| Network Working Group | Wilkinson |
| Internet-Draft | YFS |
| Intended status: Informational | March 2013 |
| Expires: September 2, 2013 |
rxgk: GSSAPI based security class for RX
draft-wilkinson-afs3-rxgk-03
rxgk is a security class for the RX RPC protocol. It uses the GSSAPI framework to provide authentication, confidentiality and integrity protection. This document provides a general description of rxgk. A further document will provide details of integration with specific RX applications.
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at http://datatracker.ietf.org/drafts/current/.
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."
This Internet-Draft will expire on September 2, 2013.
Copyright (c) 2013 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.
rxgk is a GSSAPI [RFC2743] based security class for the rx [RX] protocol. It provides authentication, confidentiality and integrity protection for rx RPC calls, using a security context established using any GSSAPI mechanism with PRF [RFC4401] support. The External Data Representation Standardard, XDR [RFC4506], is used to represent data structures on the wire and in the code fragments contained within this document.
Architecturally, rxgk is split into two parts. The rxgk rx security class provides strong encryption using previously negotiated ciphers and keys. It builds on the Kerberos crypto framework for its encryption requirements, but is authentication mechanism independent -- the class itself does not require the use of either Kerberos, or GSSAPI. The security class simply uses a previously negotiated encryption type, and master key. The master key is never directly used, but instead a per-connection key is derived for each new secure connection that is established.
The second portion of rxgk is a service which permits the negotiation of an encryption algorithm, and the establishment of a master key. This is done via a separate RPC exchange with a server, prior to the setup of any rxgk connections. The exchange establishes an rxgk token, and a master key shared between client and server. This exchange is protected within a GSSAPI security context.
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in RFC 2119 [RFC2119].
rxgk expresses absolute time as a 64-bit integer. This contains the time relative to midnight, or 0 hour, January 1, 1970 UTC, represented in increments of 100 nanoseconds, excluding any leap seconds. Negative times, whilst permitted by the representation, MUST NOT be used within rxgk.
Bulk data encryption within rxgk is performed using the encryption framework defined by RFC3961 [RFC3961]. Any algorithm which is defined using this framework and supported by both client and server may be used.
In order to avoid using the same key for multiple tasks, key derivation is employed. To avoid any conflicts with other users of these keys, key usage numbers are allocated within the application space documented in section 4 of RFC4120 [RFC4120]. Section 8.
rxgk supports the negotiation of a range of different security levels. These, along with the protocol constants that represent them during key negotiation, are:
The authentication only, or clear, security level provides faster throughput, at the expense of connection security. The 'clear' security level is vulnerable to a man in the middle altering the data passed over the connection, whereas the 'integrity' security level prevents such attacks by sending a cryptographic checksum of the data being transmitted.
An rxgk token is an opaque identifier which is specific to a particular application's implementation of rxgk. The token is completely opaque to the client, which just transmits it from server to server. The token MUST permit the receiving server to identify the corresponding user and session key for the incoming connection -- whether that be by decrypting the information within the token, or making the token a large random identifier which keys a lookup table on the server, or some other mechanism. It is assumed that such mechanisms will conceptually "encrypt" a token by somehow associating the "encrypted" token with the associated unencrypted data, and will "decrypt" an encrypted token by using that association to find the unencrypted data. As such, this document will use "encrypt" and "decrypt" to refer to these operations on tokens. If the token is an encrypted blob, it should be encrypted using the key usage RXGK_SERVER_ENC_TOKEN.
The token MUST NOT expose the session key on the wire. The token MUST be sufficiently random that an attacker cannot predict suitable token values by observing other connections. An attacker MUST NOT be able to forge tokens which convey a particular session key or identity.
rxgk uses an independent RX RPC service for key negotiation. The location of this service is application dependent. Within a given application protocol, a client MUST be able to locate the key negotiation service, and that service MUST be able to create tokens which can be read by the application server. The simplest deployment has the negotiation service running on every application server, on the same transport endpoints, but using a separate, dedicated, rx service ID.
The rxgk key negotiation service uses the service ID 34567.
GSS security context negotiation requires that the initiator specify a principal name for the acceptor; in the absence of application-specific knowledge, when using rxgk over a port number registered with IANA, the registered service name SHOULD be used to construct the target principal name as <service name>@%lt;hostname%gt; using the name type GSS_C_NT_HOSTBASED_SERVICE.
The key negotiation protocol is defined by the following RPC-L:
The client populates RXGK_StartParams with its preferred options. The enctypes and levels parameters are lists of values supported by the client, and MUST be ordered from best to worst, with the client's favoured option occurring first within the list. The parameters are:
The client then calls gss_init_sec_context() to obtain an output token to send to the server. The GSS service name is application dependent.
The client then calls RXGK_GSSNegotiate, as defined above. This takes the following parameters:
The client proceeds through a GSS security context initialization loop, with alternating calls to gss_init_sec_context() and the GSSNegotiate() RPC, until an error or success condition is reached. Each call to GSSNegotiate will return both an output token from GSS_Accept_sec_context() and an output opaque that are to be used as an inputs for a subsequent call to GSSNegotiate, if such a subsequent call is necessary.
Different GSS mechanisms will require a different number of full (or half) round trips. The structure of the loop, with success and error conditions noted (noting that RX level errors may occur as well but are not mentioned as part of the loop structure), is as follows:
Upon successful completion, rxgk_info contains the XDR representation of a RXGK_ClientInfo structure, encrypted using gss_wrap() with confidentiality protection. The client should decrypt this structure using gss_unwrap().
ClientInfo contains the following server populated fields:
Upon receiving the server's response, the client must verify that the mic contained within it matches the MIC of the XDR representation of the StartParams structure it sent to the server (this prevents a man in the middle from performing a downgrade attack). The client SHOULD also verify that the server's selected connection properties match those proposed by the client.
The client may then compute K0, by taking the nonce it sent to the server (client_nonce) and the one it has just received (server_nonce), combining them together, and passing them to gss_psuedo_random() with the GSS_C_PRF_KEY_FULL option:
|| is the concatenation operation.
K, the desired output length, is the key generation seed length as specified in the RFC3961 profile of the negotiated enctype.
The ouput of gss_pseudo_random must then be passed through the random-to-key operation specified in the RFC3961 profile for the negotiated enctype in order to obtain the actual key K0.
The gss_pseudo_random() operation is deterministic, ensuring that the client and server generate the same K0. The gssapi_context parameter is the same context used in the client's gss_init_sec_context() call and the server's gss_accept_sec_context() call.
A client may elect to combine multiple rxgk tokens in its possession into a single token. This allows an rx connection to be secured using a combination of multiple, individually established identities, which provides additional security for a number of application protocols.
Token combination is performed using the CombineTokens RPC call. The client has two keys -- K0 and K1, and two tokens, T0 and T1. The client calls the CombineTokens RPC with T0 and T1 and negotiates the enctype and security level of the new token, received as Tn. Tn contains the new key Kn, as computed by the server. Using the negotiated enctype returned by the server, the client then locally combines the two keys using a defined combination algorithm to produce Kn.
Assume that the tokens being combined are T0 and T1, with initial keys K0 and K1. The new initial key for the combined token, Kn is computed using the KRB-FX-CF2 operation, described in section 5.1 of [RFC6113]. The PRF+ operations will correspond to their respective key enctypes, and the random-to-key operation will correspond to the negotiated new enctype. The constants pepper1 and pepper2 required by this operation are defined as the ASCII strings "AFS" and "rxgk" respectively.
The combine keys RPC is defined as:
The server receives token0 and token1 from the RPC call, as well as the options suggested by the client. Upon receipt, the server decrypts these tokens using its private key. Providing this decryption is successful, it now has copies of the initial key (K0) from both tokens. The server then chooses an enctype and security level from the lists supplied by the client in the options argument. The server SHOULD select the first entry from each list which is acceptable in the server's configuration, so as to respect any preferences indicated by the client. The server then performs the key combination algorithm detailed above to obtain the new key, Kn. The server then constructs a new token as follows. The expiration is set to the minimum of the expiration values of the original tokens. The lifetime, bytelife, and any application-specific data fields are each combined so that the result is the most restrictive of the two values in each of the original tokens. The identity information associated with the tokens are combined in an application-specific manner to yield the identity information in the combined token (the identity combining operation may be non-commutative). This new token contains the derived key, Kn. The new token is encrypted with the server's private key, as normal, and returned to the client. The enctype and level chosen by the server are returned in the info parameter, along with the computed lifetime, bytelife, and expiration. If the server is unable to perform the CombineTokens operation with the given arguments, a nonzero value is returned in the errorcode element of the info parameter; errorcode is zero for a successful CombineTokens operation. If errorcode is nonzero, the values of the other fields in the RXGK_TokenInfo structure and the value of new_token are undefined. Nonzero values for errorcode should be com_err codes [COMERR], from an RX, RXGK, or application-specific table. See section Section 10 for RXGK error codes. section Section 8.3. For example,
As detailed within the overview, the client calls the CombineTokens RPC using two tokens, T0 and T1, within its possession, as well as an RXGK_CombineOptions structure containing a list of acceptable enctypes and a list of acceptable security levels for the new token. The client SHOULD supply these lists sorted by preference, with the most preferred option appearing first in the list. The client then receives a new token, Tn, from this call, as well as an RXGK_TokenInfo structure containing information relating to Tn. The client needs the level element of the info parameter to determine what security level to use the new token at, and the enctype parameter to know which enctype's random-to-key function to use in generating Kn. With the negotiated enctype, the client can then perform the key combination algorithm described in section Section 8.3. The client can only make use of Tn to establish an rxgk protected connection if it can derive Kn, which it can only do if it already knows K0 and K1.
Clients MUST use an rxgk secured connection for the CombineTokens operation.
When a new connection using rxgk is created by the client, the client stores the current timestamp as an rxgkTime (start_time for the rest of this discussion), and then uses this, along with other connection information, to derive a transport key from the current user's master key (see Section 8.3).
This key is then used to protect the first message the client sends to the server. The server follows the standard RX security establishment protocol, and responds to the client with a challenge [RX]. rxgk challenges simply contain a random nonce selected by the server.
Upon receiving this challenge, the client uses the transport key to encrypt an authenticator, which contains the server's nonce, and some other connection information. The client sends this authenticator, together with start_time and the current user's rxgk token, back to the server.
The server decrypts the rxgk token to determine the master key in use, uses this to derive the transport key, which it in turn uses to decrypt the authenticator, and thus validate the connection.
As part of connection negotiation, the server and client agree upon advisory lifetimes (both time, and data, based) for connection keys. Each connection has a key number, which starts at 0. When a connection exceeds one of its lifetimes, either side may elect to increment the key number. When the other endpoint sees a key number increment, it should reset all of its connection counters. Endpoints should accept packets encrypted with either the current, previous, or next key number, to allow for resends around the rekeying process.
The key version number is contained within the 16 bit spare field of the RX header (used by previous security layers as a checksum field), and expressed as an unsigned value in network byte order. If rekeying would cause this value to wrap, then the key version number MAY be stored locally as a 32-bit integer on both endpoints with only the low 16 bits transmitted on the wire. If an endpoint cannot store a per-connection 32-bit key version number when the 16-bit key version number would wrap, that endpoint MUST terminate the connection.
In order to avoid the sharing of keys between multiple connections, each connection has its own transport key, TK, which is derived from the master key, K0. Derivation is performed using the PRF+ function defined in [RFC4402], combined with the random-to-key function of K0's encryption type, as defined in RFC3961. The PRF input data is the concatenation of the rx epoch, connection ID, start_time and key number, all in network byte order. This gives:
L is the key generation seed length as specified in the RFC3961 profile.
epoch, cid and key_number are passed as 32 bit quantities, start_time is a 64 bit value.
Note that start_time is selected by the client when it receives the server's challenge, and shared with the server as part of its response. Thus both sides of the negotiation are guaranteed to use the same value for start_time.
The rxgk challenge is an XDR encoded structure with the following signature:
The rxgk response is an XDR encoded structure, with the following signature:
To check the validity of an rxgk response, the authenticator should be decrypted, the nonce from the decrypted authenticator compared with the nonce sent in the RXGK_Challenge, and the connection ID and epoch compared with that of the current connection. The call number vector (call_numbers) should be supplied to the rx implementation. Failure of any of these steps MUST result in the failure of the security context.
The way in which the rxgk security class handles packets depends upon the requested security level. As noted in Section 4, 3 levels are currently defined -- authentication only, integrity protection and encryption.
When running at the clear security level, RXGK_LEVEL_CLEAR, no manipulation of the payload is performed by the security class.
Packet payloads transmitted in the auth security level, RXGK_LEVEL_AUTH, consist of an opaque blob of MIC data followed by the unencrypted original payload data.
The MIC data is generated by calling the RFC3961 get_mic operation using a key and a data input. The RXGK_CLIENT_MIC_PACKET key usage number MUST be used for packets transmitted from the client to the server. The RXGK_SERVER_MIC_PACKET key usage number MUST be used for packets transmitted from the server to the client. The following data structure is the get_mic operation data input:
All fields MUST be in network byte order. The data length field specifies the length of the original packet payload excluding padding required for encryption routines.
The packet is transmitted with the following payload:
Note: The length of the MIC depends on which RFC3961 encryption type is used. In particular, the original packet payload may not be word-aligned.
Note: The data prepended to the original packet payload during the MIC generation is not transmitted.
Using the encryption security level, RXGK_LEVEL_CRYPT, provides both integrity and confidentiality protection.
The existing payload is prefixed with a psuedo header, to produce the following plaintext data for encryption before transmission. All fields MUST be represented in network byte order for encryption.
The data length is the length of the following data in octets, and is necessary so the receiving end can remove any padding added by the encryption routines.
This plaintext is encrypted using an RFC3961 style encrypt() function, with the connection's transport key, using key usage RXGK_CLIENT_ENC_PACKET for messages from client to server, and RXGK_SERVER_ENC_PACKET for messages from server to client, and the encrypted block transmitted to the peer.
This document specifies several erorr codes for use by RXGK implementations (see section Section 10 for the com_err table). In general, when an endpoint receives any such error code, it should abort the current operation. The various codes allow some information about why the operation failed to be conveyed to the peer so that future requests will be more likely to succeed. The circumstances in which each error code should be used are as follows:
This document requests that the AFS-3 registrar include a com_err error table for the RXGK module, as follows:
The error table base should be 1233242880, with codes within the table assigned relative numbers starting from 0 in the order appearing above.
This memo includes no request to IANA.
RX Abort packets are not protected by the RX security layer. Therefore caution should be exercised when relying on their results. In particular, clients MUST NOT use an error from GSSNegotiate or CombineTokens to determine whether to downgrade to another security class
This document permits tokens to be issued with expiration times after the expiration time of the underlying GSSAPI credential, though implementations SHOULD NOT do so. Allowing the expiration time of a credential to be artificially increased can break the invariants assumed by a security system, with potentially disastrous consequences. For example, with the krb5 GSSAPI mechanism, access revocation may be implemented by refusing to issue new tickets (or renew existing tickets) for a principal; all access is assumed to be revoked once the maximum ticket lifetime has passed. If an rxgk token is created with a longer lifetime than the kerberos ticket, this assumption is invalid, and the user whose access has supposedly been revoked may gain access to sensitive materials. An application should only allow token expiration times to be extended after a security review of the assumptions made about credential expiration for the GSSAPI mechanism(s) in use with that application. Such a review is needed to confirm that allowing token expiration times to be extended will not introduce vulnerabilities into the security eocsystem in which the application operates.
| [RX] | Zeldovich, N., "RX protocol specification" |
| [COMERR] | Raeburn, K., "A Common Error Description Library for UNIX" This paper is available as com_err.texinfo within com_err.tar.Z. |
rxgk was originally developed over a number of AFS Hackathons. The editor of this document has assembled the protocol description from a number of notes taken at these meetings, and from a partial implementation in the Arla AFS client.
Thanks to Derrick Brashear, Jeffrey Hutzelman, Love Hornquist Astrand and Chaskiel Grundman for their original design work, and comments on this document, and apologies for any omissions or misconceptions in my archaeological work.
Marcus Watts and Jeffrey Altman provided invaluable feedback on an earlier version of this document at the 2009 Edinburgh AFS Hackathon.
The text describing the rxgkTime type is based on language from Andrew Deason.
Add a reference to RFC4402, which describes the PRF+ algorithm we are using.
Change references to RXGK_Token to RXGK_Data for clarity, and add a definition of that type.
Rename the 'ticket' member of RXGK_ClientInfo to 'token', for consistency, and make it a simple opaque.
Add a length field to the packet header, so that we can remove padding.
Remove versioning in the challenge and the response.
Clarify that both bytelife and lifetime are advisory.
Remove the RXGK_CLIENT_COMBINE_ORIG and RXGK_SERVER_COMBINE_NEW key derivations, as these are no longer used.
Update the reference to draft-ietf-krb-wg-preauth-framework.
Require that CombineTokens be offered over an rxgk authenticated connection.
Pull our time definition out into its own section and define a type for it.
Define an enum for the security level, and use that throughout.
Spell check.
Remove a couple of stray references to afs_ types.
Update start_time text to clarify that it uses rxgkTime.
Make expiration also be an rxgkTime.
Add a definition for RXGK_LEVEL_BIND.
Add reference to RX.
Add reference to XDR.
Rename the gss_status output parameter from the GSSNegotiate RPC to gss_major_status, and update the supporting text.
Add a new gss_minor_status output paramter to the GSSNegotiate RPC, but make clear that it is there for informational use only.
Edit for grammar and punctuation.
Remove RXGK_LEVEL_BIND.
Make CombineTokens negotiate level and enctype.
Allow key version rollover at 16 bits when rekeying.
Add Security Considerations for increasing token expiry.
Clarify behavior at RXGK_LEVEL_AUTH.
Add RXGK com_err table and descriptions.
Clean up call number vector and maxcalls support.
Improve the description of the GSS negotiation loop.
Give suggestions for acceptor principal names.