DNSEXT Working Group                                        Yuji Kamite
INTERNET-DRAFT                                       NTT Communications
<draft-ietf-dnsext-tkey-renewal-mode-02.txt>
<draft-ietf-dnsext-tkey-renewal-mode-03.txt>            Masaya Nakayama
Expires: Mar. 24, Sep. 3, 2003                           The University of Tokyo
                                                          Sep. 24, 2002
                                                           Mar. 3, 2003

                      TKEY Secret Key Renewal Mode

Status of this Memo

  This document is an Internet-Draft and is in full conformance with all
  provisions of Section 10 of RFC2026.

  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/ietf/1id-abstracts.txt

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

Abstract

   This document defines a new mode in TKEY [RFC2930] and proposes an
   atomic method for changing secret keys used for TSIG [RFC2845]
   periodically. Originally, TKEY provides methods of setting up shared
   secrets other than manual exchange, but it cannot control timing of
   key renewal very well though it can add or delete shared keys
   separately. This proposal is a systematical key renewal procedure
   intended for preventing signing DNS messages with old and non-safe
   keys permanently.

                           Table of Contents

1  Introduction  . . . . . . . . . . . . . . . . . . . . . . . . . .   3
  1.1  Defined Words . . . . . . . . . . . . . . . . . . . . . . . .   3
  1.2  New Format and Assigned Numbers . . . . . . . . . . . . . . .   4
  1.3  Overview of Secret Key Renewal Mode . . . . . . . . . . . . .   4
2  Shared Secret Key Renewal . . . . . . . . . . . . . . . . . . . .   5
  2.1  Key Usage Time Check  . . . . . . . . . . . . . . . . . . . .   5
  2.2  Partial Revocation  . . . . . . . . . . . . . . . . . . . . .   6
  2.3  Key Renewal Message Exchange  . . . . . . . . . . . . . . . .   6
    2.3.1  TKEY RR structure for Key Renewal . . . . . . . . . . . .   8
  2.4  Key Adoption  . . . . . . . . . . . . . . . . . . . . . . . .   9
    2.4.1  Query for Key Adoption  . . . . . . . . . . . . . . . . .   9
    2.4.2  Response for Key Adoption . . . . . . . . . . . . . . . .  10
  2.5  Keying Schemes  . . . . . . . . . . . . . . . . . . . . . . .  10  11
    2.5.1  DH Exchange for Key Renewal . . . . . . . . . . . . . . .  11
    2.5.2  Server Assigned Keying for Key Renewal  . . . . . . . . .  12
    2.5.3  Resolver Assigned Keying for Key Renewal  . . . . . . . .  13
  2.6  Considerations about Non-compliant Hosts  . . . . . . . . . .  14
3  Secret Storage  . . . . . . . . . . . . . . . . . . . . . . . . .  14
4  Compulsory Key Revocation by Server . . . . . . . . . . . . . . .  14  15
  4.1  Example . . . . . . . . . . . . . . . . . . . . . . . . . . .  15
5  Special Considerations for Two Servers' Case  . . . . . . . . . .  15  16
  5.1  To Cope with Collisions of Renewal Requests . . . . . . . . .  16
6  Key Name Considerations . . . . . . . . . . . . . . . . . . . . .  16  17
7  Example Usage of Secret Key Renewal Mode  . . . . . . . . . . . .  17
8  Security Considerations . . . . . . . . . . . . . . . . . . . . .  19
9  IANA Considerations . . . . . . . . . . . . . . . . . . . . . . .  20
10 References  . . . . . . . . . . . . . . . . . . . . . . . . . . .  20
Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . . .  21  22

1.  Introduction

   TSIG [RFC2845] provides DNS message integrity and the
   request/transaction authentication by means of message authentication
   codes (MAC). TSIG is a practical solution in view of calculation
   speed and availability. However, TSIG does not have exchanging
   mechanism of shared secret keys between server and resolver, and
   administrators might have to exchange secret keys manually.  TKEY
   [RFC2930] is introduced to solve such problem and it can exchange
   secrets for TSIG via networks.

   In various modes of TKEY, a server and a resolver can add or delete a
   secret key be means of TKEY message exchange. However, the existing
   TKEY does not care fully about the management of keys which became
   too old, or dangerous after long time usage.

   It is ideal that the number of secret which a pair of hosts share
   should be limited only one, because having too many keys for the same
   purpose might not only be a burden to resolvers for managing and
   distinguishing according to servers to query, but also does not seem
   to be safe in terms of storage and protection against attackers.
   Moreover, perhaps holding old keys long time might give attackers
   chances to compromise by scrupulous calculation.

   Therefore, when a new shared secret is established by TKEY, the
   previous old secret should be revoked immediately. To accomplish
   this, DNS servers must support a protocol for key renewal. This
   document specifies procedure to refresh secret keys between two hosts
   which is defined within the framework of TKEY, and it is called "TKEY
   Secret Key Renewal Mode".

   The key words "MUST", "MUST NOT", "SHOULD", "SHOULD NOT" NOT", "MAY" and
   "OPTIONAL" in this document are to be interpreted as described in
   [RFC 2119].
   [RFC2119].

1.1.  Defined Words

   * Inception Time: Beginning of the shared secret key lifetime. This
   value is determined when the key is generated.

   * Expiry Limit: Time limit of the key's validity. This value is
   determined when a new key is generated. After Expiry Limit, server
   and client (resolver) must not authenticate TSIG signed with the key.
   Therefore, Renewal to the next key should be carried out before
   Expiry Limit.

   * Partial Revocation Time: Time when server judges the key is too old
   and must be updated. It must be between Inception Time and Expiry
   Limit.  This value is determined by server freely following its
   security policy. e.g., If the time from Inception to Partial
   Revocation is short, renewal will be carried out more often, which
   might be safer.

   * Revocation Time: Time when the key becomes invalid and can be
   removed. This value is not determined in advance because it is the
   actual time when revocation is completed.

   * Adoption Time: Time when the new key is adopted as the next key
   formally. After Adoption, the key is valid and server and client can
   generate or verify TSIG making use of it. Adoption Time also means
   the time when it becomes possible to remove the previous key, so
   Revocation and Adoption are usually done at the same time.

                  Partial
    Inception     Revocation    Revocation         Expiry Limit
     |                |              |             |
     |----------------|- - - - - - >>|- (revoked) -|
     |                |              |             |
    previous key      |      |       |
                             |- - - -|-------------------->> time
                             |       |               new key
                         Inception   Adoption

1.2.  New Format and Assigned Numbers

   TSIG
         ERROR = (PartialRevoke), to be defined

   TKEY
         Mode  = (server assignment for key renewal), to be defined
         Mode  = (Diffie-Hellman exchange for key renewal), to be
   defined
         Mode  = (resolver assignment for key renewal), to be defined
         Mode  = (key adoption), to be defined

1.3.  Overview of Secret Key Renewal Mode

   When a server receives a query from a client signed with a TSIG key,
   It always checks if the present time is within the range of usage
   duration it considers safe. If it is judged that the key is too old,
   i.e., after Partial Revocation Time, the server comes to be in
   Partial Revocation state about the key. key, and this key is called
   partially revoked.

   In this state, whenever a client sends a normal query (e.g., question
   about A RR) other than TKEY Renewal request with TSIG signed with the
   old key, the server returns an error message to notify that the time
   to renew has come. This is called "PartialRevoke" error message.

   The client which got this error is able to notice that it is
   necessary to refresh the secret. To make a new shared secret, it
   sends a TKEY Renewal request, in which several keying methods are
   available. It can make use of TSIG authentication signed with the
   partially revoked key mentioned above.

   After new secret establishment, the client sends a TKEY Adoption request.
   request for renewal confirmation. This can also be authenticated with
   the partially revoked key. If this is admitted by the server, the new
   key is formally adopted, and at the same time the corresponding old
   secret is invalidated. Then the client can send the first query again
   signed with the new key.

   Key renewal procedure is executed based on two-phase commit
   mechanism. The first phase is the TKEY Renewal request and its
   response, which means preparatory confirmation for key update. The
   second phase is Adoption request and its response. If the server gets
   request and client receives the response successfully, they can
   finish renewal process. If any error happens and renewal process
   fails during these phases, client should roll back to the beginning
   of the first phase, and send TKEY Renewal request again. This
   rollback can be done until the Expiry Limit of the key.

2.  Shared Secret Key Renewal

   Suppose a server and a client agree to change their TSIG keys
   periodically. Key renewal procedure is defined between two hosts.

2.1.  Key Usage Time Check

   Whenever a server receives a query with TSIG and can find a key that
   is used for signing it, the server checks its Inception Time, Partial
   Revocation Time and Expiry Limit (this information is usually
   memorized by the server).

   When the present time is before Inception Time, the server MUST NOT
   verify TSIG with the key, and server acts the same way as when no
   valid key is found, following [RFC2845].

   When the present time is equal to Inception Time, or between
   Inception Time and Partial Revocation Time, the behavior of the
   server is the same as when a valid key is found, required in
   [RFC2845].

   When the present time is the same as the Partial Revocation Time, or
   between the Partial Revocation Time and Expiry Limit, the server
   comes to be in Partial Revocation state about the TSIG key and
   behaves according to the next section.

   When the present time is the same as the Expiry Time or after it, the
   server MUST NOT verify TSIG with the key, and returns error messages
   in the same way as when no valid key is found, following [RFC2845].

2.2.  Partial Revocation

   In Partial Revocation state, we say the server has partially revoked
   the key and the key has become a "partially revoked key".

   Server which

   If server has received queries a query signed with the partially revoked key MUST NOT answer them except when they are
   for TKEY Renewal requests request (See section 2.3.) or "key adoption" requests request
   (See section 2.4.). Instead,
   it returns 2.4.), then server does proper process following each
   specification.

   If server receives other types of query signed with the partially
   revoked key and the corresponding MAC is verified, then server SHOULD
   return TSIG error messages message for response whose error codes are code is
   "PartialRevoke" (See section 9.) 9.). However, if it is for TKEY key
   deletion request ([RFC2930] 4.2), server MAY process usual deletion
   operation defined therein.

   PartialRevoke error messages have the role to inform clients of the
   keys' partial revocation and urge them to send TKEY Renewal requests.
   These error responses MUST be signed with those partial revoked keys
   if the queries are signed with them. They are sent only when the keys
   used for queries' TSIG are found to be partially revoked. If the MAC
   of TSIG cannot be verified with the partially revoked keys, servers
   MUST NOT return PartialRevoke error with MAC, but should return
   another error such as "BADSIG" without MAC; in other words, a server
   informs its key's partial revocation only when the MAC in the
   received query is valid.

2.3.  Key Renewal Message Exchange

   If a client has received a PartialRevoke error and authenticated the
   response based on TSIG MAC, it sends a TKEY query for Key Renewal (in
   this document, we call it Renewal request, too.) to the server. The
   request MUST be signed with TSIG or SIG(0) [RFC2931] for
   authentication. If TSIG is selected, the client can sign it with the
   partial revoked key.

   Key Renewal can use one of several keying methods which is indicated
   in "Mode" field of TKEY RR, and its message structure is dependent on
   that method.

   The server which has received Key Renewal request first tries to
   verify TSIG or SIG(0) accompanying it. If the TSIG is signed and
   verified with the partially revoked key, the request MUST be
   authenticated.

   After authentication, server must check existing old key's validity.
   If the partially revoked key indicated in the request TKEY's OldName
   and OldAlgorithm field (See section 2.3.1.) does not really exist at
   the server, "BADKEY" [RFC 2845] [RFC2845] is given in Error field for response.
   If any other error happens, server returns appropriate error messages
   following the specification described in section 2.5. If there are no
   errors, server returns a Key Renewal answer. This answer MUST be
   signed with TSIG or SIG(0) for authentication.

   When this answer is successfully returned and no error is detected by
   client, a new shared secret can be established. The details of
   concrete keying procedure are given in the section 2.5.

   As a result of this message exchange, client comes to know the newly
   generated key's attributes such as key's name, Inception Time and
   Expiry Limit. They are decided finally by the server, and are told to
   the client; in particular, however, once the server has decided
   Expiry Limit and returned a response, it should obey the decision as
   far as it can. In other words, they SHOULD NOT change time values for
   checking Expiry Limit in the future without any special reason, such
   as security issue like "Emergency Compulsory Revocation" described in
   section 8.

   On the other hand, Partial Revocation Time of this generated key is
   not decided based on the request, and not informed to the client. The
   server can determine any value as long as it is between Inception
   Time and Expiry Limit. However, it is recommended that the period
   from Inception to Partial Revocation should be fixed as the server
   side's configuration or should be set the same as the corresponding
   old key's one.

   Note:
      Even after the response is returned to client, possibly server
      sometimes receives another Renewal TKEY request whose OldName
      indicates the same partial revoked key. Mostly such messages
      originate in client's resending or rollback because of some kinds
      of errors. In this case, the server processes keying again and
      MUST replace the associated secret with the newest produced
      secret. The secret key produced before comes to be invalid and
      should be removed from memory; this process is called secret
      overwrite. Moreover, This regenerated key's name MUST always be
      different from the one which it overwrites for secret key's
      uniqueness and distinction. See section 6, too.

      Even if client sends Key Renewal request though the key described
      in OldName has not been partially revoked yet, server must do
      renewal processes.  But at the moment when the server accepts such
      requests with valid authentication, it MUST forcibly consider the
      key is already partially revoked, that is, the key's Partial
      Revocation Time must be changed into the present time (i.e., the
      time when the server receives the request).

2.3.1.  TKEY RR structure for Key Renewal

   TKEY RR for Key Renewal message has the structure given below. In
   principle, format and definition for each field follows [RFC2930].
   Note that each keying scheme sometimes needs different interpretation
   of RDATA field; for detail, see section 2.5.

        Field        Type         Comment
        -------      ------       -------
        NAME         domain       used for a new key, see below
        TYPE         u_int16_t    (defined in [RFC2930])
        CLASS        u_int16_t    (defined in [RFC2930])
        TTL          u_int32_t    (defined in [RFC2930])
        RDLEN        u_int16_t    (defined in [RFC2930])
        RDATA:
         Algorithm:   domain       algorithm for a new key
         Inception:   u_int32_t    about the keying material
         Expiration:  u_int32_t    about the keying material
         Mode:        u_int16_t    scheme for key agreement
                                   see section 9.
         Error:       u_int16_t    see description below
         Key Size:    u_int16_t    see description below
         Key Data:    octet-stream
         Other Size:  u_int16_t    (defined in [RFC2930])
                                   size of other data
         Other Data:               newly defined: see description below
     For "NAME" field, both non-root and root name are allowed. It may
     be used for a new key's name in the same manner as [RFC2930] 2.1.

     "Algorithm" specifies which algorithm is used for agreed keying
     material, which is used for identification of the next key.

     "Inception" and "Expiration" are used for the valid period of
     keying material. The meanings differ somewhat according to whether
     the message is request or answer, and its keying scheme.

     "Key Data" has different meanings according to keying schemes.

     "Mode" field stores the value in accordance with the keying method,
     and see section 2.5. Servers and clients supporting TKEY Renewal
     method MUST implement "Diffie-Hellman exchange for key renewal"
     scheme. All other modes are OPTIONAL.

     "Error" is an extended RCODE which includes "PartialRevoke" value
     too. See section 9.

     "Other Data" field has the structure given below.  They describe
     attributes of the key to be renewed.

        in Other Data filed:

          Field          Type     Comment
          -------        ------   -------
          OldNAME        domain   name of the old key
          OldAlgorithm   domain   algorithm of the old key

          "OldName" indicates the name of the previous key (usually,
          this is partially revoked key's name that client noticed by
          PartialRevoke answer from server), and "OldAlogirthm"
          indicates its algorithm.

2.4.  Key Adoption

2.4.1.  Query for Key Adoption

   After receiving a TKEY Renewal answer, the client gets the same
   secret as the server. Then, it sends a TKEY Adoption request. The
   request's question section's QNAME field is the same as the NAME
   filed of TKEY written below. In additional section, there is one TKEY
   RR that has the structure and values described below.

     "NAME" field is the new key's name to be adopted which was already
     generated by Renewal message exchange. "Algorithm" is its algorithm
     name. algo-
     rithm. "Inception" means the key's Inception Time, and "Expiration"
     means Expiry Limit.

     "Mode" field is the value of "key adoption". See section 9.

     "Other Data" field in Adoption has the same structure as that of
     Renewal request message. "OldName" means the previous old key, and
     "OldAlogirthm" means its algorithm.

   Key Adoption request MUST be signed with TSIG or SIG(0) for
   authentication. The client can sign TSIG with the previous key. Note
   that until Adoption is finished, the new key is treated as invalid,
   thus it cannot be used for authentication immediately.

2.4.2.  Response for Key Adoption

   The server which has received Adoption request, it verifies TSIG or
   SIG(0) accompanying it. If the TSIG is signed with the partially
   revoked key and can be verified, the message MUST be authenticated.

   If the next new key indicated by the request TKEY's "NAME" is not
   really present at the server, BADNAME [RFC 2845] [RFC2845] is given in Error
   field and the error message is returned.

   If the next key really exists and it has not been adopted formally
   yet, the server confirms the previous key's existence indicated by
   the "OldName" and "OldAlgorithm" field. If it succeeds, the server
   executes Adoption of the next key and Revocation of the previous key.
   Response message duplicates the request's TKEY RR with NOERROR,
   including "OldName" and "OldAlgorithm" that indicate the revoked key.

   If the next key exists but it is already adopted, the server returns
   a response message regardless of the substance of the request TKEY's
   "OldName". In this response, Response TKEY RR has the same data as
   the request's one except as to its "Other Data" that is changed into
   null (i.e., "Other Size" is zero), which is intended for telling the
   client that the previous key name was ignored, and the new key is
   already available.

   Client sometimes has to retry Adoption request. Suppose the client
   sent request signed with the partially revoked key, but its response
   did not return successfully (e.g., due to the drop of UDP packet).
   Client will probably retry Adoption request; however, the request
   will be refused in the form of TSIG "BADKEY" error because the
   previous key was already revoked. In this case, client should
   retransmit Adoption request signed with the next key, and expect a
   response which has null "Other Data" for confirming the completion of
   renewal.

2.5.  Keying Schemes

   In Renewal message exchanges, there are no limitations as to which
   keying method is actually used. The specification of keying
   algorithms is independent of the general procedure of Renewal that is
   described in section 2.3.

   Now this document specifies three algorithms in this section, but
   other future documents can make extensions defining other methods.

2.5.1.  DH Exchange for Key Renewal

   This scheme is defined as an extended method of [RFC2930] 4.1. This
   specification only describes the difference from it and special
   notice; assume that all other points, such as keying material
   computation, are the exactly same as the specification of [RFC2930]
   4.1.

   Query
      In Renewal request for type TKEY with this mode, there is one TKEY
      RR and one KEY RR in the additional information section. KEY RR is
      the client's Diffie-Hellman public key [RFC2539].

      QNAME in question section is the same as that of "NAME" field in
      TKEY RR, i.e., it means the requested new key's name.

      TKEY "Mode" field stores the value of "DH exchange for key
      renewal". See section 9.

      TKEY "Inception" and "Expiration" are those requested for the
      keying material, that is, requested usage period of a new key.

      TKEY "Key Data" is used as a random, following [RFC2930] 4.1.

   Response
      The server which received this request first verifies the TSIG or
      SIG(0). After authentication, the old key's existence validity is
      checked, following section 2.3. If any incompatible DH key is
      found in the request, "BADKEY" [RFC 2845] [RFC2845] is given in Error field
      for response. "FORMERR" is given if the query included no DH KEY.

      If there are no errors, the server processes a response according
      to Diffie-Hellman algorithm and returns the answer. In this
      answer, there is one TKEY RR in answer section and KEY RR(s) in
      additional section.

      As long as no error has occurred, all values of TKEY are equal to
      that of the request message except TKEY NAME, TKEY RDLEN, RDATA's
      Inception, Expiration, Key Size and Key Data.

      TKEY "NAME" field in the answer specifies the name of newly
      produced key which the client will have to use.

      TKEY "Inception" and "Expiration" mean the periods of the produced
      key usage. "Inception" should be set to be the time when the new
      key is actually generated or the time before it, and it will be
      regarded as Inception Time. "Expiration" is determined by the
      server, and it will be regarded as Expiry Limit.

      TKEY "Key Data" is used as an additional nonce, following
      [RFC2930] 4.1.

      The resolver supplied Diffie-Hellman KEY RR SHOULD be echoed in
      the additional section and a server Diffie-Hellman KEY RR will
      also be present in the answer section, following [RFC2930] 4.1.

2.5.2.  Server Assigned Keying for Key Renewal

   This scheme is defined as an extended method of [RFC2930] 4.4. This
   specification only describes the difference from it and special
   notice; assume that all other points, such as secret encrypting
   method, are the exactly same as the specification of [RFC2930] 4.4.

   Query
      In Renewal request for type TKEY with this mode, there is one TKEY
      RR and one KEY RR in the additional information section. KEY RR is
      used in encrypting the response.

      QNAME in question section is the same as that of "NAME" field in
      TKEY RR, i.e., it means the requested new key's name.

      TKEY "Mode" field stores the value of "server assignment for key
      renewal". See section 9.

      TKEY "Inception" and "Expiration" are those requested for the
      keying material, that is, requested usage period of a new key.

      TKEY "Key Data" is provided following the specification of
      [RFC2930] 4.4.

   Response
      The server which received this request first verifies the TSIG or
      SIG(0). Resolver KEY RR is also authenticated by means of
      verifying that TSIG or SIG(0). After authentication, the old key's
      existence validity is checked, following section 2.3. "FORMERR" is
      given if the query specified no encryption key.

      If there are no errors, the server response contains one TKEY RR
      in the answer section, and echoes the KEY RR provided in the query
      in the additional information section.

      TKEY "NAME" field in the answer specifies the name of newly
      produced key which the client will have to use.

      TKEY "Inception" and "Expiration" mean the periods of the produced
      key usage. "Inception" should be set to be the time when the new
      key is actually generated or the time before it, and it will be
      regarded as Inception Time. "Expiration" is determined by the
      server, and it will be regarded as Expiry Limit.

      TKEY "Key Data" is the assigned keying data encrypted under the
      public key in the resolver provided KEY RR, which is the same as
      [RFC2930] 4.4.

2.5.3.  Resolver Assigned Keying for Key Renewal

   This scheme is defined as an extended method of [RFC2930] 4.5. This
   specification only describes the difference from it and special
   notice; assume that all other points, such as secret encrypting
   method, are the exactly same as the specification of [RFC2930] 4.5.

   Query
      In Renewal request for type TKEY with this mode, there is one TKEY
      RR and one KEY RR in the additional information section. TKEY RR
      has the encrypted keying material and KEY RR is the server public
      key used to encrypt the data.

      QNAME in question section is the same as that of "NAME" field in
      TKEY RR, i.e., it means the requested new key's name.

      TKEY "Mode" field stores the value of "resolver assignment for key
      renewal". See section 9.

      TKEY "Inception" and "Expiration" are those requested for the
      keying material, that is, requested usage period of a new key.

      TKEY "Key Data" is the encrypted keying material.

   Response
      The server which received this request first verifies the TSIG or
      SIG(0). After authentication, the old key's existence validity is
      checked, following section 2.3. "FORMERR" is given if the server
      does not have the corresponding private key for the KEY RR that
      was shown in the request.

      If there are no errors, the server returns response. The response
      must have a TKEY RR in the answer section to tell the shared key's
      name and its usage time values.

      TKEY "NAME" field in the answer specifies the name of newly
      produced key which the client will have to use.

      TKEY "Inception" and "Expiration" mean the periods of the produced
      key usage. "Inception" should be set to be the time when the new
      key is actually generated or the time before it, and it will be
      regarded as Inception Time. "Expiration" is determined by the
      server, and it will be regarded as Expiry Limit.

2.6.  Considerations about Non-compliant Hosts

   Key Renewal requests and responses must be exchanged between hosts
   which can understand them and do proper processes. "PartialRevoke"
   error messages will be only ignored if they should be returned to
   non-compliant hosts.

   Note that server does not inform actively the necessity of renewal to
   clients, but inform it as responses invoked by client's query.
   Server needs not care whether the "PartialRevoke" errors has reached
   client or not. If client has not received yet because of any reasons
   such as packet drops, it will resend the queries, and finally will be
   able to get "PartialRevoke" information.

3.  Secret Storage

   Every server should keep all secrets and attached information, e.g.,
   Inception Time, Expiry Limit, etc. safely to be able to recover from
   unexpected stop. To accomplish this, formally adopted keys should be
   memorized not only on memory, but also be stored in the form of some
   files. Make sure that this should be protected strongly not to be
   read by others. If possible, they should be stored in encrypted form.

4.  Compulsory Key Revocation by Server

   There is a rare but possible case that although servers have already
   partially revoked keys, clients do not try to send any Renewal
   requests. If this state continues, in the future it will become the
   time of Expiry Limit. After Expiry Limit, the keys will be expired
   and completely removed, so this is called Compulsory Key Revocation
   by server.

   If Expiry Limit is too distant from the Partial Revocation Time, then
   even though very long time passes, clients will be able to refresh
   secrets only if they add TSIG signed with those old partially revoked
   keys into requests, which is not safe.

   On the other hand, if Expiry Limit is too close to Partial Revocation
   Time, perhaps clients might not be able to notice their keys' Partial
   Revocation by getting "PartialRevoke" errors.

   Therefore, servers should set proper Expiry Limit to their keys,
   considering both their keys' safety, and enough time for clients to
   send requests and process renewal.

4.1.  Example

   It might be ideal to provide both SIG(0) and TSIG as authentication
   methods. For example:

     A client and a server start SIG(0) authentication at first, to
     establish TSIG shared keys by means of "Query for Diffie-Hellman
     Exchanged Keying" as described in [RFC 2930] [RFC2930] 4.1. Once they get
     shared secret, they keep using TSIG for usual queries and
     responses. After a while the server returns a "ParitalRevoke" error
     and they begin a key renewal process. Both TSIG signed with
     partially revoked keys and SIG(0) are okay for authentication, but
     TSIG would be easier to use considering calculation efficiency.

     If any troubles should happen such as client host's long suspension
     against expectation, the server will do Compulsory Revocation.
     After recovery if the client sends a key Renewal request to refresh
     the old key, it will fail because the key is removed from the
     server. So, the client will send "Query for Diffie-Hellman
     Exchanged Keying" with SIG(0) to make a new shared key again.

5.  Special Considerations for Two servers' Case

   This section refers to the case where both two hosts are DNS servers
   which can act as full resolvers as well. If one server (called Server
   A) comes to want to refresh a shared key (called "Key A-B"), it will
   await a TKEY Renewal request from the other server (called Server B).
   But perhaps Server A will have to send queries with TSIG immediately
   to Server B to resolve some queries if it is asked by other clients.

   At this time, Server A is allowed to send a Renewal request to Server
   B, if Server A finds the key to use now is too old and should be
   renewed. To provide this function, both servers should be able to
   receive and process key renewal request from each other if they agree
   to refresh their shared secret keys.

   Note that the initiative in key renewal belongs to Server A because
   it can notice the Partial Revocation Time and decide key renewal. If
   Server B has information about Partial Revocation Time as well, it
   can also decide for itself to send Renewal request to Server A. But
   it is not essential for both two servers have information about key
   renewal timing.

5.1.  To Cope with Collisions of Renewal Requests

   At least one of two hosts which use Key Renewal must know their key
   renewal information such as Partial Revocation Time. It is okay that
   both hosts have it.

   Provided that both two servers know key renewal timing information,
   there is possibility for them to begin partial revocation and sending
   Renewal requests to each other at the same time. Such collisions will
   not happen so often because Renewal requests are usually invoked when
   hosts want to send queries, but it is possible.

   When one of two servers try to send Renewal requests, it must protect
   old secrets that it has partially revoked and prevent it from being
   refreshed by any requests from the other server (i.e., it must lock
   the old secret during the process of renewal). While the server is
   sending Renewal requests and waiting responses, it ignores the other
   server's Renewal requests.

   Therefore, servers might fail to change secrets by means of their own
   requests to others. After failure they will try to resend, but they
   should wait for random delays by the next retries. If they get any
   Renewal requests from others while they are waiting, their shared
   keys may be refreshed, then they do not need to send any Renewal
   requests now for themselves.

6.  Key Name Considerations

   Since both servers and clients have only to distinguish new secrets
   and old ones, keys' names do not need to be specified strictly. But
   it is recommended that some serial number or key generation time
   should be added to the name and that the names of keys between the
   same pair of hosts should have some common labels among their keys.
   For example, suppose A.example.com. and B.example.com. share the key
   "<serial number>.A.example.com.B.example.com." such as
   "10010.A.example.com.B.example.com.". After key renewal, they change
   their secret and name into "10011.A.example.com.B.example.com." If
   secret overwrite occurs as a result of client's retransmission,
   server must change the next key's name somehow; for example, server
   increases serial number forcibly, or add some random numbers to the
   name.

   Servers and clients must be able to use keys properly according to
   servers to query. Because TSIG secret keys themselves do not have any
   particular IDs to be distinguished and would be identified by their
   names and algorithm, it must be understood correctly what keys are
   refreshed.

7.  Example Usage of Secret Key Renewal Mode

   This is an example of Renewal mode usage where a Server,
   server.example.com, and a Client, client.exmple.com have an initial
   shared secret key named "00.client.example.com.server.example.com".

     (1) The time values about key
     "00.client.example.com.server.example.com" was set as follows:
     Inception Time is at 6:00, Expiry Limit is at 11:00.

     (2) At Server, a time value about renewal timing has been set too:
     Partial Revocation Time is at 8:00.

     (3) Suppose the present time is 7:00. If Client sends a query
     signed with key "00.client.example.com.server.example.com" to ask
     the IP address of "www.somedomain.com", finally it will get a
     proper answer from Server with valid TSIG (NOERROR).

     (4) At 9:00. Client sends a query signed with key
     "00.client.example.com.server.example.com" to ask the IP address of
     "www.otherdomain.com". But it will not get a proper answer from
     Server. The response does not have any IP address information about
     "www.otherdomain.com". Instead, it includes valid TSIG MAC signed
     with "00.client.example.com.server.example.com", and its Error Code
     indicates PartialRevoke.

     (5) At 9:01. Client sends a Renewal request to Server. This request
     is signed with key "00.client.example.com.server.example.com". It
     includes data such as:

      Question Section:
         QNAME = 01.client.example.com. (Client can set this freely)
         TYPE  = TKEY

      Additional Section:
         01.client.example.com. TKEY
          Algorithm    = hmac-md5-sig-alg.reg.int.
          Inception    = (value which means 8:55)
          Expiration   = (value which means 14:00)
          Mode         = (DH exchange for key renewal)
          OldName      = 00.client.example.com.server.example.com.
          OldAlgorithm = hmac-md5-sig-alg.reg.int.

      Additional Section also contains a KEY RR for DH and a TSIG RR.

     (6) As soon as Server receives this request, it verifies TSIG. It
     is signed with the partially revoked key
     "00.client.example.com.server.example.com". and Server accepts the
     request.  It creates a new key by Diffie-Hellman calculation and
     returns an answer which includes data such as:

      Answer Section:
         01.client.example.com.server.example.com. TKEY
          Algorithm    = hmac-md5-sig-alg.reg.int.
          Inception    = (value meaning 8:55)
          Expiration   = (value meaning 14:00)
          Mode         = (DH exchange for key renewal)
          OldName      = 00.client.example.com.server.example.com.
          OldAlgorithm = hmac-md5-sig-alg.reg.int.
      Answer Section also contains KEY RRs for DH.

      Additional Section also contains a TSIG RR.
     This response is signed with key
     "00.client.example.com.server.example.com" without error.

     At the same time, Server decides to set the Partial Revocation Time
     of this new key "01.client.example.com.server.example.com." as
     11:00.

     (7) Client gets the response and checks TSIG MAC, and calculates
     Diffie-Hellman. It will get a new key, and it has been named
     "01.client.example.com.server.example.com" by Server.

     (8) At 9:02. Client sends an Adoption request to Server. This
     request is signed with the previous key
     "00.client.example.com.server.example.com". It includes:

      Question Section:
         QNAME = 01.client.example.com.server.example.com.
         TYPE  = TKEY

      Additional Section:
         01.client.example.com.server.example.com. TKEY
          Algorithm    = hmac-md5-sig-alg.reg.int.
          Inception    = (value which means 8:55)
          Expiration   = (value which means 14:00)
          Mode         = (key adoption)
          OldName      = 00.client.example.com.server.example.com.
          OldAlgorithm = hmac-md5-sig-alg.reg.int.

     Additional Section also contains a TSIG RR.

     (9) Server verifies the query's TSIG. It is signed with the
     previous key and authenticated. It returns a response whose TKEY RR
     is the same as the request's one. The response is signed with key
     "00.client.example.com.server.example.com.". As soon as the
     response is sent, Server revokes and removes the previous key. At
     the same time, key "01.client.example.com.server.example.com." is
     validated.

     (10) Client acknowledges the success of Adoption by receiving the
     response.  Then, it will retry to send an original question about
     "www.otherdomain.com". It is signed with the adopted key
     "01.client.example.com.server.example.com", so Server authenticates
     it and returns an answer.

     (11) This key is used until 11:00. After that, it will be partially
     revoked again.

8.  Security Considerations

   This document considers about how to refresh shared secret. Secret
   changed by this method is used at servers in support of TSIG
   [RFC2845].

   [RFC2104] says that current attacks to HMAC do not indicate a
   specific recommended frequency for key changes but periodic key
   refreshment is a fundamental security practice that helps against
   potential weaknesses of the function and keys, and limits the damage
   of an exposed key. TKEY Secret Key Renewal provides the method of
   periodical key refreshment.

   TKEY Secret Key Renewal forbids clients to send queries as long as
   they do not change old keys. This means that when keys become old,
   clients must spend rather lots of time to get answers they wanted
   originally because at first they must send key Renewal requests. Thus
   the usage period of secrets should be considered carefully based on
   both TKEY processing performance and security.

   This document specifies the procedure of periodical key renewal, but
   actually there is possibility for servers to have no choice other
   than revoking their secret keys immediately especially when the keys
   are found to be compromised by attackers. This is called "Emergency
   Compulsory Revocation". For example, suppose the original Expiry
   Limit was set at 15:00, Partial Revocation Time at 12:00 and
   Inception Time at 10:00.  if at 11:00 the key is found to be
   compromised, the server sets Expiry Limit forcibly to be 11:00 or
   before it.

   Consequently, once Compulsory Revocation (See section 4.) is carried
   out, normal renewal process described in this document cannot be done
   any more as far as the key is concerned. But, after such accidents
   happened, the two hosts are able to establish secret keys and begin
   renewal procedure only if they have other (non-compromised) shared
   TSIG keys or safe SIG(0) keys for the authentication of initial
   secret establishment such as Diffie-Hellman Exchanged Keying.

9.  IANA Considerations

   IANA needs to allocate a value for "DH exchange for key renewal",
   "server assignment for key renewal", "resolver assignment for key
   renewal" and "key adoption" in the mode filed of TKEY. It also needs
   to allocate a value for "PartialRevoke" from the extended RCODE
   space.

10.  References

[RFC2104]
     H. Krawczyk, M.Bellare, R. Canetti, "Keyed-Hashing for Message
     Authentication", RFC2104, February 1997.

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

[RFC2539]
     D. Eastlake 3rd, "Storage of Diffie-Hellman Keys in the Domain Name
     System (DNS)", RFC 2539, March 1999.

[RFC2845]
     Vixie, P., Gudmundsson, O., Eastlake, D. and B.  Wellington,
     "Secret Key Transaction Authentication for DNS (TSIG)", RFC 2845,
     May 2000.

[RFC2930]
     D. Eastlake 3rd, ``Secret Key Establishment for DNS (TKEY RR)'',
     RFC 2930, September 2000.

[RFC2931]
     D. Eastlake 3rd, "DNS Request and Transaction Signatures (SIG(0)s
     )", RFC 2931, September 2000.

Authors' Addresses

   Yuji Kamite
   NTT Communications Corporation
   Innovative IP Architecture Center,
   Tokyo Opera City Tower 21F,
   20-2, 3-chome, Nishi-Shinjuku, Shinjuku-ku,
   Tokyo, 163-1421, Japan.
   EMail: y.kamite@ntt.com

   Masaya Nakayama
   The University of Tokyo
   Information Technology Center,
   2-11-16 Yayoi, Bunkyo-ku,
   Tokyo, 113-8658, Japan.
   EMail: nakayama@nc.u-tokyo.ac.jp