Authentication parameters are opaque but open-ended to the rest of the RPC protocol. This section defines some flavors of authentication that have already been implemented. Other sites are free to invent new authentication types, with the same rules of flavor number assignment for program number assignment. Sun Microsystems maintains and administers a range of authentication flavors. To have authentication numbers (like RPC program numbers) allocated (or registered to them), contact the Sun RPC Administrator, as described in "Program Number Registration".
Calls are often made where the caller does not authenticate itself and the server disregards who the caller is. In these cases, the flavor value (the "discriminant" of the opaque_auth "union") of the RPC message's credentials, verifier, and response verifier is AUTH_NONE. The body length is zero when AUTH_NONE authentication flavor is used.
This is the same as the authentication flavor previously known as AUTH_UNIX. The caller of a remote procedure may wish to identify itself using traditional UNIX process permissions authentication. The flavor of the opaque_auth of such an RPC call message is AUTH_SYS. The bytes of the body encode the following structure:
struct auth_sysparms { unsigned int stamp; string machinename<255>; uid_t uid; gid_t gid; gid_t gids<10>; };
stamp is an arbitrary ID that the caller machine may generate.
machinename is the name of the caller's machine.
uid is the caller's effective user ID.
gid is the caller's effective group ID.
gids is a counted array of groups in which the caller is a member.
The flavor of the verifier accompanying the credentials should be AUTH_NONE.
When using AUTH_SYS authentication, the flavor of the response verifier received in the reply message from the server may be AUTH_NONE or AUTH_SHORT.
If AUTH_SHORT, the bytes of the response verifier's string encode a short_hand_verf structure. This opaque structure may now be passed to the server instead of the original AUTH_SYS credentials.
The server keeps a cache that maps shorthand opaque structures (passed back by way of an AUTH_SHORT style response verifier) to the original credentials of the caller. The caller can save network bandwidth and server cpu cycles by using the new credentials.
The server may flush the shorthand opaque structure at any time. If this happens, the remote procedure call message will be rejected owing to an authentication error. The reason for the failure will be AUTH_REJECTEDCRED. At this point, the caller may wish to try the original AUTH_SYS style of credentials. See Figure B-1.
AUTH_SYS authentication has the following problems:
Caller identification cannot be guaranteed to be unique if machines with differing operating systems are on the same network.
There is no verifier, so credentials can easily be faked. AUTH_DES authentication attempts to fix these two problems.
The first problem is handled by addressing the caller by a simple string of characters instead of by an operating system specific integer. This string of characters is known as the netname or network name of the caller. The server should not interpret the caller's name in any way other than as the identity of the caller. Thus, netnames should be unique for every caller in the naming domain.
It is up to each operating system's implementation of AUTH_DES authentication to generate netnames for its users that ensure this uniqueness when they call remote servers. Operating systems already distinguish users local to their systems. It is usually a simple matter to extend this mechanism to the network. For example, a user with a user ID of 515 might be assigned the following netname: "UNIX.515@sun.com". This netname contains three items that serve to ensure it is unique. Going backward, there is only one naming domain called sun.com in the Internet. Within this domain, there is only one UNIX user with user ID 515. However, there may be another user on another operating system, for example VMS, within the same naming domain who, by coincidence, happens to have the same user ID. To ensure that these two users can be distinguished you add the operating system name. So one user is "UNIX.515@sun.com" and the other is "VMS.515@sun.com".
The first field is actually a naming method rather than an operating system name. It just happens that there is almost a one-to-one correspondence between naming methods and operating systems. If the world could agree on a naming standard, the first field could be a name from that standard, instead of an operating system name.
Unlike AUTH_SYS authentication, AUTH_DES authentication does have a verifier so the server can validate the client's credential (and vice versa). The contents of this verifier are primarily an encrypted timestamp. The server can decrypt this timestamp, and if it is close to its current real time, then the client must have encrypted it correctly. The only way the client could encrypt it correctly is to know the conversation key of the RPC session. If the client knows the conversation key, it must be the real client.
The conversation key is a DES [5] key that the client generates and notifies the server of in its first RPC call. The conversation key is encrypted using a public key scheme in this first transaction. The particular public key scheme used in AUTH_DES authentication is Diffie-Hellman [3] with 192-bit keys. The details of this encryption method are described later.
The client and the server need the same notion of the current time for this to work. If network time synchronization cannot be guaranteed, then client can synchronize with the server before beginning the conversation. rpcbind provides a procedure, RPCBPROC_GETTIME, which may be used to obtain the current time.
A server can determine if a client timestamp is valid. For any transaction after the first, the server checks for two things:
The timestamp is greater than the one previously seen from the same client.
The timestamp has not expired. A timestamp is expired if the server's time is later than the sum of the client's timestamp plus what is known as the client's window. The window is a number the client passes (encrypted) to the server in its first transaction. The window can be thought of as a lifetime for the credential.
For the first transaction, the server checks that the timestamp has not expired. As an added check, the client sends an encrypted item in the first transaction known as the window verifier which must be equal to the window minus 1, or the server will reject the credential.
The client must check the verifier returned from the server to be sure it is legitimate. The server sends back to the client the encrypted timestamp it received from the client, minus one second. If the client gets anything other than this, it will reject it.
After the first transaction, the server's AUTH_DES authentication subsystem returns in its verifier to the client an integer nickname that the client may use in its further transactions instead of passing its netname, encrypted DES key and window every time. The nickname is most likely an index into a table on the server that stores for each client its netname, decrypted DES key and window. It should however be treated an opaque data by the client.
Though originally synchronized, client and server clocks can get out of sync. If this happens, the client RPC subsystem most likely will receive an RPC_AUTHERROR at which point it should resynchronize.
A client may still get the RPC_AUTHERROR error even though it is synchronized with the server. The reason is that the server's nickname table is a limited size, and it may flush entries whenever it wants. The client should resend its original credential and the server will give it a new nickname. If a server crashes, the entire nickname table will be flushed, and all clients will have to resend their original credentials.
In this scheme, there are two constants, PROOT and HEXMODULUS. The particular values chosen for these for the DES authentication protocol are:
const PROOT = 3; const HEXMODULUS = /* hex */ "d4a0ba0250b6fd2ec626e7efd637df76c716e22d0944b88b";
The way this scheme works is best explained by an example. Suppose there are two people "A" and "B" who want to send encrypted messages to each other. A and B each generate a random secret key that they do not disclose to anyone. Let these keys be represented as SK(A) and SK(B). They also publish in a public directory their public keys. These keys are computed as follows:
PK(A) = (PROOT ** SK(A)) mod HEXMODULUS PK(B) = (PROOT ** SK(B)) mod HEXMODULUS
The ** notation is used here to represent exponentiation.
Now, both A and B can arrive at the common key between them, represented here as CK(A,B), without disclosing their secret keys.
A computes:
CK(A, B) = (PK(B) ** SK(A)) mod HEXMODULUS
while B computes:
CK(A, B) = (PK(A) ** SK(B)) mod HEXMODULUS
These two can be shown to be equivalent: (PK(B)**SK(A)) mod HEXMODULUS = (PK(A)**SK(B)) mod HEXMODULUS. Drop the mod HEXMODULUS parts and assume modulo arithmetic to simplify the process:
PK(B) ** SK(A) = PK(A) ** SK(B)
Then replace PK(B) by what B computed earlier and likewise for PK(A).
((PROOT ** SK(B)) ** SK(A) = (PROOT ** SK(A)) ** SK(B)
which leads to:
PROOT ** (SK(A) * SK(B)) = PROOT ** (SK(A) * SK(B))
This common key CK(A,B) is not used to encrypt the timestamps used in the protocol. It is used only to encrypt a conversation key that is then used to encrypt the timestamps. The reason for doing this is to use the common key as little as possible, for fear that it could be broken. Breaking the conversation key is a far less serious offense, because conversations are comparatively short-lived.
The conversation key is encrypted using 56-bit DES keys, yet the common key is 192 bits. To reduce the number of bits, 56 bits are selected from the common key as follows. The middle-most 8 bytes are selected from the common key, and then parity is added to the lower order bit of each byte, producing a 56-bit key with 8 bits of parity.
To avoid compiling Kerberos code into the operating system kernel, the kernel used in the S implementation of AUTH_KERB uses a proxy RPC daemon called kerbd. The daemon exports three procedures. Refer to the kerbd(1M) manpage for more details.
KGETKCRED is used by the server-side RPC to check the authenticator presented by the client.
KSETKCRED returns the encrypted ticket and DES session key, given a primary name, instance, and realm.
KGETUCRED is UNIX-specific. It returns the user's ID, the group ID, and groups list, assuming that the primary name is mapped to a user name known to the server.
The best way to describe how Kerberos works is to use an example based on a service currently implementing Kerberos: the network file system (NFS). The NFS service on server s is assumed to have the well-known principal name nfs.s A privileged user on client c is assumed to have the primary name root and an instance c. Note that (unlike AUTH_DES) when the user's ticket-granting ticket has expired, kinit() must be reinvoked. NFS service for Kerberos mounts will fail until a new ticket-granting ticket is obtained.
This section follows an NFS mount request from start to finish using AUTH_KERB. Since mount requests are executed as root, the user's identity is root.c.
Client c makes a MOUNTPROC_MOUNT request to the server s to obtain the file handle for the directory to be mounted. The client mount program makes an NFS mount system call, handing the client kernel the file handle, mount flavor, time synchronization address, and the server's well-known name, nfs.s. Next the client kernel contacts the server at the time synchronization host to obtain the client-server time bias.
The client kernel makes the following RPC calls: (1) KSETKCRED to the local kerbd to obtain the ticket and session key, (2) NFSPROC_GETATTR to the server's NFS service, using the full name credential and verifier. The server receives the calls and makes the KGETKCRED call to its local kerbd to check the client's ticket.
The server's kerbd and the Kerberos library decrypt the ticket and return, among other data, the principal name and DES session key. The server checks that the ticket is still valid, uses the session key to decrypt the DES-encrypted portions of the credential and verifier, and checks that the verifier is valid.
The possible Kerberos authentication errors returned at this time are:
AUTH_BADCRED is returned if the verifier is invalid (the decrypted win in the credential and win +1 in the verifier do not match), or the timestamp is not within the window range
If no errors are received, the server caches the client's identity and allocates a nickname (small integer) to be returned in the NFS reply. The server then checks if the client is in the same realm as the server. If it is, the server calls KGETUCRED to its local kerbd to translate the principal's primary name into UNIX credentials. If it is not translatable, the user is marked anonymous. The server checks these credentials against the file system's export information. There are three cases to consider:
If the KGETUCRED call fails and anonymous requests are allowed, the UNIX credentials of the anonymous user are assigned.
If the KGETUCRED call fails and anonymous requests are not allowed, the NFS call fails with the AUTH_TOOWEAK.
If the KGETUCRED call succeeds, the credentials are assigned, and normal protection checking follows, including checking for root permission.
Next the server sends an NFS reply, including the nickname and server's verifier. The client receives the reply, decrypts and validates the verifier, and stores the nickname for future calls. The client makes a second NFS call to the server, and the calls to the server described earlier are repeated. The client kernel makes an NFSPROC_STATVFS call to the server's NFS service, using the nickname credential and verifier described previously. The server receives the call and validates the nickname. If it is out of range, the error AUTH_BADCRED is returned. The server uses the session key just obtained to decrypt the DES-encrypted portions of the verifier and validates the verifier.
The possible Kerberos authentication errors returned at this time are:
AUTH_REJECTEDVERF is returned if the timestamp is invalid, a replay is detected, or if the timestamp is not within the window range
AUTH_TIMEEXPIRE is returned if the service ticket is expired
If no errors are received, the server uses the nickname to retrieve the caller's UNIX credentials. Then it checks these credentials against the file system's export information, and sends an NFS reply that includes the nickname and the server's verifier. The client receives the reply, decrypts and validates the verifier, and stores the nickname for future calls. Last, the client's NFS mount system call returns, and the request is finished.
Example B-3 (AUTH_KERB) has many similarities to the one for AUTH_DES, shown in Example B-2. Note the differences.