.

2.2.2 Header Values

ID          See note 2.2.2.a
QR          MUST be 1 (Response)
OPCODE      MUST be 0 (Standard Query)
AA          See note 2.2.2.b
TC          MUST be 0 (Not truncated)
RD          RECOMMENDED copy request's value, MAY be set to 0
RA          See note 2.2.2.c
Z           See note 2.2.2.d
AD          See note 2.2.2.e
CD          See note 2.2.2.e
RCODE       See note 2.2.2.f
QDCOUNT     MUST be 1 in the first message; MUST be 0 or 1 in all
            following
ANCOUNT     See note 2.2.2.g
NSCOUNT     MUST be 0
ARCOUNT     See note 2.2.2.h

Note 2.2.2.a Because some old implementations behave differently than
is now desired, the requirement on this field is stated in detail.
New DNS servers MUST set this field to the value of the AXFR query
ID in each AXFR response message for the session.  AXFR clients MUST
be able to manage sessions resulting from the issuance of multiple
outstanding queries, whether AXFR queries or other DNS queries.  A
client SHOULD discard responses that do not correspond (via the
message ID) to any outstanding queries.

Unless the client is sure that the server will consistently set the ID
field to the query's ID, the client is NOT RECOMMENDED to issue any
other queries until the end of the zone transfer.  A client MAY become
aware of a server's abilities via a configuration setting.

Note 2.2.2.b If the RCODE is 0 (no error), then the AA bit MUST be 1.
For any other value of RCODE, the AA bit MUST be set according to rules
for that error code.  If in doubt, it is RECOMMENDED that it be set
to 1.  It is RECOMMENDED that the value be ignored by the AXFR client.

Note 2.2.2.c It is RECOMMENDED that the server set the value to 0,
the client MUST ignore this value.

The server MAY set this value according to the local policy regarding
recursive service, but doing so might confuse the interpretation of the
response as AXFR can not be retrieved recursively.  A client MAY note
the server's policy regarding recursive service from this value, but
SHOULD NOT conclude that the AXFR response was obtained recursively
even if the RD bit was 1 in the query.

Note 2.2.2.d The server MUST set this bit to 0, the client MUST ignore
it.

Note 2.2.2.e If the implementation supports the DNS Security Extensions
(see below) then this value MUST be set according to the rules in RFC
4035, section 3.1.6, "The AD and CD Bits in an Authoritative Response".
If the implementation does not support the DNS Security Extensions,
then this value MUST be set to 0 and MUST be ignored upon receipt.

The DNS Security Extensions (DNSSEC) is defined in these base
documents:
- "DNS Security Introduction and Requirements" [RFC4033]
- "Resource Records for the DNS Security Extensions" [RFC4034]
- "Protocol Modifications for the DNS Security Extensions" [RFC4035]
- "Use of SHA-256 in DNSSEC Delegation Signer RRs" [RFC4509]
- "DNS Security Hashed Authenticated Denial of Existence" [RFC5155]

as well pending documents, such as these:

- "Use of SHA-2 algorithms with RSA in DNSKEY and RRSIG Resource
  Records for DNSSEC" [DRAFT1]
- "Clarifications and Implementation Notes for DNSSECbis" [DRAFT2]

Note 2.2.2.f In the absence of an error, the server MUST set the value
of this field to NoError.  If a server is not authoritative for the
queried zone, the server SHOULD set the value to NotAuth.  (Reminder,
consult the appropriate IANA registry [DNSVALS].)  If a client
receives any other value in response, it MUST act according to the
error.  For example, a malformed AXFR query or the presence of an EDNS0
OPT resource record sent to an old server will garner a FormErr value.
This value is not set as part of the AXFR-specific response processing.
The same is true for other error-indicating values.

Note 2.2.2.g The count of answer records MUST equal the number of
resource records in the AXFR Answer Section.  When a server is aware
that a client will only accept one resource record per response
message, then the value MUST be 1.  A server MAY be made aware of a
client's limitations via configuration data.

Note 2.2.2.h The client MUST set this field to be the number of
resource records appearing in the additional section.  The
considerations in Note 2.1.1.d above apply equally; see Section
2.2.6 "Additional Section" below for more details.

2.2.3 Query Section

In the first response message, this section MUST be copied from the
query.  In subsequent messages, this section MAY be copied from the
query or it MAY be empty.  The content of this section MAY be used to
determine the context of the message, that is, the name of the zone
being transferred.

2.2.4 Answer Section

MUST be populated with the zone contents.  See later section on
encoding zone contents.

2.2.5 Authority Section

MUST be empty.

2.2.6 Additional Section

The contents of this section MUST follow the guidelines for EDNS0,
TSIG, SIG(0), or what ever other future record is possible here.  The
contents of section 2.1.5 apply here as well.

2.3 TCP Connection Aborts

If an AXFR client sends a query on a TCP connection and the connection
is closed at any point, the AXFR client MUST consider the AXFR session
terminated.  The message ID MAY be used again on a new connection,
even if the question and AXFR server are the same.  Facing a dropped
connection a client SHOULD try to make some determination whether the
connection closure was the result of network activity or a decision
by the AXFR server.  This determination is not an exact science.  It
is up to the AXFR client implementor to react, but the reaction
SHOULD NOT be an endless cycle of retries nor an increasing (in
frequency) retry rate.

An AXFR server implementor SHOULD take into consideration the dilemma
described above when a connection is closed with an outstanding query
in the pipeline.  For this reason, a server ought to reserve this
course of action for situations in which it believes beyond a doubt
that the AXFR client is attempting abusive behavior.

 Zone Contents

The objective of the AXFR session is to request and transfer the
contents of a zone.  The objective is to permit the AXFR client to
reconstruct the zone as it exists at the server for the given zone
serial number.  Over time the definition of a zone has evolved from
denoting a static set of records to also cover a dynamically updated
set of records, and then a potentially continually regenerated set of
records as well.

3.1 Records to Include

In the answer section of AXFR response messages the resource records
within a zone for the given serial number MUST appear.  The definition
of what belongs in a zone is described in RFC 1034, Section 4.2, "How
the database is divided into zones", in particular, section 4.2.1,
"Technical considerations", and it has been clarified in Section 6 of
RFC 2181.

Unless the AXFR server knows that the AXFR client is old and expects
just one resource record per AXFR response message, an AXFR server
SHOULD populate an AXFR response message with as many complete
resource record sets as will fit within a DNS message.

Zones for which it is impractical to list the entire zones for a serial
number are not suitable for AXFR retrieval.  A typical (but not
limiting) description of such a zone is a zone consisting of responses
generated via other database lookups and/or computed based upon ever
changing data.

3.2 Delegation Records

In RFC 1034, section 4.2.1, this text appears (keep in mind that the
"should" in the quotation predates [BCP14], cf. section 1.1) "The RRs
that describe cuts ... should be exactly the same as the corresponding
RRs in the top node of the subzone."  There has been some controversy
over this statement and the impact on which NS resource records are
included in a zone transfer.

The phrase "that describe cuts" is a reference to the NS set and
applicable glue records.  It does not mean that the cut point and apex
resource records are identical.  For example, the SOA resource record
is only found at the apex.  The discussion here is restricted to just
the NS resource record set and glue as these "describe cuts".

DNSSEC resource records have special specifications regarding their
occurrence at a zone cut and the apex of a zone.  This has for the
first time been described in Sections 5.3 ff. and 6.2 of RFC 2181
(for the initial specification of DNSSEC), which now is historical.
The current DNSSEC core document set (see Note 2.2.2.e above) gives
the full details for DNSSEC(bis) resource record placement, and
Section 3.1.5 of RFC 4035 normatively specifies their treatment during
AXFR; the alternate NSEC3 resource record defined later in RFC 5155
behaves identically as the NSEC RR, for the purpose of AXFR.

Informally:
o  The DS RRSet only occurs at the parental side of a zone cut and is
   authoritative data in the parent zone, not the secure child zone.
o  The DNSKEY RRSet only occurs at the APEX of a signed zone and is
   authoritative part of the zone it serves.
o  Independent RRSIG RRSets occur at the signed parent side and of a
   zone cut and at the apex of a signed zone; they are authoritative
   part of the respective zone; simple queries for RRSIG resource
   records may return bth RRSets at once if the same server is
   authoritative for the parent zone and the child zone (Section
   3.1.5 of RFC 4035 describes how to distinguish these RRs); this
   seeming ambiguity does not occur for AXFR, since each such RRSIG
   RRset belongs to a single zone.
o  Different NSEC [RFC4034] or NSEC3 [RFC5155] resource records
   equally may occur at the parental siede of a zone cut and at the
   apex of a zone; each such resource record belongs to exactly one
   of these zones and is to be included in the AXFR of that zone.

The issue is that in operations there are times when the NS resource
records for a zone might be different at a cut point in the parent and
at the apex of a zone.  Sometimes this is the result of an error and
sometimes it is part of an ongoing change in name servers.  The DNS
protocol is robust enough to overcome inconsistencies up to (but not
including) there being no parent indicated NS resource record
referencing a server that is able to serve the child zone.  This
robustness is one quality that has fueled the success of the DNS.
Still, the inconsistency is an error state and steps need to be taken
to make it apparent (if it is unplanned) and to make it clear once
the inconsistency has been removed.

Another issue is that the AXFR server could be authoritative for a
different set of zones than the AXFR client.  It is possible that the
AXFR server be authoritative for both halves of an inconsistent cut
point and that the AXFR client is authoritative for just the parent of
the cut point.

The question that arises is, when facing a situation in which a cut
point's NS resource records do not match the authoritative set, whether
an AXFR server responds with the NS resource record set that is in the
zone being transferred or is at the authoritative location.

The AXFR response MUST contain the cut point NS resource record set
registered with the zone whether it agrees with the authoritative set
or not.  "Registered with" can be widely interpreted to include data
residing in the zone file of the zone for the particular serial
number (in zone file environments) or as any data configured to be in
the zone (database), statically or dynamically.

The reasons for this requirement are:

1) The AXFR server might not be able to determine that there is an
inconsistency given local data, hence requiring consistency would mean
a lot more needed work and even network retrieval of data.  An
authoritative server ought not be required to perform any queries.

2) By transferring the inconsistent NS resource records from a server
that is authoritative for both the cut point and the apex to a client
that is not authoritative for both, the error is exposed.  For example,
an authorized administrator can manually request the AXFR and inspect
the results to see the inconsistent records.  (A server authoritative
for both halves would otherwise always answer from the more
authoritative set, concealing the error.)

3) The inconsistent NS resource record set might indicate a problem
in a registration database.

4) This requirement is necessary to ensure that retrieving a given
(zone,serial) pair by AXFR yields the exact same set of resource
records no matter which of the zone's authoritative servers is
chosen as the source of the transfer.

If an AXFR server were allowed to respond with the authoritative
NS RRset of a child zone instead of a glue NS RRset in the zone
being transferred, the set of records returned could vary depending
on whether or not the server happens to also be authoritative for
the child zone.

The property that a given (zone,serial) pair corresponds to a
single, well-defined set of records is necessary for the correct
operation of incremental transfer protocols such as IXFR
[RFC1995].  For example, a client may retrieve a zone by AXFR from
one server, and then apply an incremental change obtained by IXFR
from a different server.  If the two servers have different ideas
of the zone contents, the client can end up attempting to
incrementally add records that already exist or to delete records
that do not exist.

3.3 Glue Records

As quoted in the previous section, section 4.2.1 of RFC 1034 provides
guidance and rationale for the inclusion of glue records as part of
an AXFR transfer.  And, as also argued in the previous section of this
document, even when there is an inconsistency between the address in a
glue record and the authoritative copy of the name server's address,
the glue resource record that is registered as part of the zone for
that serial number is to be included.

This applies to glue records for any address family [RFC1700].

The AXFR response MUST contain the appropriate glue records as
registered with the zone.  The interpretation of "registered with"
in the previous section applies here.  Inconsistent glue records are
an operational matter.

3.4 Name Compression

Compression of names in DNS messages is described in RFC 1035, section 4.1.4, "Message compression".  The issue highlighted here relates to a
comment made in RFC 1034, section 3.1, "Name space specifications and
terminology" which says "When you receive a domain name or label, you
should preserve its case." ("Should" in the quote predates [BCP14].)

Name compression in an AXFR message MUST preserve the case of the
original domain name.  That is, although when comparing a domain name,
"a" equals "A", when comparing for the purposes of message compression,
"a" is not equal to "A".  Note that this is not the usual definition
of name comparison in the DNS protocol and represents a new
requirement on AXFR servers.

Rules governing name compression of RDATA in an AXFR message MUST
abide by the specification in "Handling of Unknown DNS Resource Record
(RR) Types" [RFC3597], specifically, section 4 on "Domain Name
Compression."

3.5 Occluded Names

Dynamic Update [RFC2136] operations, and in particular its interaction
with DNAME [RFC2672], can have a side effect of occluding names in a
zone.  The addition of a delegation point via dynamic update will
render all subordinate domain names to be in a limbo, still part of
the zone but not available to the lookup process.  The addition of a
DNAME resource record has the same impact.  The subordinate names are
said to be "occluded."

Occluded names MUST be included in AXFR responses.  An AXFR client MUST
be able to identify and handle occluded names.  The rationale for this
action is based on a speedy recovery if the dynamic update operation
was in error and is to be undone.

 Transport

AXFR sessions are currently restricted to TCP by section 4.3.5 of RFC
 that states: "Because accuracy is essential, TCP or some other
reliable protocol must be used for AXFR requests."  The restriction to
TCP is also mentioned in section 6.1.3.2. of "Requirements for Internet
Hosts - Application and Support" [RFC1123].

The most common scenario is for an AXFR client to open a TCP connection
to the AXFR server, send an AXFR query, receive the AXFR response, and
then close the connection.  There are variations on this, such as a
query for the zone's SOA resource record first, and so on.  Note that
this is documented as a most common scenario.

The assumption that a TCP connection is dedicated to the single AXFR
session is incorrect, this has led to implementation choices that
prevent either multiple concurrent zone transfers or the use of the
open connection for other queries.

Being able to have multiple concurrent zone transfers is considered
desirable by operators who have sets of name servers that are
authoritative for a common set of zones.  It would be desirable
if the name server implementations did not have to wait for one
zone to transfer before the next could begin.  The desire here is to
tighten the specification, not a change, but adding words to the
unclear areas, to define what is needed to permit two servers to
share a TCP connection among concurrent AXFR sessions.  The challenge
is to design this in a way that can fall back to the old behavior if
either the AXFR client or AXFR server is incapable of performing
multiple concurrent AXFR sessions.

With the addition of EDNS0 and applications which require many
small zones such as in web hosting and some ENUM scenarios, AXFR
sessions on UDP would now be possible and seem desirable.  However,
there are still some aspects of the AXFR session that are not easily
translated to UDP.  This document leaves AXFR over UDP undefined.

4.1 TCP

In the original definition there is an implicit assumption (probably
unintentional) that a TCP connection is used for one and only one
AXFR session.  This is evidenced in no requirement to copy neither
the Query Section nor the message ID in responses, no explicit
ordering information within the AXFR response messages and the lack
of an explicit notice indicating that a zone transfer continues in the
next message.

The guidance given here is intended to enable better performance of
the AXFR exchange as well as guidelines on interactions with older
software.  Better performance includes being able to multiplex DNS
message exchanges including zone transfer sessions.  Guidelines for
interacting with older software are generally applicable to new AXFR
clients.  In the reverse situation, older AXFR client and newer AXFR
server ought to induce the server to operate within the specification
for an older server.

4.1.1 AXFR client TCP

An AXFR client MAY request a connection to an AXFR server for any
reason.  An AXFR client SHOULD close the connection when there is
no apparent need to use the connection for some time period. The
AXFR server ought not have to maintain idle connections, the burden
of connection closure ought to be on the client.  Apparent need for
the connection is a judgment for the AXFR client and the DNS
client. If the connection is used for multiple sessions, or if it is
known sessions will be coming or if there is other query/response
traffic anticipated or currently on the open connection, then there
is "apparent need."

An AXFR client MAY cancel delivery of a zone only by closing the
connection. However, this action will also cancel all other outstanding
activity using the connection. There is no other mechanism by which
an AXFR response can be cancelled.

When a TCP connection is closed remotely (relative to the client),
whether by the AXFR server or due to a network event, the AXFR client
MUST cancel all outstanding sessions and non-AXFR transactions.
Recovery from this situation is not straightforward.  If the disruption
was a spurious event, attempting to restart the connection would be
proper.  If the disruption was caused by a medium or long term
disruption, the AXFR client would be wise to not spend too many
resources trying to rebuild the connection.  Finally, if the connection
was dropped because of a policy at the AXFR server (as can be the case
with older AXFR servers), the AXFR client would be wise to not retry
the connection.  Unfortunately, knowing which of the three cases above
(momentary disruption, failure, policy) applies is not possible with
certainty, and can only be assessed by heuristics.

An AXFR client MAY use an already opened TCP connection to start an
AXFR session.  Using an existing open connection is RECOMMENDED over
opening a new connection.  (Non-AXFR session traffic can also use an
open connection.)  If in doing so the AXFR client realizes that
the responses cannot be properly differentiated (lack of matching
query IDs for example) or the connection is terminated for a remote
reason, then the AXFR client SHOULD NOT attempt to reuse an open
connection with the specific AXFR server until the AXFR server is
updated (which is of course, not an event captured in the DNS
protocol).

4.1.2  AXFR server TCP

An AXFR server MUST be able to handle multiple AXFR sessions on a
single TCP connection, as well as handle other query/response
transactions.

If a TCP connection is closed remotely, the AXFR server MUST cancel
all AXFR sessions in place.  No retry activity is necessary; that is
initiated by the AXFR client.

Local policy MAY dictate that a TCP connection is to be closed.  Such
an action SHOULD be in reaction to limits such as those placed on
the number of outstanding open connections.  Closing a connection in
response to a suspected security event SHOULD be done only in extreme
cases, when the server is certain the action is warranted.  An
isolated request for a zone not on the AXFR server SHOULD receive
a response with the appropriate return code and not see the connection
broken.

4.2 UDP

AXFR sessions over UDP transport are not defined.

 Authorization

A zone administrator has the option to restrict AXFR access to a zone.
This was not envisioned in the original design of the DNS but has
emerged as a requirement as the DNS has evolved.  Restrictions on AXFR
could be for various reasons including a desire (or in some instances,
having a legal requirement) to keep the bulk version of the zone
concealed or to prevent the servers from handling the load incurred in
serving AXFR.  All reasons are arguable, but the fact remains that
there is a requirement to provide mechanisms to restrict AXFR.

A DNS implementation SHOULD provide means to restrict AXFR sessions to
specific clients.

An implementation SHOULD allow access to be granted to Internet
Protocol addresses and ranges, regardless of whether a source address
could be spoofed.  Combining this with techniques such as Virtual
Private Networks (VPN) [RFC2764] or Virtual LANs has proven to be
effective.

A general purpose implementation is RECOMMENDED to implement access
controls based upon "Secret Key Transaction Authentication for DNS"
[RFC2845] and/or "DNS Request and Transaction Signatures ( SIG(0)s )"
[RFC2931].

A general purpose implementation SHOULD allow access to be open to
all AXFR requests.  I.e., an operator ought to be able to allow any
AXFR query to be granted.

A general purpose implementation SHOULD NOT have a default policy
for AXFR requests to be "open to all."  For example, a default could
be to restrict transfers to addresses selected by the DNS
administrator(s) for zones on the server.

 Zone Integrity

An AXFR client MUST ensure that only a successfully transferred
copy of the zone data can be used to serve this zone.  Previous
description and implementation practice have introduced a two-stage
model of the whole zone synchronization procedure: Upon a trigger
event (e.g., polling of SOA resource record detects change in the
SOA serial number, or via DNS NOTIFY [RFC1996]), the AXFR session
is initiated, whereby the zone data are saved in a zone file or
data base (this latter step is necessary anyway to ensure proper
restart of the server); upon successful completion of the AXFR
operation and some sanity checks, this data set is 'loaded' and
made available for serving the zone in an atomic operation, and
flagged 'valid' for use during the next restart of the DNS server;
if any error is detected, this data set MUST be deleted, and the
AXFR client MUST continue to serve the previous version of the zone,
if it did before.  The externally visible behavior of an AXFR client
implementation MUST be equivalent to that of this two-stage model.

If a server rejects data contained in an AXFR session, the server
SHOULD remember the serial number and MAY attempt to retrieve the
same zone version again.  The reason the same retrieval could make
sense is that the reason for the rejection could be rooted in an
implementation detail of one AXFR server used for the zone and not
in another AXFR server used for the zone.

Ensuring that an AXFR client does not accept a forged copy of a zone is
important to the security of a zone.  If a zone operator has the
opportunity, protection can be afforded via dedicated links, physical
or virtual via a VPN among the authoritative servers.  But there are
instances in which zone operators have no choice but to run AXFR
sessions over the global public Internet.

Besides best attempts at securing TCP connections, DNS implementations
SHOULD provide means to make use of "Secret Key Transaction
Authentication for DNS" [RFC2845] and/or "DNS Request and Transaction
Signatures ( SIG(0)s )" [RFC2931] to allow AXFR clients to verify the
contents.  These techniques MAY also be used for authorization.

 Backwards Compatibility

Describing backwards compatibility is difficult because of the lack of
specifics in the original definition.  In this section some hints at
building in backwards compatibility are given, mostly repeated from the
earlier sections.

Backwards compatibility is not necessary, but the greater the extent of
an implementation's compatibility the greater it's interoperability.
For turnkey implementations this is not usually a concern.  For general
purpose implementations this takes on varying levels of importance
depending on the implementer's desire to maintain interoperability.

It is unfortunate that a need to fall back to older behavior cannot be
discovered, hence needs to be noted in a configuration file.  An
implementation SHOULD, in it's documentation, encourage operators to
periodically review AXFR clients and servers it has made notes about as
old software periodically gets updated.

7.1 Server

An AXFR server has the luxury of being able to react to an AXFR
client's abilities with the exception of knowing if the client can
accept multiple resource records per AXFR response message.  The
knowledge that a client is so restricted apparently cannot be
discovered, hence it has to be set by configuration.

An implementation of an AXFR server MAY permit configuring, on a per
AXFR client basis, a need to revert to single resource record per
message; in that case, the default SHOULD be to use multiple records

7.2 Client

An AXFR client has the opportunity to try other features (i.e., those
not defined by this document) when querying an AXFR server.

Attempting to issue multiple DNS queries over a TCP transport for an
AXFR session SHOULD be aborted if it interrupts the original request,
and SHOULD take into consideration whether the AXFR server intends to
close the connection immediately upon completion of the original
(connection-causing) zone transfer.

 Security Considerations

Concerns regarding authorization, traffic flooding, and message
integrity are mentioned in "Authorization" (section 5), "TCP" (section
4.2) and "Zone Integrity" (section 6).

 IANA Considerations

No new registries or new registrations are included in this document.

 Internationalization Considerations

The AXFR protocol is transparent to the parts of DNS zone content that
can possibly be subject to Internationalization considerations.
It is assumed that for DNS labels and domain names, the issue has been
solved via "Internationalizing Domain Names in Applications (IDNA)"
[RFC3490].


 Acknowledgements

Earlier editions of this document have been edited by Andreas
Gustafsson. In his latest version, this acknowledgement appeared.

"Many people have contributed input and commentary to earlier versions
of this document, including but not limited to Bob Halley, Dan
Bernstein, Eric A. Hall, Josh Littlefield, Kevin Darcy, Robert Elz,
Levon Esibov, Mark Andrews, Michael Patton, Peter Koch, Sam Trenholme,
and Brian Wellington."

Comments since the -05 version have come from these individuals:
Alfred Hoenes, Mark Andrews, Paul Vixie, Wouter Wijngaards, Iain
Calder, Tony Finch, Ian Jackson, Andreas Gustafsson, Brian Wellington,
and other participants of the DNSEXT working group.

 References

All references prefixed by "RFC" can be obtained from the RFC Editor
web site at the URLs:   http://rfc-editor.org/rfc.html
or                      http://rfc-editor.org/rfcsearch.html ;
information regarding this organization can be found at the following
URL:                    http://rfc-editor.org/

12.1 Normative

[RFC0793] Postel, J., "Transmission Control Protocol", STD 7, RFC 793,
          September 1981.
[RFC0768] Postel, J., "User Datagram Protocol", STD 6, RFC 768, August
          1980.
[RFC1034] Mockapetris, P., "Domain names - concepts and facilities",
          STD 13, RFC 1034, November 1987.
[RFC1035] Mockapetris, P., "Domain names - implementation and
          specification", STD 13, RFC 1035, November 1987.
[RFC1123] Braden, R., "Requirements for Internet Hosts - Application
          and Support", STD 3, RFC 1123, October 1989.
[RFC1995] Ohta, M., "Incremental Zone Transfer in DNS", RFC 1995,
          August 1996.
[RFC1996] Vixie, P., "A Mechanism for Prompt Notification of Zone
          Changes (DNS NOTIFY)", RFC 1996, August 1996.
[RFC2136] Vixie, P., Ed., Thomson, S., Rekhter, Y., and J. Bound,
          "Dynamic Updates in the Domain Name System (DNS UPDATE)", RFC
          2136, April 1997.
[RFC2181] Elz, R. and R. Bush, "Clarifications to the DNS
          Specification", RFC 2181, July 1997.
[RFC2671] Vixie, P., "Extension Mechanisms for DNS (EDNS0)", RFC 2671,
          August 1999.
[RFC2672] Crawford, M., "Non-Terminal DNS Name Redirection", RFC 2672,
          August 1999.
[RFC2845] Vixie, P., Gudmundsson, O., Eastlake 3rd, D., and B.
          Wellington, "Secret Key Transaction Authentication for DNS
          (TSIG)", RFC 2845, May 2000.
[RFC5395] Eastlake 3rd, "Domain Name System (DNS) IANA Considerations",
          BCP 42, RFC 5395, November 2008.
[RFC2930] Eastlake 3rd, D., "Secret Key Establishment for DNS (TKEY
          RR)", RFC 2930, September 2000.
[RFC2931] Eastlake 3rd, D., "DNS Request and Transaction Signatures
          ( SIG(0)s )", RFC 2931, September 2000.
[RFC3425] Lawrence, D., "Obsoleting IQUERY", RFC 3425, November 2002.
[RFC3597] Gustafsson, A., "Handling of Unknown DNS Resource Record
          (RR) Types", RFC 3597, September 2003.
[RFC4033] Arends, R., Austein, R., Larson, M., Massey, D., and S.
          Rose, "DNS Security Introduction and Requirements", RFC 4033,
          March 2005.
[RFC4034] Arends, R., Austein, R., Larson, M., Massey, D., and S.
          Rose, "Resource Records for the DNS Security Extensions",
          RFC 4034, March 2005.
[RFC4509] Hardaker, W., "Use of SHA-256 in DNSSEC Delegation Signer
          (DS) Resource Records (RRs)", RFC 4509, May 2006
[RFC5155] Laurie, B., Sisson, G., Arends, R., and D. Blacka, "DNS
          Security (DNSSEC) Hashed Authenticated Denial of Existence",
          RFC 5155, March 2008
[RFC4035] Arends, R., Austein, R., Larson, M., Massey, D., and S.
          Rose, "Protocol Modifications for the DNS Security
          Extensions", RFC 4035, March 2005.
[RFC4635] Eastlake 3rd, D., "HMAC SHA (Hashed Message Authentication
          Code, Secure Hash Algorithm) TSIG Algorithm Identifiers",
          RFC 4635, August 2006.
[DNSFLGS] http://www.iana.org/assignments/dns-header-flags
[DNSVALS] http://www.iana.org/assignments/dns-parameters

12.2 Informative

[BCP14]   Bradner, S., "Key words for use in RFCs to Indicate
          Requirement Levels", BCP 14, RFC 2119, March 1997.
[RFC1700] J. Reynolds and J. Postel, "Assigned Numbers", RFC 1700,
          October 1994.
[RFC2764] Gleeson, B., Lin, A., Heinanen, J., Armitage, G., and A.
          Malis, "A Framework for IP Based Virtual Private Networks",
          RFC 2764, February 2000.
[RFC3490] Faltstrom, P., Hoffman, P., and A. Costello,
          "Internationalizing Domain Names in Applications (IDNA)", RFC
          3490, March 2003.
[DRAFT1]  Jansen, J., "Use of SHA-2 algorithms with RSA in DNSKEY and
          RRSIG Resource Records for DNSSEC",
          draft-ietf-dnsext-dnssec-rsasha256-12, work in progress.
[DRAFT2]  Weiler, S., and D. Blacka, "Clarifications and Implementation
          Notes for DNSSECbis",
          draft-ietf-dnsext-dnssec-bis-updates-08, work in progress.

 Editor's Address

Edward Lewis
 Center Oak Plaza
Sterling, VA, 22033, US
+1-571-434-5468
ed.lewis@neustar.biz